CN108768876A - A kind of traffic scheduling method of Machine oriented learning framework - Google Patents
A kind of traffic scheduling method of Machine oriented learning framework Download PDFInfo
- Publication number
- CN108768876A CN108768876A CN201810569876.6A CN201810569876A CN108768876A CN 108768876 A CN108768876 A CN 108768876A CN 201810569876 A CN201810569876 A CN 201810569876A CN 108768876 A CN108768876 A CN 108768876A
- Authority
- CN
- China
- Prior art keywords
- stream
- group
- priority
- detection
- module
- 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
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L47/00—Traffic control in data switching networks
- H04L47/10—Flow control; Congestion control
-
- 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
- H04L47/00—Traffic control in data switching networks
- H04L47/50—Queue scheduling
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L47/00—Traffic control in data switching networks
- H04L47/50—Queue scheduling
- H04L47/62—Queue scheduling characterised by scheduling criteria
- H04L47/625—Queue scheduling characterised by scheduling criteria for service slots or service orders
- H04L47/6275—Queue scheduling characterised by scheduling criteria for service slots or service orders based on priority
Abstract
The present invention proposes a kind of traffic scheduling method of Machine oriented learning framework, it is distributed machines learning framework flow scheduling mechanism in a kind of efficient data center, in the case where the scene using stream information can not be obtained, efficient scheduling strategy is realized using the self-similarity of machine learning flow in group level of stream.The mechanism organically combines the rate control of stream with flow scheduling, effective stream information has been helped to speculate the completion in stream transmission procedure by timely rate control, while the scheduling strategy based on estimation result has reasonably guided rate control of the stream under different network environments.
Description
Technical field
The present invention relates to a kind of schemes improving machine learning application traffic scheduling performance in data center network, belong to meter
Calculation machine network field.
Background technology
It is artificial to its that the technological break-through in machine learning field in recent years so that more and more large scale business companies increase
The input of intelligent use is researched and developed.In order to promote research and development progress, each company to be proposed different machine learning frames and carry out fully profit
With the computing resource of physical computer cluster.The scheduling of resource is particularly significant to being efficiently completed machine learning task in cluster,
The distribution of wherein Internet resources is particularly critical.It usually will produce a large amount of flows in machine learning tasks make progress, these flows
It is easy to cause network congestion into cluster network (data center network) and then extends the deadline of task.Network congestion generates
Reason of both mainly having：(1) lack application semantics and perceive mechanism, traditional data center network is due to cannot be distinguished difference
Using the differentiated demand for network, the fair services provided can cause so that network becomes the bottle that application performance is promoted
Neck；(2) traditional network transmission speed controlling mechanism is not suitable for data center network, the traffic aggregation in data center network
Pattern is easy to cause data-bag lost, to reduce transmission performance.
Before machine learning task occupies the extensive work load of cluster, group stream has proven to a kind of and effectively carries
Rise the valid model of the network performance of distributed computing framework in data center.It is better than traditional base based on a group scheduling scheme for stream
It is in the reason of scheme of stream, group stream contains real-time requirement of the Distributed Application to network.For example it for, comes from
The a plurality of stream of the same Distributed Application reaches the same recipient by different links, and application requirement recipient completes
Next calculation stages could be entered after the transmission of all streams；Congestion occurs for link where wherein one stream at this time, then in base
In the scheduling scheme of stream, the congestion signal of a stream can only have an impact the rate control mechanism of this stream itself；And it is right
For seeing above-mentioned a plurality of stream the group stream scheduling scheme of a group stream as in logic, then the speed of other streams can be suitably reduced
Rate avoids unnecessary bandwidth occupancy.
It therefore, it is expected to, realize that the flow scheduling in distributed machines learning framework group fluid layer face will bring great net
Network performance boost.
But regrettably, effect is limited if the group stream scheme for lacking stream information.This is because realizing distributed
After the flow scheduling in machine learning frame set fluid layer face, group stream scheduling strategy determines group stream scheduling performance when congestion occurs.
Different from the scheduling scheme based on stream, the definition of group stream itself determines that it has optimal scheduling strategy：The priority of group stream
It is determined by stream most slow in group stream, this is because the deadline of group stream depends on the completion of the last one stream.The prior art because
Lack effective group stream scheme of stream information, therefore effect is limited, the quantity according to stream transmission data packet is needed to carry out pre- flow measurement
Size simultaneously places it in different priority queries, and scheduling performance can not dependent on the rate control that the accuracy of prediction flows simultaneously
Explicitly controlled.
Invention content
The purpose of the present invention is to solve group stream scheme the problem of, propose a kind of flow tune of Machine oriented learning framework
Degree method.
In the traffic scheduling method of Machine oriented learning framework of the present invention, the machine learning is for passing through data
Parallel model obtains different machine learning models under large-scale data set；In the machine learning, large-scale number
Multiple distributed nodes are divided into according to collection to be stored；The working example of distributed node is run on according to local part number
According to collection training and the Grad of model parameter is obtained, and is sent to the update that hyper parameter server carries out model；Hyper parameter service
Device converges multigroup Grad and carries out model training, and will send back to working example under updated model parameter；It is characterized in that,
The stream that multiple working examples are sent to the same hyper parameter server is organized into a group stream, hyper parameter server is sent to multiple
The stream tissue of example becomes another group stream, realizes distributed machines learning framework in group flow scheduling in fluid layer face.
