CN110365535A - A kind of model loading method and device based on multi-party computations - Google Patents

A kind of model loading method and device based on multi-party computations Download PDF

Info

Publication number
CN110365535A
CN110365535A CN201910636095.9A CN201910636095A CN110365535A CN 110365535 A CN110365535 A CN 110365535A CN 201910636095 A CN201910636095 A CN 201910636095A CN 110365535 A CN110365535 A CN 110365535A
Authority
CN
China
Prior art keywords
state
calculate node
model
partial model
online
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910636095.9A
Other languages
Chinese (zh)
Other versions
CN110365535B (en
Inventor
陈星宇
王磊
黄群山
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Advanced New Technologies Co Ltd
Advantageous New Technologies Co Ltd
Original Assignee
Alibaba Group Holding Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Alibaba Group Holding Ltd filed Critical Alibaba Group Holding Ltd
Priority to CN201910636095.9A priority Critical patent/CN110365535B/en
Publication of CN110365535A publication Critical patent/CN110365535A/en
Application granted granted Critical
Publication of CN110365535B publication Critical patent/CN110365535B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/50—Network service management, e.g. ensuring proper service fulfilment according to agreements
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L43/00—Arrangements for monitoring or testing data switching networks
    • H04L43/10—Active monitoring, e.g. heartbeat, ping or trace-route

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Health & Medical Sciences (AREA)
  • Cardiology (AREA)
  • General Health & Medical Sciences (AREA)
  • Debugging And Monitoring (AREA)

Abstract

This specification embodiment provides a kind of model loading method and device based on multi-party computations；The described method includes:, in response to the first heartbeat message of calculate node arbitrary in multiple calculate nodes, sending the first feedback message to the calculate node when detecting that the online order of master mould is with new model loading command in command table；In response to the third heartbeat message of arbitrary calculate node, state record table is updated to the third state；And third is sent to the calculate node and feeds back message；The third state indicates that first partial model is in stress state in each calculate node, and third feeds back message and indicates online first partial model；In response to the 5th heartbeat message of arbitrary calculate node, the state record table is updated to the 5th state, and send the 5th feedback message to the calculate node；First partial model is in online state in each calculate node described in 5th state instruction, and the 5th feedback message indicates online first partial model.