Further, further include that group stream information speculates mechanism, for detecting a group potential congestion for stream in as fast as possible
Ability, group stream information speculate that mechanism includes the following steps：S1, after machine learning task starts, it is logical based on a group stream Scheduling Framework
The quantity of statistics active stream is crossed to get the number n flowed in a group stream；S2, the random stream selected in group stream are as spy
Flow measurement simultaneously ensures that it completes the transmission of data packet as early as possible, to obtain the size f of its stream；S3, in conjunction with machine study group stream from
The size of similitude, a group stream can be obtained by n*f；The size of the group stream is equivalent to for the potential of shared forwarding node
Congestion ability, and by a foundation of the priority as one group stream of judgement；S4, priority update is carried out.
Further, come from group stream in newly generated machine learning task to need just be added by information supposition
In being closed to active group adfluxion；During information speculates, according to the edge switch belonging to receiving terminal hyper parameter server with
Machine chooses a group stream in multiple groups of streams in new task；For the group stream being selected, residing for transmitting terminal working example
Physical host randomly selects a stream as detection stream.
Further, the data packet from detection stream is labeled with label and enters explorer queue, enjoys highest priority；It visits
By being randomly assigned in the group stream that randomly chooses, that is, the selection for detecting stream is designed flow measurement using dual random.
Further, the highest priority combination flexible rate control algolithm of stream is detected to ensure detection stream transmission rate
Rapid growth.
Further, before priority update startup, the non-detection stream from new task enters active queue, to ensure
The minimum transmission rate that itself do not died of hunger under current network configuration is given out a contract for a project.
Further, when flow priority update timer expiry, all non-detection stream, comprising do not complete group stream,
The result of detection flowed according to the detection belonging to it enters corresponding priority query；Detection stream size passes through influence group stream size
Speculate to determine that the priority of non-detection stream, priority update ensure the scheduling strategy of short job priority.
Further, flexible rate control algolithm includes：In the case where there is enough available bandwidths, the rate of stream is in again
Increase quickly to occupy available bandwidth to promote link utilization.
Further, each RTT information for flowing passed through link is available to estimate in end system acquisition group stream
Bandwidth realizes the rate control based on end system；And in such a way that similar setting sends window size come with existing TCP/IP
Protocol-compliant.
Further, for a stream f, the desired number of data packets sent of targeted rate, that is, next cycle, by
Following formula is calculated：
elastic_gain*f.max_rtt/f.min_rtt
Wherein preset value elastic_gain determines that the minimum transmission rate of stream, f.max_rtt indicate institute's energy in link
The maximum queue length of receiving, size are related to network configuration；F.min_rtt is indicated in a measurement period in link
Queue length.
The invention also includes a kind of flow scheduling systems of Machine oriented learning framework, for realizing above-mentioned method, packet
Controller is included, the newest stream information periodically reported for receiving transmitting terminal, realization group stream semantic analysis function；Controller
In comprising a group stream information acquisition module, communication pattern matching module and group stream size speculate module；Group stream information acquisition module
For flowing the stream tissue from machine learning task in groups according to existing stream information, and identify the group from different cycles of training
Stream；Communication pattern matching module is for recording completed group of stream to match current unfinished group stream；The group of successful match
Stream then need not enter group stream a size speculate a module, belonging to stream directly issue decision to receiving terminal using matched result；Group
It flows size and speculates that module speculates algorithm for realizing above-mentioned group stream information, new group stream is divided into detection stream to flow with non-detection
And according to group result of stream information acquisition module come update it is non-detection stream priority；
Further, controller is additionally operable to periodically issue scheduling decision to receiving terminal；Module positioned at terminal passes through
Received group Flow Policy is realized in elastomeric flow scheduling, wherein comprising priority labeling module, measurement result update module with
And transport layer Rate control module；Priority labeling module receives priority update information of the controller for each stream in group stream
Afterwards, corresponding priority is marked to the data packet of each stream and ensures its multi-level feedback priority query entered in operating system
To realize shortest-job-first strategy；Measurement result update module acquire it is end-to-end between link RTT information and a upper week
Transmission packet quantity in phase and receiver packet number amount, to provide data basis for the transport layer Rate control module of next step；Transmission
Layer Rate control module is used to calculate the transmission packet quantity in next cycle, and with the help of rate limiter of speed, it is ensured that has
Corresponding transmission packet is normally sent out from network interface card.
Compared with prior art, the present invention proposes distributed machines learning framework flow in a kind of efficient data center
Scheduling mechanism realizes efficient scheduling strategy in the case where that can not obtain the scene using stream information in group level of stream.