Description

A kind of model loading method and device based on multi-party computations
Technical field
This specification one or more embodiment is related to computer information processing field, more particularly to a kind of more based on safety The model loading method and device just calculated.
Background technique
Multi-party computations (Secure Multi-Party Computation, MPC) are that a kind of protection data safety is hidden Private multi-party calculation method.Multi-party computations allow multiple data owners to carry out collaboration meter in the case where mutual mistrust It calculates, exports calculated result, and guarantee other any information that either party is unable to get in addition to deserved calculated result. The available data use value of MPC technology, does not reveal original data content but.
It is divided into multiple modules based on the model that MPC scheme trains, is stored in multiple calculate nodes respectively.It is multiple Calculate node belongs to different mechanisms (participant).Each mechanism only possesses a part of model of one's own module composition, And the model part of other mechanisms can not be obtained.When carrying out prediction service, only the model of some can not be formed effectively Model service, it is necessary to which combining after all mechanisms dispose respective model part could work normally.
Due to there are the demand of model iteration, how the online new model based on MPC is asking for a urgent need to resolve Topic.
Summary of the invention
This specification one or more embodiment describes the model loading method based on multi-party computations, can effectively into Row model iteration.
According in a first aspect, provide a kind of model loading method based on multi-party computations, it is applied to control node, The control node is for controlling multiple calculate nodes, and each calculate node is respectively provided with new model in the multiple calculate node Second partial model of first partial model and master mould, the first partial model of each calculate node form complete new mould Type, the local model group of the second of each calculate node is at complete master mould；Wherein, the control node maintenance has command table And state record table, the command table be used for record cast operational order, the state record table by record it is described it is each based on Model state in operator node；The described method includes:
When detecting that the online order of master mould is with new model loading command in the command table, in response to the multiple meter First heartbeat message of arbitrary calculate node in operator node sends the first feedback message to the calculate node；Wherein, first heart It jumps message and indicates that the second partial model is in online state, the first feedback message instruction load first partial in the calculate node Model, and maintain the second partial model online；
In response to the third heartbeat message of arbitrary calculate node, the state record table is updated to the third state, and Third, which is sent, to the calculate node feeds back message；Wherein, third heartbeat message indicates in the calculate node at first partial model In stress state, the third state indicates that first partial model is in stress state in each calculate node, the Three feedback messages indicate online first partial model, and the second partial model is made to enter stress state；
In response to the 5th heartbeat message of arbitrary calculate node, the state record table is updated to the 5th state, and The 5th feedback message is sent to the calculate node；Wherein, the 5th heartbeat message indicates first partial mould in the calculate node Type is in online state, and first partial model, which is in, in each calculate node described in the 5th state instruction has gone up threadiness State, the 5th feedback message indicate online first partial model, and offline second partial model.
In one embodiment, the method also includes: the method also includes: send the to the calculate node described Before three feedback messages, in response to the second heartbeat message of the calculate node, the state record table is updated to the second state, And the second feedback message is sent to the calculate node, wherein the second heartbeat message indicates the first partial model of the calculate node In stress state or the state in load；At least one in the multiple calculate node of second state instruction calculates First partial model is not in stress state, the second feedback message instruction load first partial model in node, and ties up It is online to hold the second partial model.
In one embodiment, the method also includes the method also includes sending the to the calculate node described Before five feedback messages, in response to the 4th heartbeat message of the calculate node, the state record table is updated to the 4th state, And the 4th feedback message is sent to the calculate node, wherein the 4th heartbeat message indicates first partial model in the calculate node In online state or in stress state；At least one in the multiple calculate node of 4th state instruction calculates First partial model is not in online state in node, and the 4th feedback message indicates online first partial model, and makes Second partial model enters stress state.
In one embodiment, the partial model in the partial model of online state and in stress state Participate in processing service request.
In one embodiment, the 5th feedback message only includes the order of online new model, does not include master mould Information.
In one embodiment, after sending the 5th feedback message to the calculate node, further includes: delete the order The online order of master mould in table, and new model loading command in the command table is revised as the online order of new model.
In one embodiment, the method also includes, it is described to the calculate node send the 5th feedback message after, Receive the 6th heartbeat message of each calculate node；
When the 6th heartbeat message of each calculate node does not include the state of the second partial model, by the shape State record sheet is updated to the 6th state, wherein the state record table in the 6th state does not include the second partial model State.
According to second aspect, a kind of model loading method based on multi-party computations is provided, is applied to first and calculates Node, first calculate node are any node in multiple calculate nodes of control node control；The multiple calculating section Each calculate node is respectively provided with first partial model and the second partial model in point；Wherein, the first game of each calculate node Portion's model forms complete new model, and the local model group of the second of each calculate node is at complete master mould；The method Include:
The first heartbeat message is sent to control node, the first heartbeat message indicates the second partial model in the first calculate node In online state；
According to the instruction of the first feedback message received from control node, in response to the first heartbeat message, load first Partial model, and maintain the second partial model online；
When the first partial model of the first calculate node is in stress state, third heartbeat report is sent to control node Text, third heartbeat message indicate that first partial model is in stress state in the first calculate node；
According to the instruction from the received third feedback message of control node, online first partial model, and make the second part Model enters stress state；Wherein, third feedback message is control node in response to third heartbeat message, and described in the judgement First partial model is in the message sent when stress state in each calculate node；
When the first partial model of the first calculate node is in online state, the 5th heartbeat report is sent to control node Text, the 5th heartbeat message indicate that the first partial model of the first calculate node is in online state；
It is indicated according to from the received 5th feedback message of control node, offline second partial model, and maintains first partial Model is online；Wherein, the 5th feedback message is control node in response to the 5th heartbeat message, and is judging each calculating First partial model is in the message sent when online state in node.
According to the third aspect, a kind of model eedle threader based on multi-party computations is provided, control node is set to, The control node is for controlling multiple calculate nodes, and each calculate node is respectively provided with new model in the multiple calculate node Second partial model of first partial model and master mould, the first partial model of each calculate node form complete new mould Type, the local model group of the second of each calculate node is at complete master mould；Wherein, the control node maintenance has command table And state record table, the command table be used for record cast operational order, the state record table by record it is described it is each based on Model state in operator node；Described device includes:
Transmission unit is configured to when detecting that the online order of master mould is with new model loading command in the command table, In response to the first heartbeat message of calculate node arbitrary in the multiple calculate node, the first feedback is sent to the calculate node Message；Wherein, the first heartbeat message indicates that the second partial model is in online state, the first feedback message in the calculate node Instruction load first partial model, and maintain the second partial model online；
The transmission unit is additionally configured to, and in response to the third heartbeat message of arbitrary calculate node, the state is remembered Record table is updated to the third state, and sends third to the calculate node and feed back message；Wherein, third heartbeat message indicates the calculating First partial model is in stress state in node, and the third state indicates first partial model in each calculate node It is in stress state, third feeds back message and indicates online first partial model, and enters the second partial model and loaded State；
The transmission unit is additionally configured to, and in response to the 5th heartbeat message of arbitrary calculate node, the state is remembered Record table is updated to the 5th state, and sends the 5th feedback message to the calculate node；Wherein, the 5th heartbeat message instruction should First partial model is in online state, first partial in each calculate node described in the 5th state instruction in calculate node Model is in online state, and the 5th feedback message indicates online first partial model, and offline second partial model.
According to fourth aspect, a kind of model eedle threader based on multi-party computations is provided, the first calculating is set to Node, first calculate node are any node in multiple calculate nodes of control node control；The multiple calculating section Each calculate node is respectively provided with first partial model and the second partial model in point；Wherein, the first game of each calculate node Portion's model forms complete new model, and the local model group of the second of each calculate node is at complete master mould；Described device It include: transmission unit, loading unit, online unit, offline unit；
The transmission unit is configured to, and sends the first heartbeat message, the first meter of the first heartbeat message instruction to control node The second partial model is in online state in operator node；
The loading unit is configured to, according to the first feedback received from control node, in response to the first heartbeat message The instruction of message loads first partial model, and maintains the second partial model online；