Further, which organically combines the rate control of stream with flow scheduling, passes through timely rate control
It makes the effective stream information of help and speculates the completion in stream transmission procedure, while the scheduling strategy based on estimation result is rational
Rate control of the stream under different network environments is guided.
Description of the drawings
Fig. 1 is the deployment in the data center of distributed machines learning framework and the signal of data center network canonical topology
Figure；
Fig. 2 is the group stream example schematic diagram in distributed machines learning framework in one embodiment of the invention；
Fig. 3 is that group stream information of the embodiment of the present invention speculates flow diagram；
Fig. 4 is rate control strategy schematic diagram of the embodiment of the present invention based on link queue；
Fig. 5 is flexible rate control flow schematic diagram of the embodiment of the present invention；
Fig. 6 is model framework schematic diagram of the embodiment of the present invention；
Fig. 7a-7c is group stream dispatching algorithm exemplary plot of the embodiment of the present invention.
Reference numeral：Tn：The edge switch that number is n；Hn-m：The object of n-th of edge switch of access that number is m
Manage host；VM1：The fictitious host computer that number is 1, D1:No. 1 working example based on Hadoop；S1:No. 1 based on Spark
Working example.
Specific implementation mode
With reference to embodiment and compares attached drawing the present invention is described in further details.Wherein identical attached drawing
Label indicates identical component, unless stated otherwise.It is emphasized that following the description is only exemplary, without
It is to limit the scope of the invention and its apply.
The deployment in the data center of distributed machines learning framework and data center network topology are as shown in Fig. 1：
One physical host H is usually run multiple distributed work examples in a manner of virtual machine VM and (for example comes from Computational frame
The D of Hadoop, or Spark S)；Multiple working examples in the same physical host share physical resource, such as CPU, interior
It deposits, hard disk and network bandwidth；Each physical host accesses internal network (data center's internal network by edge switch T
Refer to being accessed by edge switch, the network organization that is built by multiple high performance switch of inside), at the same edge switch to
Affiliated physical host distributes the flow from internal network.
It has been generally acknowledged that congestion will not occur for internal network, congestion occurs mainly in data center by edge switch and object
Manage the network portion that host is constituted, i.e. external network.Common data center network congestion can be divided into transmitting terminal congestion and side
Edge switch congestion.The former is since the multiple working examples run in individual host cause data packet can not seizing for bandwidth
It leaves host and enters network；The latter is made because the flow from internal network has been more than the transfer capability of edge switch
Final jump of the data packet in routed path is obtained to be dropped.
It can see in conjunction with the example of above-mentioned group of stream, if we can be by multiple work from the same physical host
The stream of example is organized as a group stream, alternatively, a plurality of stream for reaching the same edge switch is organized as a group stream, so that it may
To give full play to a group scheduling advantage for stream.
Fortunately, the calculating semanteme support group stream scheduling scheme of distributed machines learning framework.Point of most of mainstreams
Cloth machine frame is built based on hyper parameter server model as shown in Fig. 2.In such systems, machine learning is appointed
The target of business is that different machine learning models is obtained under large-scale data set by data parallel model.Large-scale number
Multiple distributed nodes are divided into according to collection to be stored；The working example of distributed node is run on according to local part number
According to collection training and the Grad of model parameter is obtained, and is sent to the update that hyper parameter server carries out model；Hyper parameter service
Device converges multigroup Grad and carries out model training, and will send back to working example under updated model parameter.Due to hyper parameter
Server needs that until a certain amount of gradient result model training could be started, and working example is needed until hyper parameter server
Issuing for updated model parameter finishes the gradient that could start new round calculating, and multiple working examples can be sent to same by we
The stream of a hyper parameter server is organized into a group stream (the group stream from task 1), and hyper parameter server is sent to multiple examples
Stream tissue become a group stream (the group stream from task 2).Compare the congestion in the above-mentioned data center network that we mention
Type can see, the former advantageously accounts for the edge switch congestion of hyper parameter server end downlink, and the latter is conducive to
Solve the transmitting terminal congestion of hyper parameter server end uplink.
The inventors discovered that different from the group stream of other Distributed Applications, machine learning group stream, which has, following several has pole
The feature of big potential value：1, it unlike other multistage application tasks, can see from above-mentioned example, machine learning task is only
Have two stages, which dictates that because the stage is different migration will not occur for the terminal of group stream to ensure that a group flow structure
The stability of (transmitting terminal and receiving terminal that each flow)；2, it is drawn during the semanteme applied is similar and a machine learning task
Data block size after point is fixed, therefore the size flowed in different cycles of training is consistent；3, due in super ginseng
In number server model, the model parameter that a hyper parameter server is responsible for is that equivalent divides, each hyper parameter server
It is mutually similar between corresponding same type group stream.Features above can be to be collectively referred to as the self similarity of machine learning group stream
Property.