The transmission unit is additionally configured to, when the first partial model of the first calculate node is in stress state, to Control node sends third heartbeat message, and first partial model, which is in, in third heartbeat message the first calculate node of instruction has loaded State；
The online unit is configured to, according to the instruction from the received third feedback message of control node, online first game Portion's model, and the second model is made to enter stress state；Wherein, third feedback message is control node in response to third heartbeat report Text, and the message that when first partial model is in stress state in judging each calculate node sends；
The transmission unit is additionally configured to, when the first partial model of the first calculate node is in online state, to Control node sends the 5th heartbeat message, and the 5th heartbeat message indicates that the first partial model of the first calculate node is in upper threadiness State；
The offline unit is configured to, and according to from the received 5th feedback message instruction of control node, offline second is local Model, and maintain first partial model online；Wherein, the 5th feedback message is control node in response to the 5th heartbeat report Text, and the message that when first partial model is in online state in judging each calculate node sends.
According to the 5th aspect, a kind of computer readable storage medium is provided, computer program is stored thereon with, when described When computer program executes in a computer, enable computer execute first aspect method or second aspect method.
According to the 6th aspect, a kind of computing terminal, including memory and processor are provided, is stored in the memory Executable code realizes the method for first aspect or the method for second aspect when the processor executes the executable code.
The method and device that this specification embodiment provides, when the partial model of new model in each calculate node is in When stress state, the partial model of the online new model of each calculate node is just indicated；When the localized mode of new model in each calculate node After type completion is online, the partial model of the offline master mould of, to be completed as early as possible new in the case where not stopping master mould service Model is online, reduce model iteration to MPC system can externally service time influence.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, required use in being described below to embodiment Attached drawing be briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for this For the those of ordinary skill of field, without creative efforts, it can also be obtained according to these attached drawings others Attached drawing.
Fig. 1 shows the MPC system structural schematic diagram according to one embodiment；
Fig. 2 shows the MPC system structural schematic diagrams according to one embodiment；
Fig. 3 shows the model state transition diagram according to one embodiment；
Fig. 4 shows the MPC system structural schematic diagram according to one embodiment；
Fig. 5 shows the signaling interaction diagram between control node and calculate node according to one embodiment；
Fig. 6 shows the signaling interaction diagram between control node and calculate node according to one embodiment；
Fig. 7 shows the MPC system structural schematic diagram according to one embodiment；
Fig. 8 shows the signaling interaction diagram between control node and calculate node according to one embodiment；
Fig. 9 shows the MPC system structural schematic diagram according to one embodiment；
Figure 10 shows the signaling interaction diagram between control node and calculate node according to one embodiment；
Figure 11 shows the MPC system structural schematic diagram according to one embodiment；
Figure 12 shows the signaling interaction diagram between control node and calculate node according to one embodiment；
Figure 13 shows the MPC system structural schematic diagram according to one embodiment；
Figure 14 shows the signaling interaction diagram between control node and calculate node according to one embodiment；
Figure 15 shows the MPC system structural schematic diagram according to one embodiment；
Figure 16 shows the signaling interaction diagram between control node and calculate node according to one embodiment；
Figure 17 shows the MPC system structural schematic diagram according to one embodiment；
Figure 18 shows the model loading method flow chart according to one embodiment；
Figure 19 shows the model loading method flow chart according to one embodiment；
Figure 20 shows the schematic block diagram of the model loading method device according to one embodiment；
Figure 21 shows the schematic block diagram of the model loading method device according to one embodiment.
Specific embodiment
With reference to the accompanying drawing, the scheme provided this specification is described.
Model based on MPC is divided into multiple partial models (model part), and different partial models is respectively arranged at not With in the calculate node of mechanism.By taking model shown in FIG. 1 as an example, can be divided into partial model 1, partial model 2, partial model 3, Partial model 4.Wherein, partial model 1 is arranged in calculate node 1, and calculate node 1 belongs to mechanism (institution) 1；Office Portion's model 2 is arranged in calculate node 1, and calculate node 2 belongs to mechanism 2；Partial model 3 is arranged in calculate node 3, calculates section Point 3 belongs to mechanism 3, and partial model 1 is arranged in calculate node 4, and calculate node 4 belongs to mechanism 4.
For the model based on MPC, each partial model after the deployment of respective calculate node is online, just be may be provided in Line service.When any one calculate node receives the service request of user, service is collectively formed in each partial model of the model Link, with for user's negotiation result, to complete the processing to service request.
In practical applications, there can be the demand of model iteration.This needs each calculate node in each mechanism online respective The corresponding partial model for belonging to new model could provide effective model service for user.
It is first that master mould is offline when carrying out model iteration in a kind of scheme, it reloads, online new model.It is easy reason Solution, model is offline, load is required to holding time, therefore carries out model iteration according to the program, will lead to each calculate node and exists Following period of time can not provide service for user.
This specification embodiment provides a kind of model loading method based on MPC, can be each when carrying out model iteration In the case that the memory of calculate node retains the partial model of old model, the localized mode of each calculate node load new model is indicated Type；After each calculate node loads the partial model success of new model, the part of the online new model of each calculate node is just indicated Model；After the partial model online success of new model in each calculate node, the offline master mould of each calculate node is just indicated Partial model.It is online in the case where not stopping master mould service, complete new model as early as possible, model iteration is reduced to MPC The influence of systems serve time.
Next, in different embodiments, being carried out to the model loading method based on MPC that this specification embodiment provides It illustrates.
Referring to Fig.2, the model loading method application architecture based on MPC that this specification embodiment provides includes control node With multiple calculate nodes.
Control node is referred to as coordinator (Coordinator), can receive the operation life of administrator (admin) It enables, operational order may include upper line model order and lower line model order, with Controlling model iterative process.
Multiple calculate nodes can belong to different mechanisms (Institution), wherein a mechanism may include more One or both of a calculate node or more calculate node.Each calculate node is for load or online corresponding part Model.The complete model of each corresponding partial model composition of calculate node, allows each calculate node to cooperate with the online clothes of offer Business, to handle service request.
In the embodiment of the present application, service request is specifically as follows the service request that any calculate node receives.To this The processing of service request needs the participation of partial model in each calculate node.
Each calculate node periodically can send heartbeat message to control node, to report the state of its partial model.Control section The state for the partial model that point can be reported according to each calculate node, feedback the corresponding command gives each calculate node, so that each calculate Node executes relevant operation.
Next, control node is described in detail in conjunction with Fig. 2.
Control node, i.e. coordinator may include a controller (controller), a record cast operational order Order (commands) table, state (status) table for recording partial model state in each calculate node.
Command table is used for the operational order of record manager input, and operational order can specifically include online order (to_ ) and loading command (to_loaded) online.The specific structure of command table can be as shown in table 1.
Table 1
Wherein, model_id is model name, for example, " old " can be the name of master mould under the scene of model iteration Claim, " new " is the title of new model." type " is command type, wherein " TO_LOADED " is load partial model or model Order, the order that " TO_ONLINE " is online partial model or model.
State table is used to record the state of the partial model in each calculate node, and structure can be as shown in table 2.
Table 2
The current state that node_id is calculate node title, model_id is model name, status is partial model, Wherein state (state for being referred to as model) flow path switch of partial model is as shown in Figure 3.
Refering to Fig. 3, be in original state, i.e., offline (offine) state, partial model can be converted to shape in load State (loading).State refers to that the partial model of model or model is loading in load.
Model or partial model in stress state can be converted to load failure (load_failed) state, can also Successfully (loaded) state is loaded to be converted to.Load success status is referred to as stress state.Any calculating is saved For point, the partial model in stress state refers to that the partial model has loaded the memory into the node, can participate in Processing to service request.
Model or partial model in stress state can be converted to online success (online) state.Online success shape State is referred to as online state.For any calculate node, the partial model in online state refers to this The content that module has been in the node can participate in the processing to service request.
Controller (controller) in control node is for handling the order that administrator (admin) inputs to command table And the heartbeat message that each calculate node is reported.Each calculate node sends heartbeat message to control node, to report the calculating section The state of each partial model in point.The controller of control node can be according to heartbeat message more new state (status) table, so that shape State table records the last state of each partial model in each calculate node.Controller can generate corresponding according to command table, state table Feedback message feeds back to each calculate node, so that each calculate node executes corresponding operating according to feedback message.
Administrator can be to control node transmission pattern (partial model) operational order (such as online order, offline order) Its process is as follows:
A. control node can be inserted by controller into command table after receiving the online order about new model of administrator Enter the command record that a command type about new model is " TO_LOADED " (load).Then it can be incited somebody to action by heartbeat message The order is sent to each calculate node.After each calculate node receives " TO_LOADED " order about new model, starts execution and add The partial model of new model is carried, and periodically sends the state of the partial model of new model to control node.When state table shows institute Have the partial model of new model in calculate node all in stress state after, control node to each calculate node transmission includes The feedback message of " TO_ONLINE " order, to indicate the partial model of the online new model of each calculate node.When state table shows institute When having in calculate node new model partial model be in online state, control node thinks the online success of new model, can will The type of the order of new model in command table is changed to online (TO_ONLINE) by loading (TO_LOADED), and deletes former mould The type of type is the order of online (TO_ONLIE).
After the completion of controller inquiry state table finds that the partial model of the new model in all calculate nodes is all online, then Will in command table about new model " TO_LOADED " order be revised as " TO_ONLINE " order, with indicate new model it is online at Function.
B. after control node receives the offline order of administrator, controller is by " the TO_ in command table about master mould ONLINE " order is deleted, with offline master mould.It will no longer include master mould in the subsequent order to each calculate node of control node Title (ID).For any calculate node, after the message that it is received from control node does not contain the title of master mould, meeting It is automatic it is offline fall master mould therein partial model.
Next, calculate node is described in detail.
Calculate node mainly includes the partial model an of actuator (executor) and a group model (new, master mould).
The current state of each partial model of calculate node where actuator can collect it, periodically sends the heart to control node It jumps, to report the state of each partial model in its place calculate node.Actuator is also from it from the received feedback of control node The order (" TO_LOADED " or " TO_ONLINE ") that control node issues is obtained in message, and corresponding behaviour is executed according to order Make.
Wherein, for TO_LOADED order, actuator can be attempted to load the corresponding partial model of the order.For any Partial model, if it in load, state is in load (LOADING)；If it is loaded successfully, state is to have loaded (LOADED)；If its load failure, state be load failure (LOAD_FAILED), and terminate subsequent process.
Wherein, for TO_ONLINE order, then actuator attempts the corresponding partial model of the online order.If executing life The state that the order corresponding partial model is found when enabling is not to have loaded (LOADED) or online (ONLINE), then executes Device can first go to load the corresponding partial model of the order, when the state of the partial model becomes LOADED, then reattempt online The partial model.
It is readily appreciated that, service request can only be by a model treatment, in other words, can only be by belonging to the part of the same model Model interoperability handles service request.For any calculate node, when the partial model and Xin Mo of its master mould The partial model of type is respectively at stress state and online state, and when the node receives service request, then should The actuator of node can indicate that the model name in the partial model of online state is arranged in the service request, and according to Model name in the service request is set, partial model in all nodes is accessed.In other nodes, no matter the service is asked The corresponding partial model of model name in asking is to be in stress state in online state, each may participate in this The processing of service request.
For example, the calculate node of certain MPC system is calculate node 1 and calculate node 2.It is set in a certain period, if The partial model of new model in devise a stratagem operator node 1 is in online state in the partial model of stress state, master mould, The partial model of new module in calculate node 2 is in stress state in the partial model of online state, master mould.It can To assume in the period, calculate node 1 receives service request, then the actuator of calculate node 1 can be by the mould of master mould Type title (such as old) is arranged into the service request, and accesses the partial model of master mould in all calculate nodes, with processing The service request.It assume that in the period, calculate node 2 receives service request, then the actuator of calculate node 2 can incite somebody to action The model name (such as new) of new model is arranged into the service request, and accesses the part of new model in all calculate nodes Model, to handle the service request.
Need to illustrate when, this specification embodiment with " old " is master mould title, " new " is new model title, " TO_ ONLINE " is for online order, " TO_LOADED " be loading command, to model name, online order, loading command progress It illustrates, does not constitute to model name, online order, the limitation of loading command.Administrator can appointing using agreement Ideographic characters, phrase, sentence, symbol indicate model name, online order, loading command.
Next, in one embodiment, to include the MPC system of a control node and two calculate nodes, to this The model loading method for the model iteration scene that specification embodiment provides is illustrated.
In the present embodiment, two calculate nodes are respectively calculate node 1 and calculate node 2, and wherein calculate node 1 belongs to Mechanism 1, calculate node 2 belong to mechanism 2.The model_id (i.e. model name) of master mould is old, and the model_id of new model is new.The format that calculate node is sent to the heartbeat message of control node is [" model_id ": " state "].Control node is sent Format to the feedback message of calculate node is [" operational order ": " model_id "].Unless otherwise specified, it hereinafter mentions Specific instructions represented by the character of expression operation (online, load etc.) order, and indicate that state (has loaded, is online, having added Carry medium) character represented by particular state can be with reference to described above.
Refering to Fig. 4, it may be set in for the first moment, the old partial model of each calculate node is deployed online and can mention For online service.Order (commands) table in the control node at the first moment is as shown in table 3, state (status) table such as table Shown in 4.
Table 3
Table 4
Refering to Fig. 5, calculate node 1 and calculate node 2 are respectively converged the state of respective partial model by heartbeat message Control node is offered, heartbeat message particular content is [" old ": " ONLINE "], the i.e. current state of the partial model of old model For ONLINE.In response to calculate node 1 and the respective heartbeat message of calculate node 2, control node is respectively to calculate node 1 and meter Operator node 2 sends feedback message, wherein feedback message particular content is [{ " TO_ONLINE ": " old " }], that is, notifies each calculating The online respective old partial model of node, since online old partial model, calculate node need not repeatedly for each calculate node Execute online operation.
At the second moment after the first moment, administrator inputs instruction, so that the online model_id of control node For the new model of new, then control node increases a command record newly into commands table, which is TO_LOADED.This When, commands table is as shown in table 5.
Table 5
Refering to Fig. 5, controller issues feedback message to each calculate node, wherein feedback message content is [{ " TO_ ONLINE ": " old " }, { " TO_LOADED ": " new " }], that is, it indicates the online old partial model of each calculate node, loads simultaneously New model partial model.After each calculate node receives the feedback message, load new model partial model operation is executed.Each meter Operator node distinguishes regular control node and sends heartbeat message, to report respective state, wherein the content of heartbeat message be [" Old ": " ONLINE " }, { " new ": " LOADING " }], i.e., old partial model is in online state, the localized mode of new model Type is in stress state.
After control node receives the heartbeat message from calculate node, status table is updated.Updated status table is such as Shown in table 6.
Table 6
At this point, the state of MPC system can be with as shown in fig. 7, wherein in Fig. 7, at the old partial model of each calculate node In online state, new partial model is in stress state.
Third moment after the second moment, the state of MPC system are as shown in Figure 9.It is each to calculate section wherein in Fig. 9 The old partial model of point is in online state, and the new partial model of calculate node 1 is in stress state, calculate node 2 New partial model be in load in state.
I.e. at the third moment, calculate node 1 first completes the partial model load operation of new model, i.e. its new model Partial model enters stress state.And the partial model load operation of new model, i.e. its new has not been completed in calculate node 2 The partial model of model is in state in load.
Refering to Fig. 8, calculate node 1 can by heartbeat message, to control node report its new partial model state and its The state of old partial model；Wherein, heartbeat message content is [{ " old ": " ONLINE " }, { " new ": " LOADED " }], i.e., The current state of new partial model is to have loaded.In response to the heartbeat message, control node sends feedback report to calculate node 1 Text, the content for feeding back message is [{ " old ": " ONLINE " }, { " new ": " LOADED " }], to indicate that calculate node 1 maintains old The online state of partial model and the stress state of new partial model.
Calculate node 2 can report state and the part its old of its new partial model to control node by heartbeat message The state of model；Wherein, heartbeat message content is [{ " old ": " ONLINE " }, { " new ": " LOADING " }], i.e. new model The current state of partial model is in load.In response to the heartbeat message, control node sends feedback message to calculate node 2, The content for feeding back message is [{ " old ": " ONLINE " }, { " new ": " LOADED " }], to indicate that calculate node 2 maintains old innings The online state of portion's model and new partial model is made to enter stress state.
After receiving the heartbeat message of calculate node, controller updates status table, first receives calculating in setting The heartbeat message of node 1, then receive the heartbeat message of calculate node 2, then according to the heartbeat message of the calculate node 2 received As shown in table 7.
Table 7
And commands table does not change.
The 4th moment after the third moment, the state of MPC system are as shown in figure 11.It is each to calculate wherein in Figure 11 The old partial model of node is in online state, new partial model are in stress state.
Refering to fig. 10, calculate node 1 and calculate node 2 report respective localized mode to control node by heartbeat message The state of type, the content of heartbeat message are [{ " old ": " ONLINE " }, { " new ": " LOADED " }].
As shown in Figure 10, be when control node only receives the content from calculate node 1 [{ " old ": " ONLINE " }, " New ": " LOADED " }] heartbeat message, do not receive also the content from calculate node 2 be [{ " old ": " ONLINE " }, " New ": " LOADED " }] heartbeat message when, control node continues in response to the heartbeat message of calculate node 1 to calculate node 1 Sending content is [{ " old ": " ONLINE " }, { " new ": " LOADED " }].