The inventors discovered that the self-similarity of machine learning group stream is conducive to simplify the design of group Flow Policy.The present invention is real
It is exactly to construct distributed machines learning framework flow scheduling mechanism in efficient data center using these discoveries to apply example.Tool
Body is described as follows：
1. group stream information speculates mechanism
Group stream information speculates that the purpose of mechanism is to detect a group potential congestion ability for stream in as fast as possible.It considers
Short job priority be currently known shortening task completion time optimal algorithm, it is assumed that stream transmission rate stablize it is constant
Under the premise of, we are using maximum group of stream as the group stream for most possible initiation congestion here.
How key problem that group stream in longest stream deadline be design efficient group stream scheduling scheme is calculated.Tradition
Group stream scheme need to calculate objective result in conjunction with available bandwidth under the scene of size each flowed in known group of stream, and
The rate each flowed is updated according to result.This scheme is difficult to be suitable for the unknown scene of stream information, and is visited dependent on bandwidth
The accuracy of survey, while large-scale update stream transmission rate will bring great resource overhead.
Due to the group flow structure in machine learning frame have stability, we can after machine learning task starts,
Based on existing group of stream Scheduling Framework the number n flowed in a group stream is got by counting the quantity of active stream.If I
The random stream selected in group stream flowed as detection and ensure that it completes the transmission of data packet as early as possible, flow to obtaining it
Size f, in conjunction with the self-similarity of machine study group stream, the size of a group stream can be obtained by n*f.Due to group stream size by etc.
Valence is the potential congestion ability for shared forwarding node, therefore organizes stream size by the priority as one group stream of judgement
Main Basiss.
Information supposition and priority update are two cores in group stream information supposition algorithm, as shown in Fig. 3.Come
It is needed in speculating that can just be added to active group adfluxion closes by information from stream is organized in newly generated machine learning task.?
During information speculates, multiple groups are randomly selected in new task according to the edge switch belonging to receiving terminal (hyper parameter server)
A group stream in stream.For the group stream being selected, one is randomly selected according to the physical host residing for transmitting terminal (working example)
A stream is flowed as detection.Data packet from detection stream is labeled with label and enters explorer queue, enjoys highest priority.Detection stream
It is randomly assigned in the group stream obtained by random selection (see step 2,3 in Fig. 3)；Detection stream choose dual random design be in order to
Collide competition bandwidth between avoiding detection to flow as possible.The highest priority of detection stream can ensure that itself data packet measures
Link Round-Trip delay (RTT) it is minimum, this point combination flexible rate control algolithm can ensure detection flows transmission rate it is fast
Speed increases.
Before priority update startup, the non-detection stream from new task enters active queue, to ensure itself working as
The minimum transmission rate that do not died of hunger under preceding network configuration is given out a contract for a project.When flow priority updates timer expiry, all is non-
Detection stream (including the group stream not completed) enters corresponding priority query according to the result of detection for detecting stream belonging to it.Detection
Stream size determines the priority of non-detection stream by the supposition of influence group stream size, therefore combines the scheduling of short job priority former
From the point of view of then, the priority of the detection stream smaller non-detection stream of size is higher.Priority update ensures the scheduling of short job priority
Implementation of strategies, because when the size of detection stream constantly increases, the priority of the non-detection stream determined can be lowered gradually
(see step 9) in Fig. 3.
Group stream supposition mechanism support group stream scheduling scheme realizes efficient scheduling strategy.It is sent compared to traditional by flowing
Data volume number carry out the mode (LAS) of decision set flow priority, the result of detection that detection stream provides can be on realization theory more
Excellent shortest-job-first strategy (SJF), as shown in attached drawing 7a.Task J1, the J2 occurred simultaneously there are two in this example,
Middle J1 includes two group streams C1 and C2, and C1 includes that the stream that three sizes are 1 shares a receiving terminal H1, others group flow structure ginseng
See Fig. 7 a, repeats no more.Assuming that each chronomere of link handles the data of a unit, come from group angle of stream deadline
It sees, this patent proposes that scheduling SJF schemes (such as Fig. 7 c) are better than scheme (such as Fig. 7 b) under traditional LAS schedule strategy.This is
Because LAS need in t=3 could group stream of the Differentiated Services from different task, and SJF can be in t=2 just by J1
It is greater than J1 to the detection of H3 stream area come J2 in group stream size earlier in H1 and J2.
2. flexible rate control algolithm
Detection stream and non-detection stream propose rate control strategy different requirements.Ideal rate control strategy is for example attached
Shown in Fig. 4.Detection stream carries the mission that group stream size speculates, additivity increases obviously not as good as the increasing of multiplying property has in traditional rate control
Effect.Therefore in the case where there is enough available bandwidths, the rate of stream should be in that multiplication Calais quickly occupies available bandwidth to be promoted
Link utilization.As shown in bandwidth detection part in Fig. 4, when the data packet number for the transmission flowed in a upper detection cycle is m
When there is not packet loss, then the data packet number sent in the period can increase to 2*m.At the same time, higher rate is not
Should be by random network fluctuation, such as a small amount of data packet discarding, and be greatly lowered, significantly so as to cause rate
Fluctuation；Rational strategy should perceive the smaller rear reduction of speed appropriate of the queue occurred in link, to protect current rate
It is horizontal.As shown in Fig. 4 medium-rates protection part, when packet loss number is n in the upper a cycle of stream, the period interior number sent
It can be adjusted to m-n according to packet number.Opposite, the target for detecting stream is as far as possible will be available in intrinsic rate and link
Bandwidth matches.Detection stream can receive significantly rate fluctuation and avoid congestion, especially when exist in link occupy it is big
When measuring the detection stream of bandwidth.As shown in Congestion Avoidance part in Fig. 4, when the hair of the m data packet sent in the upper a cycle of stream
The transmission for averagely wanting waiting n data packet is sent, the data packet number sent in the period can be adjusted to m/n.
The flexible rate control algolithm that this patent proposes can as far as possible meet in the case of smaller overhead
Above-mentioned requirements, as shown in Fig. 5.Since machine learning group stream shares the same receiving terminal (or transmitting terminal), we can hold
Each RTT information for flowing passed through link estimates available bandwidth in system acquisition group stream.Rate control based on end system
System avoids calculating and communication overhead caused by centralized rate calculations, and can be very good to send window by similar setting
The mode of mouthful size come it is compatible with existing ICP/IP protocol, to reduce development cost.It should be noted that the transmission rate of stream
The upper limit do not do here pressure limitation.
Flexible rate control algolithm, which regards arbitrarily complicated link as, has the direct-connected of identical RTT sizes and bottleneck bandwidth
Link.The receptible maximum queue length of institute, size and network configuration phase in link are indicated for stream a f, f.max_rtt
It closes；F.min_rtt indicates the queue length in link in a measurement period, this is because in Modern High-Speed data center net
The forward delay of data packet, processing delay and transmission delay can be ignored compared to the queueing delay that queue is brought in network.
Currently transmitted rate can be equivalent to indicate the number of data packets in upper a cycle transmitted by transmitting terminal (due to week in logic
Phase size is changeless, and the data packet more multilist of transmission shows that the rate of transmission is faster)；Currently receive in rate representation one
The number of data packets that receiving terminal is received in period, to realize rate Preservation tactics；Targeted rate indicates that next cycle is wished
It hopes the number of data packets sent, (is calculated see step 1) in Fig. 5 by formula in Fig. 5：
elastic_gain*f.max_rtt/f.min_rtt
Wherein preset value elastic_gain determines the minimum transmission rate of stream, because the queue in link reaches
In limited time, f.max_rtt/f.min_rtt=1；Last counted transmission rate indicates that next cycle is practical and sends out the data sent
Packet number, wherein multi_gain indicate the radical degree of bandwidth detection.
The present embodiment speculates a mechanism by group stream information, avoid group stream scheduling scheme to acquisition group stream information in advance according to
Rely, the complexity of scheduling scheme design is reduced to the excavation of machine learning group stream self-similarity, and simplifies existing group of stream
Consequently facilitating disposing, flexible rate controls the performance in Optimized Operation strategy while improving the utilization rate of link Scheduling Framework,
Reduce the average completion time of machine learning group stream.
This programme needs to simplify existing centralization group stream Scheduling Framework, and part of module is migrated to receiving terminal
It realizes, as shown in Fig. 6.
Algorithm shown in attached drawing 3 is located at group stream semantic module, algorithm shown in attached drawing 5 in Fig. 6 and is located at elastomeric flow tune in Fig. 6
Spend module.The main realization group stream semantic analysis function of controller, wherein comprising a group stream information acquisition module, communication pattern matches mould
Block and group stream size speculate module.For transmitting terminal periodically to the newest stream information of controller report, group stream information acquires mould
Block flows the stream tissue from machine learning task according to existing stream information in groups, and identifies the group from different cycles of training
Stream.Communication pattern matching module is responsible for recording completed group of stream to match current unfinished group stream；The group of successful match
Stream then need not enter group stream a size speculate a module, belonging to stream directly issue decision to receiving terminal using matched result.Group
Stream size speculates that module mainly realizes that above-mentioned group stream information speculates algorithm, and new group stream is divided into detection stream to flow with non-detection
And according to group result of stream information acquisition module come update it is non-detection stream priority.
Controller periodically issues scheduling decision to receiving terminal.Module positioned at terminal is realized by elastomeric flow scheduling
Received group Flow Policy, wherein including priority labeling module, measurement result update module and transport layer rate control
Module.After priority labeling module receives controller for the priority update information of each stream in group stream, to the number of each stream
It marks corresponding priority according to packet and ensures it into the multi-level feedback priority query in operating system and realize that short operation is excellent
First strategy.Measurement result update module acquire it is end-to-end between link RTT information and the transmission packet number in upper a cycle
Amount and receiver packet number amount, to provide data basis for the transport layer Rate control module of next step.Transport layer Rate control module
It is main to realize algorithm described function shown in attached drawing 5, calculate the transmission packet quantity in next cycle and in rate limiter of speed
With the help of, it is ensured that there is corresponding transmission packet normally to be sent out from network interface card.