It is [{ " old ": " ONLINE " }, { " new ": " when control node receives the content from calculate node 2 LOADED " }], control node updates status table according to heartbeat message, and updated status table is as shown in table 8.
Table 8
When the new partial model of calculate node each in status table is in online state, control node is being received After each heartbeat message, in response to each heartbeat message, the anti-of online new partial model is used to indicate to the transmission of corresponding calculate node Message is presented, content is [{ " TO_ONLINE ": " new " }, { " TO_LOADED ": " old " }].
The state at the 5th moment after the 4th moment, MPC system is as shown in figure 13.Wherein in Figure 11, section is calculated The old partial model of point 1 is in online state, new partial model and is in stress state, the old localized mode of calculate node 2 Type is in stress state, new partial model and is in online state.
Refering to fig. 12, control node be in response to the content that calculate node 1 is sent [{ " old ": " ONLINE " }, " New ": " LOADED " }] heartbeat message, to calculate node 1 send content be [{ " TO_ONLINE ": " new " }, { " TO_ LOADED ": " old " }] feedback message.Control node is [{ " old ": " in response to the content that calculate node 2 is sent LOADED " }, { " new ": " ONLINE " }] heartbeat message, to calculate node 2 send content be [{ " TO_ONLINE ": " New " }, { " TO_LOADED ": " old " }] feedback message.
Wherein, [{ " old ": " LOADED " }, { " new's ": " ONLINE " }] that control node is sent according to calculate node 1 It is as shown in table 9 to feed back the updated status table of message.
Table 9
The state at the 6th moment after the 5th moment, MPC system is as shown in figure 15.It is each to calculate wherein in Figure 15 The new partial model of node is in online state, old partial model are in stress state.
At this point, status table is as shown in table 10.
Table 10
At this point, reference table 11, control node can delete the online order of old partial model in commands table, and will The command type of new partial model is changed to TO_ONLINE.
Table 11
Refering to fig. 14, after the 6th moment, control node sends content [{ " TO_ONLINE ": " to each calculate node New " }] feedback message, wherein no longer include the order in relation to old partial model.When the feedback message that calculate node receives In when not including the related command of old partial model, then calculate node can automatic offline old partial model.
The state at the 7th moment after the 6th moment, MPC system is as shown in figure 17.It is each to calculate wherein in Figure 17 The new partial model of node is in online state, old partial model are in stress state.
Refering to fig. 16, each calculate node sends the heartbeat message that content is [{ " new ": " ONLINE " }] to control node. The information of old partial model is not included in heartbeat message, controller deletes the correlated condition of old partial model, to update Status table, updated status are as shown in table 12.
Table 12
Commands table is as shown in table 13.
Table 13
So far, each calculate node is completed that new partial model is online, and old partial model is offline, entire MPC system portion Administration when new model online service, complete model iteration.
In this embodiment, it when the partial model of new model in each calculate node is in stress state, just indicates The partial model of each online new model of calculate node；After the partial model of new model in each calculate node completes online, under The partial model of line master mould reduces model so that completion new model is online as early as possible in the case where not stopping master mould service Iteration to MPC system can externally service time influence.
For synthesis, this specification embodiment provides a kind of model loading method based on multi-party computations, application In control node, the control node is for controlling multiple calculate nodes, each calculate node difference in the multiple calculate node Second partial model of first partial model and master mould with new model, the first partial model group of each calculate node At complete new model, the local model group of the second of each calculate node is at complete master mould；Wherein, the control node Maintenance has command table and state record table, and the command table is used for record cast operational order, and the state record table is for remembering Record the model state in each calculate node.Refering to fig. 18, which comprises step 1800, when detecting the life When enabling in table the online order of master mould and new model loading command, in response to arbitrary calculate node in the multiple calculate node The first heartbeat message, to the calculate node send first feedback message；Wherein, the first heartbeat message indicates in the calculate node Second partial model is in online state, the first feedback message instruction load first partial model, and maintains the second localized mode Type is online；Step 1802, the third heartbeat message in response to arbitrary calculate node, are updated to third for the state record table State, and send third to the calculate node and feed back message；Wherein, third heartbeat message indicates first partial in the calculate node Model is in stress state, and first partial model, which is in, in the third state instruction each calculate node has loaded shape State, third feeds back message and indicates online first partial model, and the second partial model is made to enter stress state；Step 1804, In response to the 5th heartbeat message of arbitrary calculate node, the state record table is updated to the 5th state, and to the calculating Node sends the 5th feedback message；Wherein, the 5th heartbeat message indicates that first partial model is in the calculate node Online state, first partial model is in online state in each calculate node described in the 5th state instruction, and described Five feedback messages indicate online first partial model, and offline second partial model.
In some embodiments, the method also includes: it is described to the calculate node send third feedback message before, In response to the second heartbeat message of the calculate node, the state record table is updated to the second state, and to the calculate node Send the second feedback message, wherein the second heartbeat message indicates that the first partial model of the calculate node is in stress state Or the state in load；First partial at least one calculate node in the multiple calculate node of second state instruction Model is not in stress state, the second feedback message instruction load first partial model, and maintains the second partial model It is online.
In some embodiments, the method also includes, it is described to the calculate node send the 5th feedback message before, In response to the 4th heartbeat message of the calculate node, the state record table is updated to the 4th state, and to the calculate node Send the 4th feedback message, wherein the 4th heartbeat message indicates that first partial model is in online state in the calculate node Or in stress state；First partial at least one calculate node in the multiple calculate node of 4th state instruction Model is not in online state, and the 4th feedback message indicates online first partial model, and make the second partial model into Enter stress state.
In some embodiments, the partial model in the partial model of online state and in stress state Participate in processing service request.
In some embodiments, the 5th feedback message only includes the order of online new model, does not include master mould Information.
In some embodiments, after sending the 5th feedback message to the calculate node, further includes: delete the order The online order of master mould in table, and new model loading command in the command table is revised as the online order of new model.
In some embodiments, the method also includes, it is described to the calculate node send the 5th feedback message after, Receive the 6th heartbeat message of each calculate node；
When the 6th heartbeat message of each calculate node does not include the state of the second partial model, by the shape State record sheet is updated to the 6th state, wherein the state record table in the 6th state does not include the second partial model State.
Embodiment of the method shown in Figure 18 specifically can refering to the introduction above to embodiment shown in Fig. 4-Figure 17, This is repeated no more.
In this specification embodiment, when the partial model of new model in each calculate node is in stress state, Just indicate the partial model of the online new model of each calculate node；When the partial model completion of new model in each calculate node is online Afterwards, the partial model of the offline master mould of, so that completion new model is online as early as possible in the case where not stopping master mould service, Reduce influence of the model iteration to MPC system serviceable time.
In addition, this specification embodiment additionally provides a kind of model loading method based on multi-party computations, it is applied to First calculate node, first calculate node are any node in multiple calculate nodes of control node control；It is described more Each calculate node is respectively provided with first partial model and the second partial model in a calculate node；Wherein, each calculate node First partial model form complete new model, the local model group of the second of each calculate node is at complete master mould； Refering to fig. 19, which comprises step 1900 sends the first heartbeat message to control node, the first heartbeat message instruction the The second partial model is in online state in one calculate node；Step 1902, according to it is received from control node, in response to the The instruction of first feedback message of one heartbeat message, loads first partial model, and maintain the second partial model online；Step 1904, when the first partial model of the first calculate node is in stress state, third heartbeat message is sent to control node, Third heartbeat message indicates that first partial model is in stress state in the first calculate node；Step 1906, according to from control The instruction of the received third feedback message of node, online first partial model, and the second partial model is made to enter stress state； Wherein, third feedback message is control node in response to third heartbeat message, and the first game in judging each calculate node Portion's model is in the message sent when stress state；Step 1908 is in the first partial model of the first calculate node When online state, the 5th heartbeat message is sent to control node, the 5th heartbeat message indicates the first game of the first calculate node Portion's model is in online state；Step 1910 indicates, offline second part according to from the received 5th feedback message of control node Model, and maintain first partial model online；Wherein, the 5th feedback message is control node in response to the 5th heartbeat report Text, and the message that when first partial model is in online state in judging each calculate node sends.
Embodiment of the method shown in Figure 18 specifically can be refering to above to embodiment shown in Fig. 4-Figure 17 and Figure 18 institute The introduction for the embodiment of the method stated, details are not described herein.
In this specification embodiment, when the partial model of new model in each calculate node is in stress state, Just indicate the partial model of the online new model of each calculate node；When the partial model completion of new model in each calculate node is online Afterwards, the partial model of the offline master mould of, so that completion new model is online as early as possible in the case where not stopping master mould service, Reduce model iteration to MPC system can externally service time influence.
This specification embodiment provides a kind of model eedle threader 20 based on multi-party computations, is set to control section Point, the control node is for controlling multiple calculate nodes, and each calculate node is respectively provided with new mould in the multiple calculate node The first partial model composition of the first partial model of type and the second partial model of master mould, each calculate node is complete New model, the local model group of the second of each calculate node is at complete master mould；Wherein, the control node maintenance has life Table and state record table are enabled, the command table is used for record cast operational order, and the state record table is described each for recording Model state in a calculate node；Refering to Figure 20, described device 20 includes:
Transmission unit 2010 is configured to detect the online order of master mould and new model loading command in the command table When, in response to the first heartbeat message of calculate node arbitrary in the multiple calculate node, first is sent to the calculate node Feed back message；Wherein, the first heartbeat message indicates that the second partial model is in online state, the first feedback in the calculate node Message instruction load first partial model, and maintain the second partial model online；
The transmission unit 2010 is additionally configured to, in response to the third heartbeat message of arbitrary calculate node, by the shape State record sheet is updated to the third state, and sends third to the calculate node and feed back message；Wherein, third heartbeat message instruction should First partial model is in stress state in calculate node, and the third state indicates first partial in each calculate node Model is in stress state, and third feeds back message and indicates online first partial model, and enters the second partial model Stress state；
The transmission unit 2010 is additionally configured to, in response to the 5th heartbeat message of arbitrary calculate node, by the shape State record sheet is updated to the 5th state, and sends the 5th feedback message to the calculate node；Wherein, the 5th heartbeat message refers to Show in the calculate node that first partial model is in online state, first in each calculate node described in the 5th state instruction Partial model is in online state, and the 5th feedback message indicates online first partial model, and offline second part Model.
In some embodiments, the transmission unit 2010 is additionally configured to, described anti-to calculate node transmission third Before presenting message, in response to the second heartbeat message of the calculate node, the state record table is updated to the second state, and to The calculate node sends the second feedback message, wherein the second heartbeat message indicates that the first partial model of the calculate node is in Stress state or the state in load；At least one calculate node in the multiple calculate node of second state instruction Middle first partial model is not in stress state, the second feedback message instruction load first partial model, and maintains the Two partial models are online.
In some embodiments, the transmission unit 2010 is additionally configured to, and sends the 5th instead to the calculate node described Before presenting message, in response to the 4th heartbeat message of the calculate node, the state record table is updated to the 4th state, and to The calculate node sends the 4th feedback message, wherein the 4th heartbeat message indicates that first partial model is in the calculate node Online state or in stress state；At least one calculate node in the multiple calculate node of 4th state instruction Middle first partial model is not in online state, and the 4th feedback message indicates online first partial model, and makes second Partial model enters stress state.
In some embodiments, the partial model in the model of online state and in stress state each may participate in Handle service request.
In some embodiments, the 5th feedback message only includes the order of online new model, does not include master mould Information.
In some embodiments, described device 20 further includes deleting unit 2020, is configured to send out to the first calculate node After sending the 5th feedback message, the online order of master mould in the command table is deleted, and new model in the command table is loaded Order is revised as the online order of new model.
In some embodiments, described device further includes receiving unit 2030, is configured to send out described to the calculate node After sending the 5th feedback message, the 6th heartbeat message of each calculate node is received；
The transmission unit 2010 is additionally configured to, when the 6th heartbeat message of each calculate node does not include second When the state of partial model, the state record table is updated to the 6th state, wherein the state note in the 6th state Record table does not include the state of the second partial model.
Each functional unit of device 20 can realize that details are not described herein again with reference to embodiment of the method shown in Figure 18.
In this specification embodiment, when the partial model of new model in each calculate node is in stress state, Just indicate the partial model of the online new model of each calculate node；When the partial model completion of new model in each calculate node is online Afterwards, the partial model of the offline master mould of, so that completion new model is online as early as possible in the case where not stopping master mould service, Reduce model iteration to MPC system can externally service time influence.
This specification embodiment provides a kind of model eedle threader 21 based on multi-party computations, is set to the first meter Operator node, first calculate node are any node in multiple calculate nodes of control node control；The multiple calculating Each calculate node is respectively provided with first partial model and the second partial model in node；Wherein, the first of each calculate node Partial model forms complete new model, and the local model group of the second of each calculate node is at complete master mould；The dress Setting 21 includes: transmission unit 2110, loading unit 2120, online unit 2130, offline unit 2140；
Single 2110 yuan of the transmission are configured to, and send the first heartbeat message to control node, the first heartbeat message instruction the The second partial model is in online state in one calculate node；
The loading unit 2120 is configured to, according to it is received from control node, in response to the first heartbeat message first The instruction of message is fed back, loads first partial model, and maintain the second partial model online；
The transmission unit 2110 is additionally configured to, and is in stress state in the first partial model of the first calculate node When, third heartbeat message is sent to control node, third heartbeat message indicates that first partial model is in the first calculate node Stress state；
The online unit 2130 is configured to, according to the instruction from control node received third feedback message, online the One partial model, and the second model is made to enter stress state；Wherein, third feedback message is control node in response to the third heart Jump message, and the message that when first partial model is in stress state in judging each calculate node sends；
The transmission unit 2110 is additionally configured to, and is in online state in the first partial model of the first calculate node When, the 5th heartbeat message is sent to control node, the 5th heartbeat message indicates that the first partial model of the first calculate node is in Online state；
The offline unit 2140 is configured to, and is indicated according to from the received 5th feedback message of control node, offline second Partial model, and maintain first partial model online；Wherein, the 5th feedback message is control node in response to the 5th heartbeat Message, and the message that when first partial model is in online state in judging each calculate node sends.
In this specification embodiment, when the partial model of new model in each calculate node is in stress state, Just indicate the partial model of the online new model of each calculate node；When the partial model completion of new model in each calculate node is online Afterwards, the partial model of the offline master mould of, so that completion new model is online as early as possible in the case where not stopping master mould service, Reduce model iteration to MPC system can externally service time influence.
On the other hand, the embodiment of this specification provides a kind of computer readable storage medium, is stored thereon with calculating Machine program enables computer execute method shown in Figure 18 or Figure 19 when the computer program executes in a computer.
On the other hand, the embodiment of this specification provides a kind of computing terminal, including memory and processor, described to deposit It is stored with executable code in reservoir, when the processor executes the executable code, realizes side shown in Figure 18 or Figure 19 Method.
All the embodiments in this specification are described in a progressive manner, same and similar portion between each embodiment Dividing may refer to each other, and each embodiment focuses on the differences from other embodiments.Especially for system reality For applying example, since it is substantially similar to the method embodiment, so being described relatively simple, related place is referring to embodiment of the method Part explanation.
It is above-mentioned that this specification specific embodiment is described.Other embodiments are in the scope of the appended claims It is interior.In some cases, the movement recorded in detail in the claims or step can be come according to the sequence being different from embodiment It executes and desired result still may be implemented.In addition, process depicted in the drawing not necessarily require show it is specific suitable Sequence or consecutive order are just able to achieve desired result.In some embodiments, multitasking and parallel processing be also can With or may be advantageous.
Those of ordinary skill in the art should further appreciate that, describe in conjunction with the embodiments described herein Each exemplary unit and algorithm steps, can be realized with electronic hardware, computer software, or a combination of the two, in order to clear Illustrate to Chu the interchangeability of hardware and software, generally describes each exemplary group according to function in the above description At and step.These functions hold track actually with hardware or software mode, depending on technical solution specific application and set Count constraint condition.Those of ordinary skill in the art can realize each specific application using distinct methods described Function, but this realization is it is not considered that exceed scope of the present application.
The step of method described in conjunction with the examples disclosed in this document or algorithm, can hold track with hardware, processor Software module or the combination of the two implement.Software module can be placed in random access memory (RAM), memory, read-only storage Device (ROM), electrically programmable ROM, electrically erasable ROM, register, hard disk, moveable magnetic disc, CD-ROM or technology neck In any other form of storage medium well known in domain.
Above-described specific embodiment has carried out further the purpose of the present invention, technical scheme and beneficial effects It is described in detail, it should be understood that being not intended to limit the present invention the foregoing is merely a specific embodiment of the invention Protection scope, all within the spirits and principles of the present invention, any modification, equivalent substitution, improvement and etc. done should all include Within protection scope of the present invention.