The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be said that
The specific implementation of the present invention is confined to these explanations.For those skilled in the art to which the present invention belongs, it is not taking off
Under the premise of from present inventive concept, several equivalent substitute or obvious modifications can also be made, and performance or use is identical, all answered
When being considered as belonging to protection scope of the present invention.
Claims (12)
1. a kind of traffic scheduling method of Machine oriented learning framework, the machine learning is used for through data parallel model big
Different machine learning models is obtained under the data set of scale；In the machine learning, large-scale data set is divided into
Multiple distributed nodes are stored；The working example of distributed node is run on according to local partial data collection training and is obtained
To the Grad of model parameter, and it is sent to the update that hyper parameter server carries out model；Hyper parameter server converges multigroup ladder
Angle value carries out model training, and will send back to working example under updated model parameter；It is characterized in that, by multiple work reality
The stream that example is sent to the same hyper parameter server is organized into a group stream, and hyper parameter server is sent to the stream tissue of multiple examples
As another group stream, realize distributed machines learning framework in group flow scheduling in fluid layer face.
2. the traffic scheduling method of Machine oriented learning framework according to claim 1, it is characterised in that including a group stream letter
Supposition mechanism is ceased, in group potential congestion ability for stream that detects as fast as possible, group stream information to speculate that mechanism includes as follows
Step：
S1, after machine learning task starts, get one by counting the quantity of active stream based on a group stream Scheduling Framework
The number n flowed in group stream；
S2, the random stream selected in group stream flow as detection and ensure that it completes the transmission of data packet as early as possible, to obtain
Its size f flowed；
S3, in conjunction with the self-similarity of machine study group stream, the size of a group stream can be obtained by n*f；The size of the group stream by etc.
Valence is the potential congestion ability for shared forwarding node, and by a foundation of the priority as one group stream of judgement；
S4, priority update is carried out.
3. the traffic scheduling method of Machine oriented learning framework according to claim 2, which is characterized in that come from new production
Stream is organized in raw machine learning task to need just be added to by information supposition in active group adfluxion conjunction；Speculate in information
In the process, one in new task in multiple groups of streams is randomly selected according to the edge switch belonging to receiving terminal hyper parameter server
Group stream；For the group stream being selected, a stream is randomly selected as detection according to the physical host residing for transmitting terminal working example
Stream.
4. the traffic scheduling method of Machine oriented learning framework according to claim 3, which is characterized in that from detection stream
Data packet be labeled with label enter explorer queue, enjoy highest priority；In the group stream that detection stream is obtained by random selection with
Machine is specified, that is, the selection for detecting stream is designed using dual random.
5. the traffic scheduling method of Machine oriented learning framework according to claim 4, which is characterized in that detect stream most
High priority combination flexible rate control algolithm is to ensure the rapid growth of detection stream transmission rate.
6. the traffic scheduling method of Machine oriented learning framework according to claim 2, which is characterized in that priority more
Before new startup, the non-detection stream from new task enters active queue, to ensure itself not starved under current network configuration
Dead minimum transmission rate is given out a contract for a project.
7. the traffic scheduling method of Machine oriented learning framework according to claim 6, which is characterized in that work as flow priority
Update timer expiry, all non-detection stream, comprising unfinished group stream, the result of detection that flows according to the detection belonging to it into
Enter corresponding priority query；Detection stream size determines the priority of non-detection stream by the supposition of influence group stream size, excellent
First grade update ensures the scheduling strategy of short job priority.
8. the traffic scheduling method of Machine oriented learning framework according to claim 5, it is characterised in that flexible rate control
Algorithm processed includes：In the case where there is enough available bandwidths, the rate of stream quickly occupies available bandwidth to carry in multiplication Calais
Rise link utilization.
9. the traffic scheduling method of Machine oriented learning framework according to claim 8, it is characterised in that：It is adopted in end system
Each RTT information for flowing passed through link estimates available bandwidth in collection group stream, realizes the rate control based on end system；
And come compatible with existing ICP/IP protocol in such a way that similar setting sends window size.
10. the traffic scheduling method of Machine oriented learning framework according to claim 9, which is characterized in that for one
F is flowed, the desired number of data packets sent of targeted rate, that is, next cycle is calculated by following formula：
elastic_gain*f.max_rtt/f.min_rtt
Wherein preset value elastic_gain determines that the minimum transmission rate of stream, f.max_rtt indicate to receive in link
Maximum queue length, size is related to network configuration；F.min_rtt indicates the queue in link in a measurement period
Length.
11. a kind of flow scheduling system of Machine oriented learning framework, special for realizing the method described in claim 1-10
Sign is, including controller, the newest stream information periodically reported for receiving transmitting terminal, realization group stream semantic analysis work(
Energy；Speculate module comprising group stream information acquisition module, communication pattern matching module and group stream size in controller；
Group stream information acquisition module is identified for flowing the stream tissue from machine learning task in groups according to existing stream information
Group stream from different cycles of training；
Communication pattern matching module is for recording completed group of stream to match current unfinished group stream；The group of successful match
Stream then need not enter group stream a size speculate a module, belonging to stream directly issue decision to receiving terminal using matched result；
Group stream size speculates that module speculates algorithm for realizing above-mentioned group stream information, by new group stream be divided into detection stream with it is non-
Detection flows and updates the priority of non-detection stream according to group result of stream information acquisition module.
12. the flow scheduling system of Machine oriented learning framework as claimed in claim 11, which is characterized in that controller is also used
In periodically issuing scheduling decision to receiving terminal；Module positioned at terminal realizes received group by elastomeric flow scheduling
Flow Policy, wherein including priority labeling module, measurement result update module and transport layer Rate control module；
After priority labeling module receives controller for the priority update information of each stream in group stream, to the data of each stream
Packet marks corresponding priority and ensures it into the multi-level feedback priority query in operating system to realize short job priority
Strategy；
Measurement result update module acquire it is end-to-end between link RTT information and transmission packet quantity in upper a cycle and
Receiver packet number amount, to provide data basis for the transport layer Rate control module of next step；
Transport layer Rate control module is used to calculate the transmission packet quantity in next cycle, and in the help of rate limiter of speed
Under, it is ensured that there is corresponding transmission packet normally to be sent out from network interface card.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810569876.6A CN108768876B (en) | 2018-06-05 | 2018-06-05 | Traffic scheduling method facing machine learning framework |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810569876.6A CN108768876B (en) | 2018-06-05 | 2018-06-05 | Traffic scheduling method facing machine learning framework |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108768876A true CN108768876A (en) | 2018-11-06 |
| CN108768876B CN108768876B (en) | 2022-01-11 |
Family
ID=63999879
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810569876.6A Active CN108768876B (en) | 2018-06-05 | 2018-06-05 | Traffic scheduling method facing machine learning framework |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108768876B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110958187A (en) * | 2019-12-17 | 2020-04-03 | 电子科技大学 | Distributed machine learning parameter-oriented synchronous differential data transmission method |
| CN111078659A (en) * | 2019-12-20 | 2020-04-28 | 腾讯科技（深圳）有限公司 | Model updating method, model updating device, computer readable storage medium and computer equipment |
| CN111612155A (en) * | 2020-05-15 | 2020-09-01 | 湖南大学 | Distributed machine learning system and communication scheduling method suitable for same |
| CN111628940A (en) * | 2020-05-15 | 2020-09-04 | 清华大学深圳国际研究生院 | Flow scheduling method, device, system, switch and computer storage medium |
| CN113194086A (en) * | 2021-04-27 | 2021-07-30 | 新华三信息安全技术有限公司 | Anti-attack method and device |
| CN113839884A (en) * | 2020-06-24 | 2021-12-24 | 华为技术有限公司 | Flow control method and device |
| CN115062771A (en) * | 2022-08-16 | 2022-09-16 | 之江实验室 | Distributed machine learning gradient convergence method and device and model training method |
| CN115271102A (en) * | 2022-09-26 | 2022-11-01 | 太极计算机股份有限公司 | Task-oriented priority method and system for machine learning engine |
| CN115499306A (en) * | 2022-07-29 | 2022-12-20 | 天翼云科技有限公司 | Method and device for constructing traffic scheduling model, electronic equipment and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20030133411A1 (en) * | 1997-10-23 | 2003-07-17 | Kabushiki Kaisha Toshiba | Communication resource management method and node control device using priority control and admission control |
| CN101692657A (en) * | 2009-10-22 | 2010-04-07 | 北京交通大学 | Differentiated service core router and data forwarding method thereof |
| US20100135158A1 (en) * | 2008-12-01 | 2010-06-03 | Razoom, Inc. | Flow State Aware QoS Management Without User Signalling |
| CN104852887A (en) * | 2014-02-17 | 2015-08-19 | 上海宽带技术及应用工程研究中心 | Network flow tracing system and method based on OpenFlow technology |
| EP3200410A1 (en) * | 2016-01-28 | 2017-08-02 | Alcatel Lucent | Method and system for queueing packets in communication networks |
-
2018
- 2018-06-05 CN CN201810569876.6A patent/CN108768876B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20030133411A1 (en) * | 1997-10-23 | 2003-07-17 | Kabushiki Kaisha Toshiba | Communication resource management method and node control device using priority control and admission control |
| US20100135158A1 (en) * | 2008-12-01 | 2010-06-03 | Razoom, Inc. | Flow State Aware QoS Management Without User Signalling |
| CN101692657A (en) * | 2009-10-22 | 2010-04-07 | 北京交通大学 | Differentiated service core router and data forwarding method thereof |
| CN104852887A (en) * | 2014-02-17 | 2015-08-19 | 上海宽带技术及应用工程研究中心 | Network flow tracing system and method based on OpenFlow technology |
| EP3200410A1 (en) * | 2016-01-28 | 2017-08-02 | Alcatel Lucent | Method and system for queueing packets in communication networks |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110958187A (en) * | 2019-12-17 | 2020-04-03 | 电子科技大学 | Distributed machine learning parameter-oriented synchronous differential data transmission method |