Claims (18)

1. a kind of model loading method based on multi-party computations is applied to control node, the control node is for controlling Multiple calculate nodes, each calculate node is respectively provided with the first partial model and master mould of new model in the multiple calculate node The second partial model, the first partial model of each calculate node forms complete new model, each calculate node Second local model group is at complete master mould；Wherein, the control node maintenance has command table and state record table, the life Enable table for record cast operational order, the state record table is used to record the model state in each calculate node； The described method includes:
When detecting that the online order of master mould is with new model loading command in the command table, saved in response to the multiple calculating First heartbeat message of arbitrary calculate node in point sends the first feedback message to the calculate node；Wherein, the first heartbeat report Text indicates that the second partial model is in online state, the first feedback message instruction load first partial mould in the calculate node Type, and maintain the second partial model online；
In response to the third heartbeat message of arbitrary calculate node, the state record table is updated to the third state, and to this Calculate node sends third and feeds back message；Wherein, third heartbeat message indicates that first partial model is in the calculate node Stress state, the third state indicate that first partial model is in stress state in each calculate node, and third is anti- It presents message and indicates online first partial model, and the second partial model is made to enter stress state；
In response to the 5th heartbeat message of arbitrary calculate node, the state record table is updated to the 5th state, and to this Calculate node sends the 5th feedback message；Wherein, the 5th heartbeat message indicates in the calculate node at first partial model In online state, first partial model is in online state, institute in each calculate node described in the 5th state instruction It states the 5th feedback message and indicates online first partial model, and offline second partial model.
2. according to the method described in claim 1, wherein, the method also includes: third is sent to the calculate node described Before feeding back message, in response to the second heartbeat message of the calculate node, the state record table is updated to the second state, and The second feedback message is sent to the calculate node, wherein the second heartbeat message indicates at the first partial model of the calculate node In stress state or the state in load；At least one in the multiple calculate node of second state instruction calculates section First partial model is not in stress state, the second feedback message instruction load first partial model in point, and maintains Second partial model is online.
3. according to the method described in claim 1, wherein, the method also includes sending the 5th to the calculate node described Before feeding back message, in response to the 4th heartbeat message of the calculate node, the state record table is updated to the 4th state, and The 4th feedback message is sent to the calculate node, wherein the 4th heartbeat message indicates in the calculate node at first partial model In online state or in stress state；At least one in the multiple calculate node of 4th state instruction calculates section First partial model is not in online state in point, and the 4th feedback message indicates online first partial model, and makes the Two partial models enter stress state.
4. according to the method described in claim 1, wherein, in the partial model of online state and in stress state Partial model each may participate in processing service request.
5. according to the method described in claim 1, wherein, the 5th feedback message only includes the order of online new model, no Information including master mould.
6. according to the method described in claim 1, wherein, after sending the 5th feedback message to the calculate node, further includes: The online order of master mould in the command table is deleted, and it is online that new model loading command in the command table is revised as new model Order.
7. according to the method described in claim 1, wherein, the method also includes sending the 5th to the calculate node described After feeding back message, the 6th heartbeat message of each calculate node is received；
When the 6th heartbeat message of each calculate node does not include the state of the second partial model, the state is remembered Record table is updated to the 6th state, wherein the state record table in the 6th state does not include the state of the second partial model.
8. a kind of model loading method based on multi-party computations is applied to the first calculate node, first calculate node For any node in multiple calculate nodes of control node control；Each calculate node is respectively provided in the multiple calculate node First partial model and the second partial model；Wherein, the first partial model of each calculate node forms complete new model, The local model group of the second of each calculate node is at complete master mould；The described method includes:
The first heartbeat message is sent to control node, the first heartbeat message indicates that the second partial model is in the first calculate node Online state；
According to the instruction of the first feedback message received from control node, in response to the first heartbeat message, first partial is loaded Model, and maintain the second partial model online；
When the first partial model of the first calculate node is in stress state, third heartbeat message is sent to control node, Third heartbeat message indicates that first partial model is in stress state in the first calculate node；
According to the instruction from the received third feedback message of control node, online first partial model, and make the second partial model Into stress state；Wherein, third feedback message is control node in response to third heartbeat message, and is judging each meter First partial model is in the message sent when stress state in operator node；
When the first partial model of the first calculate node is in online state, the 5th heartbeat message is sent to control node, 5th heartbeat message indicates that the first partial model of the first calculate node is in online state；
It is indicated according to from the received 5th feedback message of control node, offline second partial model, and maintains first partial model It is online；Wherein, the 5th feedback message is control node in response to the 5th heartbeat message, and is judging each calculate node Middle first partial model is in the message sent when online state.
9. a kind of model eedle threader based on multi-party computations is set to control node, the control node is for controlling Multiple calculate nodes, each calculate node is respectively provided with the first partial model and master mould of new model in the multiple calculate node The second partial model, the first partial model of each calculate node forms complete new model, each calculate node Second local model group is at complete master mould；Wherein, the control node maintenance has command table and state record table, the life Enable table for record cast operational order, the state record table is used to record the model state in each calculate node； Described device includes:
Transmission unit is configured to the response when detecting that the online order of master mould is with new model loading command in the command table The first heartbeat message of arbitrary calculate node in the multiple calculate node sends the first feedback report to the calculate node Text；Wherein, the first heartbeat message indicates that the second partial model is in online state in the calculate node, and the first feedback message refers to Show load first partial model, and maintains the second partial model online；
The transmission unit is additionally configured to, in response to the third heartbeat message of arbitrary calculate node, by the state record table It is updated to the third state, and sends third to the calculate node and feeds back message；Wherein, third heartbeat message indicates the calculate node Middle first partial model is in stress state, and the third state indicates that first partial model is located in each calculate node In stress state, third feeds back message and indicates online first partial model, and the second partial model is made to enter stress state；
The transmission unit is additionally configured to, in response to the 5th heartbeat message of arbitrary calculate node, by the state record table It is updated to the 5th state, and sends the 5th feedback message to the calculate node；Wherein, the 5th heartbeat message indicates the calculating First partial model is in online state, first partial model in each calculate node described in the 5th state instruction in node It is in online state, the 5th feedback message indicates online first partial model, and offline second partial model.
10. device according to claim 9, wherein the transmission unit is additionally configured to, and is sent out described to the calculate node Before sending third to feed back message, in response to the second heartbeat message of the calculate node, the state record table is updated to second State, and the second feedback message is sent to the calculate node, wherein the second heartbeat message indicates the first partial of the calculate node Model is in stress state or the state in load；At least one in the multiple calculate node of second state instruction First partial model is not in stress state in calculate node, and the second feedback message instruction loads first partial model, And maintain the second partial model online.
11. device according to claim 9, wherein the transmission unit is additionally configured to, and is sent out described to the calculate node Before sending the 5th feedback message, in response to the 4th heartbeat message of the calculate node, the state record table is updated to the 4th State, and the 4th feedback message is sent to the calculate node, wherein the 4th heartbeat message indicates first partial in the calculate node Model is in online state or in stress state；At least one in the multiple calculate node of 4th state instruction First partial model is not in online state in calculate node, and the 4th feedback message indicates online first partial model, And the second partial model is made to enter stress state.
12. device according to claim 9, wherein the office in the model of online state and in stress state Portion's model each may participate in processing service request.
13. device according to claim 9, wherein the 5th feedback message only includes the order of online new model, no Information including master mould.
14. device according to claim 9, wherein described device further includes deleting unit, is configured to calculate to first After node sends the 5th feedback message, the online order of master mould in the command table is deleted, and by mould new in the command table Type loading command is revised as the online order of new model.
15. device according to claim 9, wherein described device further includes receiving unit, is configured to described to the meter After operator node sends the 5th feedback message, the 6th heartbeat message of each calculate node is received；
The transmission unit is additionally configured to, when the 6th heartbeat message of each calculate node does not include the second partial model State when, the state record table is updated to the 6th state, wherein the state record table in the 6th state is not wrapped Include the state of the second partial model.
16. a kind of model eedle threader based on multi-party computations is set to the first calculate node, first calculate node For any node in multiple calculate nodes of control node control；Each calculate node is respectively provided in the multiple calculate node First partial model and the second partial model；Wherein, the first partial model of each calculate node forms complete new model, The local model group of the second of each calculate node is at complete master mould；Described device include: transmission unit, loading unit, Online unit, offline unit；
The transmission unit is configured to, and sends the first heartbeat message to control node, the first heartbeat message instruction first calculates section The second partial model is in online state in point；
The loading unit is configured to, according to the first feedback message received from control node, in response to the first heartbeat message Instruction, load first partial model, and maintain the second partial model it is online；
The transmission unit is additionally configured to, when the first partial model of the first calculate node is in stress state, to control Node sends third heartbeat message, and first partial model, which is in, in third heartbeat message the first calculate node of instruction has loaded shape State；
The online unit is configured to, according to the instruction from the received third feedback message of control node, online first partial mould Type, and the second model is made to enter stress state；Wherein, third feedback message is control node in response to third heartbeat message, And the message sent when first partial model is in stress state in judging each calculate node；
The transmission unit is additionally configured to, when the first partial model of the first calculate node is in online state, to control Node sends the 5th heartbeat message, and the 5th heartbeat message indicates that the first partial model of the first calculate node is in online state；
The offline unit is configured to, and is indicated according to from the received 5th feedback message of control node, offline second partial model, And maintain first partial model online；Wherein, the 5th feedback message is control node in response to the 5th heartbeat message, and Judge that first partial model is in the message sent when online state in each calculate node.
17. a kind of computer readable storage medium, is stored thereon with computer program, when the computer program in a computer When execution, computer perform claim is enabled to require method or method according to any one of claims 8 described in any one of 1-7.
18. a kind of computing terminal, including memory and processor, executable code, the processing are stored in the memory When device executes the executable code, method of any of claims 1-7 or side according to any one of claims 8 are realized Method.
CN201910636095.9A 2019-07-15 2019-07-15 Model online method and device based on secure multi-party computation Active CN110365535B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910636095.9A CN110365535B (en) 2019-07-15 2019-07-15 Model online method and device based on secure multi-party computation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910636095.9A CN110365535B (en) 2019-07-15 2019-07-15 Model online method and device based on secure multi-party computation

Publications (2)

Publication Number Publication Date
CN110365535A true CN110365535A (en) 2019-10-22
CN110365535B CN110365535B (en) 2022-03-04

Family

ID=68219437

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910636095.9A Active CN110365535B (en) 2019-07-15 2019-07-15 Model online method and device based on secure multi-party computation

Country Status (1)

Country Link
CN (1) CN110365535B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021103792A1 (en) * 2019-11-28 2021-06-03 支付宝(杭州)信息技术有限公司 Secure multi-party computation-based machine learning model training method and apparatus, and prediction method and apparatus

Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101251797A (en) * 2008-03-31 2008-08-27 中国船舶重工集团公司第七〇九研究所 Element real time initiative transferring method based on domain model
US20110098982A1 (en) * 2009-10-22 2011-04-28 Dassault Systemes Method and system for updating a modeled object in a product lifecycle management system
CN103809438A (en) * 2012-11-05 2014-05-21 洛克威尔自动控制技术股份有限公司 Secure models for model-based control and optimization
WO2014149536A2 (en) * 2013-03-15 2014-09-25 Animas Corporation Insulin time-action model
CN105207366A (en) * 2015-10-22 2015-12-30 江苏省电力公司南京供电公司 IEC61850-based power distribution terminal, namely plug and play model configuration tool and realization method of IEC61850-based power distribution terminal
CN106446278A (en) * 2016-10-24 2017-02-22 北京亚控科技发展有限公司 Method for searching data target on basis of spatial-temporal database
CN107139931A (en) * 2016-03-01 2017-09-08 福特全球技术公司 Autonomous vehicle operation based on interactive model predictive control
CN107301221A (en) * 2017-06-16 2017-10-27 华南理工大学 A kind of data digging method of multiple features dimension heap fusion
CN108646553A (en) * 2018-04-20 2018-10-12 华中科技大学 A method of statistics on-line monitoring closed-loop control system model quality
CN109299124A (en) * 2018-09-30 2019-02-01 北京字节跳动网络技术有限公司 Method and apparatus for more new model
CN109657727A (en) * 2018-12-20 2019-04-19 四川新网银行股份有限公司 A kind of dynamic fusion method and device of machine learning model

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101251797A (en) * 2008-03-31 2008-08-27 中国船舶重工集团公司第七〇九研究所 Element real time initiative transferring method based on domain model
US20110098982A1 (en) * 2009-10-22 2011-04-28 Dassault Systemes Method and system for updating a modeled object in a product lifecycle management system
CN103809438A (en) * 2012-11-05 2014-05-21 洛克威尔自动控制技术股份有限公司 Secure models for model-based control and optimization
WO2014149536A2 (en) * 2013-03-15 2014-09-25 Animas Corporation Insulin time-action model
CN105207366A (en) * 2015-10-22 2015-12-30 江苏省电力公司南京供电公司 IEC61850-based power distribution terminal, namely plug and play model configuration tool and realization method of IEC61850-based power distribution terminal
CN107139931A (en) * 2016-03-01 2017-09-08 福特全球技术公司 Autonomous vehicle operation based on interactive model predictive control
CN106446278A (en) * 2016-10-24 2017-02-22 北京亚控科技发展有限公司 Method for searching data target on basis of spatial-temporal database
CN107301221A (en) * 2017-06-16 2017-10-27 华南理工大学 A kind of data digging method of multiple features dimension heap fusion
CN108646553A (en) * 2018-04-20 2018-10-12 华中科技大学 A method of statistics on-line monitoring closed-loop control system model quality
CN109299124A (en) * 2018-09-30 2019-02-01 北京字节跳动网络技术有限公司 Method and apparatus for more new model
CN109657727A (en) * 2018-12-20 2019-04-19 四川新网银行股份有限公司 A kind of dynamic fusion method and device of machine learning model

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021103792A1 (en) * 2019-11-28 2021-06-03 支付宝(杭州)信息技术有限公司 Secure multi-party computation-based machine learning model training method and apparatus, and prediction method and apparatus

Also Published As

Publication number Publication date
CN110365535B (en) 2022-03-04

Similar Documents

Publication Publication Date Title
US7818679B2 (en) Method, system, and apparatus for enabling near real time collaboration on an electronic document through a plurality of computer systems
CN104054301B (en) Remotely access the session transmission and hang-up in application framework
CN104247380B (en) CRUD type agreements are bound in distributed agreement protocol
JP2007067725A (en) Contents distribution method, contents distribution server, communication terminal device, and contents distribution system
Grier et al. A social history of Bitnet and Listserv, 1985-1991
CN105159603A (en) Repair method for distributed data storage system
CN102667728B (en) Delivery with reconciliation on client side
CN104750738B (en) A kind of method for updating the data information, back end, management node and system
EP2440999A1 (en) Consensus based distributed task execution
CN105740047A (en) Cloud computing based virtual machine template management system and method
CN105991752B (en) A kind of data block update method and equipment
CN108989399A (en) A kind of cluster common data querying method, system, device and readable storage medium storing program for executing
CN110365535A (en) A kind of model loading method and device based on multi-party computations
CN108664521A (en) A kind of method, apparatus, electronic equipment and readable storage medium storing program for executing that sequence number is provided
JP2012058960A (en) Shared file management system, control method of the same and program
CN109831323B (en) Server information management method, management system and server
Sethi et al. Design and Optimization of a Zookeeper and Kafka-Based Messaging Broker for Asynchronous Processing in High Availability and Low Latency Applications
CN109491767A (en) The processing method and distributed system of distributed transaction
Hu et al. Transactional mobility in distributed content-based publish/subscribe systems
CN109359851A (en) A kind of task processing method based on micro- assistance the superior and the subordinate office group
CN107766156A (en) Task processing method and device
CN103347253A (en) Large-scale off-line short message distributed sending system and method
CN108460078A (en) Execution method, apparatus, storage medium and the terminal of miscellaneous function
CN114168545A (en) Data migration method and device
CN115221895A (en) Cross-scene task type dialogue method and system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
TA01 Transfer of patent application right
TA01 Transfer of patent application right

Effective date of registration: 20201010

Address after: Cayman Enterprise Centre, 27 Hospital Road, George Town, Grand Cayman Islands

Applicant after: Innovative advanced technology Co.,Ltd.

Address before: Cayman Enterprise Centre, 27 Hospital Road, George Town, Grand Cayman Islands

Applicant before: Advanced innovation technology Co.,Ltd.

Effective date of registration: 20201010

Address after: Cayman Enterprise Centre, 27 Hospital Road, George Town, Grand Cayman Islands

Applicant after: Advanced innovation technology Co.,Ltd.

Address before: A four-storey 847 mailbox in Grand Cayman Capital Building, British Cayman Islands

Applicant before: Alibaba Group Holding Ltd.

GR01 Patent grant
GR01 Patent grant