| CN110958187B (en) * | 2019-12-17 | 2021-05-18 | 电子科技大学 | Distributed machine learning parameter-oriented synchronous differential data transmission method |
| CN111078659A (en) * | 2019-12-20 | 2020-04-28 | 腾讯科技（深圳）有限公司 | Model updating method, model updating device, computer readable storage medium and computer equipment |
| CN111078659B (en) * | 2019-12-20 | 2023-04-21 | 腾讯科技（深圳）有限公司 | Model updating method, device, computer readable storage medium and computer equipment |
| CN111612155A (en) * | 2020-05-15 | 2020-09-01 | 湖南大学 | Distributed machine learning system and communication scheduling method suitable for same |
| CN111628940A (en) * | 2020-05-15 | 2020-09-04 | 清华大学深圳国际研究生院 | Flow scheduling method, device, system, switch and computer storage medium |
| CN111612155B (en) * | 2020-05-15 | 2023-05-05 | 湖南大学 | Distributed machine learning system and communication scheduling method suitable for same |
| CN113839884A (en) * | 2020-06-24 | 2021-12-24 | 华为技术有限公司 | Flow control method and device |
| CN113839884B (en) * | 2020-06-24 | 2023-08-22 | 华为技术有限公司 | Flow control method and device |
| CN113194086A (en) * | 2021-04-27 | 2021-07-30 | 新华三信息安全技术有限公司 | Anti-attack method and device |
| CN113194086B (en) * | 2021-04-27 | 2022-05-27 | 新华三信息安全技术有限公司 | Anti-attack method and device |
| CN115499306A (en) * | 2022-07-29 | 2022-12-20 | 天翼云科技有限公司 | Method and device for constructing traffic scheduling model, electronic equipment and storage medium |
| CN115499306B (en) * | 2022-07-29 | 2024-03-12 | 天翼云科技有限公司 | Method and device for constructing flow scheduling model, electronic equipment and storage medium |
| CN115062771B (en) * | 2022-08-16 | 2022-11-25 | 之江实验室 | Distributed machine learning gradient convergence method and device and model training method |
| CN115062771A (en) * | 2022-08-16 | 2022-09-16 | 之江实验室 | Distributed machine learning gradient convergence method and device and model training method |
| CN115271102B (en) * | 2022-09-26 | 2022-12-16 | 太极计算机股份有限公司 | Task-oriented priority method and system for machine learning engine |
| CN115271102A (en) * | 2022-09-26 | 2022-11-01 | 太极计算机股份有限公司 | Task-oriented priority method and system for machine learning engine |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108768876B (en) | 2022-01-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108768876A (en) | A kind of traffic scheduling method of Machine oriented learning framework | |
| Meng et al. | Online deadline-aware task dispatching and scheduling in edge computing | |
| Chen et al. | Round-robin synchronization: Mitigating communication bottlenecks in parameter servers | |
| CN104714852B (en) | A kind of parameter synchronization optimization method and its system suitable for distributed machines study | |
| CN104993941B (en) | One kind is based on Openflow network high fault tolerance virtual network mapping algorithms | |
| CN102724103B (en) | Proxy server, hierarchical network system and distributed workload management method | |
| CN108282415A (en) | A kind of dispatching method and equipment | |
| CN109104373A (en) | The processing method of network congestion, apparatus and system | |
| CN110381541A (en) | A kind of smart grid slice distribution method and device based on intensified learning | |
| CN103825838B (en) | A kind of data center removes bandwidth fragmentation stream scheduling method | |
| CN109492753A (en) | A kind of method of the stochastic gradient descent of decentralization | |
| Zhang et al. | Tuning the aggressive TCP behavior for highly concurrent HTTP connections in intra-datacenter | |
| CN110990140B (en) | Method for scheduling distributed machine learning flow in photoelectric switching network | |
| CN106934454B (en) | Test-schedule method in network on three-dimensional chip based on Petri network | |
| CN109445386A (en) | A kind of most short production time dispatching method of the cloud manufacturing operation based on ONBA | |
| CN103699433A (en) | Method and system for performing dynamic adjustment on number of tasks in Hadoop platform | |
| CN111767146A (en) | Distributed machine learning system acceleration method based on network reconfiguration | |
| CN109714795A (en) | A kind of method for managing resource, resource management system and device based on SDN network slice | |
| CN114938372B (en) | Federal learning-based micro-grid group request dynamic migration scheduling method and device | |
| CN109976873A (en) | The scheduling scheme acquisition methods and dispatching method of containerization distributed computing framework | |
| CN105591790B (en) | Data communication connection pool management device | |
| Tao et al. | DRL-Driven Digital Twin Function Virtualization for Adaptive Service Response in 6G Networks | |
| Zhang et al. | ATFQ: a fair and efficient packet scheduling method in multi-resource environments | |
| CN108616569B (en) | A kind of Based on Distributed calculates the light Measurement Request dispatching method of application | |
| CN106533756B (en) | A kind of communication feature extracts, flow generation method and device |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |