CN104767778A - Task processing method and device - Google Patents
Task processing method and device Download PDFInfo
- Publication number
- CN104767778A CN104767778A CN201410006581.XA CN201410006581A CN104767778A CN 104767778 A CN104767778 A CN 104767778A CN 201410006581 A CN201410006581 A CN 201410006581A CN 104767778 A CN104767778 A CN 104767778A
- Authority
- CN
- China
- Prior art keywords
- node
- cluster
- communication path
- function
- task
- 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.)
- Withdrawn
Links
Landscapes
- Data Exchanges In Wide-Area Networks (AREA)
Abstract
The invention discloses a task processing method and a device. The method comprises steps: a connection path of nodes in a cluster is planned, wherein the cluster is at least determined by the task for executing the task, and the cluster comprises multiple nodes; and information of the connection path is sent to nodes bearing an exchange function in the cluster to build a cluster network, wherein the nodes bearing the exchange function are nodes for forwarding data according to received connection path information in the multiple nodes. Thus, the problems that stability and efficiency of a computing cluster system are not high due to unreasonable dividing of the network and storage resources can be solved, and the utilization rate of the network resources is improved.
Description
Technical field
The present invention relates to the communications field, in particular to task processing method and device.
Background technology
Along with the development of information age, occurred an increasing calculating difficult problem, the price modeling as fluid dynamics, impact simulation, seismic processing, figure comparison, financial modeling, weather simulation and complexity all needs a large amount of computational resource.High-performance calculation (HPC) scheme in correlation technique is all adopt large-scale cluster to realize scientific algorithm, and it is mainly considered to strengthen single node performance, increases the parallelization degree of system interior joint quantity and raising calculation task to pursue performance boost.
In correlation technique, the inside PC cluster software of High-Performance Computing Cluster, only for the scheduling of individual node ability, does not consider the actual influence of Internet resources restriction to high-performance calculation.Under this actual networking condition, between different computing tasks, between different user, the careless slightly unordered harmful competition namely forming the network bandwidth, forms bottleneck.
In the related, existing high-performance calculation, what store internal network and data does not require.But in actual applications, it is not often restriction due to computational resource that high performance computing system performance cannot reach requirement, but divide the access bottleneck formed due to network, the unreasonable of storage resources.Such as numerous at number of nodes, the inner complex condition interconnected by multistage network, because the disk of burst large discharge that often occurs of user operation reason and network operation in the system, well can not adapt to the change of Internet resources actual conditions, largely have impact on stability and the efficiency of HPCC system.
For in correlation technique, because the unreasonable division of network, storage resources causes the problem that the stability of computing cluster system and efficiency are not high, at present effective solution is not yet proposed.
Summary of the invention
The invention provides a kind of task processing method and device, at least to solve network, the unreasonable division of storage resources causes the problem that the stability of computing cluster system and efficiency are not high.
According to an aspect of the present invention, provide a kind of task processing method, comprising: the communication path planning the node in described cluster, wherein, described cluster is that described cluster comprises multiple node at least according to the cluster for performing described task that task is determined; The information of described communication path sent to the node bearing function of exchange in described cluster to build described cluster network, wherein, the node bearing function of exchange described in is the node for the information forwarding data according to the described communication path received in described multiple node.
Preferably, at least determine that the cluster performing described task comprises according to task: the cluster at least determining to perform described task according to the load of described task and network.
Preferably, after the described cluster network of structure, described method also comprises: the communication path again planning the node in described cluster according to the situation change of described cluster; The information of the communication path again planned is sent to the node bearing function of exchange in described cluster.
Preferably, after again planning the communication path of the node in described cluster according to the situation change of described cluster, described method also comprises: send the instruction cancelling the information of former communication path, wherein, the described node bearing function of exchange carries out buffer memory according to described instruction for the message of the information matches of the former communication path of cancelling, after the information receiving the communication path again planned, reopen the forwarding of buffer memory message; The described node bearing function of exchange, when not receiving the information of the communication path again planned, is deleted the message of buffer memory or reports the information not receiving the communication path again planned.
Preferably, the communication path of the node again planned in described cluster according to the change of the situation of described cluster comprise following one of at least: the load determining the node in described cluster, and the communication path load of the node of described cluster being planned again the node in described cluster as weights; One or more node needing lower electricity or power on is determined according to the load of the node in described cluster; Again communication path is planned according to the node in lower electricity or the described cluster after powering on.
Preferably, described in bear function of exchange node determine that the network load of self exceedes threshold value, then carry out flow control according to the priority of task.
Preferably, node in described cluster comprise in Types Below one of at least: multiple computing node, multiple memory node, multiplely deposit node, multiple switching node, wherein, described computing node is for performing calculation function, or, for performing calculation function and bearing function of exchange; Described memory node is for performing data storage function, and described meter deposits node for performing for performing calculation function and data storage function, or, for performing calculation function, data storage function and bearing function of exchange.
Preferably, node realizes described switching node function by intelligent network adapter and/or virtual switch deposited by described meter.
Preferably, the mode that the information of described communication path is shown by stream sends.
Preferably, described method also comprises: after task completes, and reclaims the list-directed Internet resources of described stream.
According to another aspect of the present invention, additionally provide a kind of Task Processing Unit, it is characterized in that comprising: planning module, for planning the communication path of the node in described cluster, wherein, described cluster is that described cluster comprises multiple node at least according to the cluster for performing described task that task is determined; Sending module, for the information of described communication path being sent to the node bearing function of exchange in described cluster to build described cluster network, wherein, the node bearing function of exchange described in is the node for the information forwarding data according to the described communication path received in described multiple node.
Preferably, described planning module is used at least determining according to the load of described task and network the cluster performing described task.
Preferably, described planning module, also for again planning the communication path of the node in described cluster according to the situation change of described cluster; Described sending module, also for the information of the communication path again planned being sent to the node bearing function of exchange in described cluster.
Preferably, described sending module is also for sending the instruction of the information cancelling former communication path, wherein, the described node bearing function of exchange carries out buffer memory according to described instruction for the message of the information matches of the former communication path of cancelling, after the information receiving the communication path again planned, reopen the forwarding of buffer memory message; And described in bear the node of function of exchange when not receiving the information of the communication path again planned, delete buffer memory message or report the information not receiving the communication path again planned.
Preferably, the load of the node of described cluster also for determining the load of the node in described cluster, and is planned the communication path of the node in described cluster by described planning module again as weights; Described planning module is also for determining one or more node needing lower electricity or power on according to the load of the node in described cluster; Again communication path is planned according to the node in lower electricity or the described cluster after powering on.
Preferably, described in bear the node of function of exchange for when determining that the network load of self exceedes threshold value, then carry out flow control according to the priority of task.
Preferably, node in described cluster comprise in Types Below one of at least: multiple computing node, multiple memory node, multiplely deposit node, multiple switching node, wherein, described computing node is for performing calculation function, or, for performing calculation function and bearing function of exchange; Described memory node is for performing data storage function; Described meter deposits node for performing for performing calculation function and data storage function, or, for performing calculation function, data storage function and bearing function of exchange.
Preferably, described in bear the node of the other types of node disjoint in described cluster of function of exchange, and/or, described in bear the node of function of exchange function realized by the node of the other types in described cluster.
Preferably, node realizes described switching node function by intelligent network adapter and/or virtual switch deposited by described meter.
Preferably, the mode that the information of described communication path is shown by stream sends.
Preferably, described device also comprises: recycling module, for after task completes, reclaims the list-directed Internet resources of described stream.
By the present invention, adopt the communication path of the node in this cluster of planning, wherein, described cluster is that described cluster comprises multiple node at least according to the cluster for performing described task that task is determined; The information of this communication path sent to the node bearing function of exchange in this cluster to build this cluster network, wherein, this node bearing function of exchange is the node for the information forwarding data according to this communication path received in the plurality of node, solve network, the unreasonable division of storage resources causes the problem that the stability of computing cluster system and efficiency are not high, improve the utilance of Internet resources.
Accompanying drawing explanation
Accompanying drawing described herein is used to provide a further understanding of the present invention, and form a application's part, schematic description and description of the present invention, for explaining the present invention, does not form inappropriate limitation of the present invention.In the accompanying drawings:
Fig. 1 is the flow chart of the task processing method according to the embodiment of the present invention;
Fig. 2 is the structured flowchart one according to the Task Processing Unit of the embodiment of the present invention;
Fig. 3 is the structured flowchart two according to the Task Processing Unit of the embodiment of the present invention;
Fig. 4 is the schematic network structure of the HPCC applying SDN technology according to the preferred embodiment of the invention;
Fig. 5 is the schematic diagram of intelligent network adapter data retransmission according to the preferred embodiment of the invention;
Fig. 6 is the flow chart of Internet resources QOS access according to the preferred embodiment of the invention;
Fig. 7 is the flow chart realizing load balancing according to the preferred embodiment of the invention;
Fig. 8 is the flow chart realizing according to the preferred embodiment of the invention saving energy and reduce the cost.
Embodiment
It should be noted that, when not conflicting, the embodiment in the application and the feature in embodiment can combine mutually.Below with reference to the accompanying drawings and describe the present invention in detail in conjunction with the embodiments.
In the present embodiment, provide a kind of task processing method, Fig. 1 is the flow chart of the task processing method according to the embodiment of the present invention, and as shown in Figure 1, this flow process comprises the steps:
Step S102, the communication path of the node in planning cluster, wherein, this cluster is that this cluster comprises multiple node at least according to the cluster for executing the task that task is determined;
Step S104, sends to the node bearing function of exchange in cluster to build cluster network by the information of communication path, wherein, this node bearing function of exchange is the node for the information forwarding data according to this communication path received in multiple node.
Pass through above-mentioned steps, can cluster be built according to the needs of task and plan communication path, the information of communication path sent to the node bearing function of exchange in cluster to build this cluster network, because the cluster set up is corresponding with performing of task, the foundation of this cluster network is efficiently balanced, relative to the way not planning cluster for task in correlation technique, it can solve network, the unreasonable division of storage resources causes the problem that the stability of computing cluster system and efficiency are not high, thus improve the utilance of Internet resources.
In the present embodiment, for performing different business, corresponding cluster can be set up according to the actual connection state of current network and load intensity, preferably, can comprise according to the cluster completing node that this required by task wants and determine to perform this task: the cluster at least determining to perform this task according to the load of this task and network.
In the present embodiment, can also adjust communication path dynamically, such as, again the communication path of the node in this cluster can be planned according to the situation of change of cluster, wherein, the situation of this cluster can comprise following one of at least: the integral load situation change of cluster, one or more load that this bears the node of function of exchange exceed increase or the minimizing of the node in threshold value, this cluster, and the information of the communication path again planned are sent to the node bearing function of exchange in this cluster.Such as, if a load bearing the node of function of exchange exceedes threshold value, then can consider in this cluster, increase the node bearing function of exchange, after the node of function of exchange is born in increase, again plan communication path.Again such as, if the burden of computing node is relatively lighter, or some computing node is even in the free time, now just these nodes can be deleted from this cluster, again can plan communication path after deletion.
Preferably, the communication path of the node again planned in this cluster according to the change of the situation of cluster can comprise following one of at least: the load determining the node in this cluster, and the load of the node of this cluster is planned again the communication path of the node in this cluster as weights, one or more node needing lower electricity or power on is determined according to the load of the node in this cluster; Again communication path is planned according to the node in lower electricity or this cluster after powering on.
The node of function of exchange is born for certain or some, it can belong to cluster corresponding to multiple different task, now this load bearing the node of function of exchange may exceed the load of self, in a preferred embodiment, the node bearing function of exchange can determine whether the network load of self exceedes threshold value, then carry out flow control according to the priority of task.
In the cluster of the present embodiment, node can comprise following classification according to the difference of function: node deposited by computing node, memory node, meter, switching node, preferably, node in above-mentioned cluster comprise following one of at least: multiple computing node, multiple memory node, multiple meter deposit node, multiple switching node, and wherein, this computing node is for performing calculation function, or, for performing calculation function and bearing function of exchange; This memory node is for performing data storage function; This meter deposits node for performing for performing calculation function and data storage function, or, for performing calculation function, data storage function and bearing function of exchange.
In the present embodiment, node realizes switching node function by intelligent network adapter and/or virtual switch deposited by meter.
In above-described embodiment, the mode that the information of communication path can be shown by stream sends, and after task completes, the list-directed Internet resources of convection current reclaim.
In the present embodiment, additionally provide a kind of Task Processing Unit, this Task Processing Unit, for realizing above-mentioned method, to have carried out explanation, had not repeated them here in above-described embodiment and preferred implementation.It should be noted that, the title of the module in following apparatus does not form the actual restriction to this module, such as, determination module can be expressed as " at least determining the module of the cluster performing this task according to task ", and following module all can realize within a processor, such as, determination module can be expressed as " a kind of processor, at least determining according to task the cluster performing this task ", or, " a kind of processor, comprises determination module " etc.
Fig. 2 is the structured flowchart one according to the Task Processing Unit of the embodiment of the present invention, and as shown in Figure 2, this device comprises: planning module 22 and sending module 24, be described this structure below.
Planning module 22, for planning the communication path of the node in cluster, wherein, this cluster is that this cluster comprises multiple node at least according to the cluster for executing the task that task is determined;
Sending module 24, for the information of communication path being sent to the node bearing function of exchange in cluster to build this cluster network, wherein, this node bearing function of exchange is the node for the information forwarding data according to this communication path received in the plurality of node.
Pass through said apparatus, can cluster be built according to the needs of task and plan communication path, the information of communication path sent to the node bearing function of exchange in cluster to build this cluster network, because the cluster set up is corresponding with performing of task, therefore cluster network runs is efficiently balanced, relative to the way not planning cluster for task in correlation technique, it can solve network, the unreasonable division of storage resources causes the problem that the stability of computing cluster system and efficiency are not high, thus improve the utilance of Internet resources.
In the present embodiment, planning module 22 can also be used for the cluster at least determining to perform this task according to the load of this task and network.
In the present embodiment, planning module 22, can also be used for the communication path again planning the node in this cluster according to the situation change of this cluster, this sending module 24, also for the information of the communication path again planned being sent to the node bearing function of exchange in this cluster.
In the present embodiment, sending module 24 can also be used for the instruction sending the information cancelling former communication path, wherein, this node bearing function of exchange carries out buffer memory according to this instruction for the message of the information matches of the former communication path of cancelling, after the information receiving the communication path again planned, reopen the forwarding of buffer memory message, and this node bearing function of exchange is not when receiving the information of the communication path again planned, deletes the message of buffer memory or reporting the information not receiving the communication path again planned.
In the present embodiment, this planning module 22 can also be used for the load of the node determined in this cluster, and the load of the node of this cluster is planned again the communication path of the node in this cluster as weights, this planning module 22, also for determining one or more node needing lower electricity or power on according to the load of the node in this cluster, plans communication path again according to the node in lower electricity or this cluster after powering on.
Fig. 3 is the structured flowchart two according to the Task Processing Unit of the embodiment of the present invention, as shown in Figure 3, this device comprises: planning module 22, sending module 24 and recycling module 32, wherein, planning module 22, sending module 24 and the planning module 22 in above-mentioned Fig. 2, sending module 24 function are identical, no longer illustrate at this, below this device is described.
Recycling module 32, for after task completes, the list-directed Internet resources of convection current reclaim.
Be described below in conjunction with preferred embodiment and execution mode.
In the preferred embodiment, relate to software defined network (Software Defined Network, referred to as SDN), SDN is that a kind of new network proposed by clean slate seminar of Stanford Univ USA innovates framework, its core technology OF(OpenFlow) controller is by separating network equipment chain of command and data surface, thus achieve the flexible control of network traffics, for the innovation of core network and application provides good platform.SDN is made up of software control and hardware data passage.Software control comprises management (CLI, SNMP) and Routing Protocol (OSPF, ISIS, BGP) etc.Data channel comprises for the inquiry of each bag, exchange and buffer memory.If the network equipments all in network to be considered as the resource be managed, so with reference to the principle of operating system, the abstract on the one hand detail of underlay network device of the concept of a network operating system (Network OS)-this network operating system can be taken out, simultaneously also for upper layer application provides unified administration view and DLL (dynamic link library).Like this, this platform of operating system Network Based, user can develop various application program, defines network topology in logic by software, to meet the different demands to Internet resources, and without the need to being concerned about the physical topological structure of bottom-layer network.
This preferred embodiment make use of SDN technology, list node needed for this calculation task reality and network actual load is completed according to High-Performance Computing Cluster resource management and storage system metadata management, for different user, calculation task determine independently sub-computing cluster, again by SDN controller according to this cluster Component units plan real network path together with, and be distributed to each network node by stream sheet form, and when actual task completes, the Internet resources under all stream table control are reclaimed.Realize user network Dynamic Programming, network and the function such as memory access load balancing, automatic energy saving consumption reduction.Improve the utilization ratio of system resource, reduce unnecessary resource occupation, in resolution system, network and storage resources access divide uneven, and the problem that during access conflict, time delay is larger.
The High-Performance Computing Cluster computing system of this preferred embodiment application SDN can be built by following three parts, and particular content is as follows:
Part I, disposes High-Performance Computing Cluster computing system;
Part II, application SDN technology provides cluster management and cluster-based storage service function;
Part III, based on the High Performance Cluster System of application SDN technology, network enabled resource QOS, load balancing, the application function expansion Design and implementation such as energy-saving and cost-reducing.
In the Part I disposing High-Performance Computing Cluster computing system, system node is divided into access node, service node, meter to deposit node, computing node and switching node by the High-Performance Computing Cluster computing system of this preferred embodiment.Wherein, node deposited by this meter is the node that Direct Attached Storage equipment provides stores service function, can avoid expensive special purpose memory devices and storage networking, the distributed storage service that the isomery that provides support stores, certain conditions permit also directly can mount special purpose memory devices.Switching node can comprise SDN switch or special intelligent network adapter, exchanges contact by the directly actuated data transfer equipment of SDN controller.Switching node can be independent equipment, also can be the function of the realization in other nodes.
Service node is responsible for the functions such as access user management, cluster resource management, calculation task scheduling, distributed storage and SDN controller (also can be called OF controller).
Meter is deposited node and is born calculating and memory function simultaneously, can resident virtual switch, common virtual bridge, and intelligent network adapter also can be used as hardware-accelerated technology, in network edge isolation or be communicated with corresponding storage networking and computing network.
Switching node is the intermediary switch node of application SDN technology, in the design architecture of SDN control and forward separating, realize data retransmission function specially.According to the stream table that SDN controller on aforementioned service node issues, calculation task is born in computing node sole duty, can dispose GPU subcard and accelerate function as calculating.
Service node and meter deposit resident distributed file system or the service of object stores service on node, and the cloud storage cluster of termination special purpose memory devices or self formation, forms storage networking thereafter.Input/output information, the information such as computational process and result of calculation of high-performance calculation all leave in this storage system.Storage networking can adopt deployment-specific network, so that isolate with supervising the network, Unified Network also can be adopted to realize.
Computing network is physically separated substantially with storage networking.Certain this division is not absolute, when network condition limits, such as, do not configure special supervising the network, storage networking also can be born a small amount of management and control flow (the external interface function etc. such as external user Operation and Maintenance, external input and output).But mutual, the intermediate object program of node main in high-performance calculation is synchronous and export and all must concentrate in the computing network of sub-computing cluster.
There is provided in the Part II of cluster management and cluster-based storage service function in application SDN technology, application SDN technology provides cluster management and cluster-based storage service function.SDN technology comprises controller and switch.Wherein SDN switch can adopt three kinds of forms, and physical exchange, virtual switch and intelligent network adapter exchange.Physical exchange is hardware SDN switch, and virtual switch is the switch of server internal software simulation, can provide function of exchange for each virtual machine on server.Intelligent network adapter runs a processor on network interface card, message is submitted to each virtual machine in server again after directly carrying out procotol process after receiving carries out Business Processing or is forwarded directly to the equipment of other node without server process, and SDN controller is managed by the mode that stream table issues and controls the forwarding capability of respective switch.
Use SDN controller, and distributed file system metadata service, cluster resource management service associated treatment together.Wherein all list nodes and business relations in file system metadata service and cluster resource management service determination sub-cluster, be handed down to SDN controller.SDN controller realizes network resource scheduling function, formulates network path according to business relations, reduces unnecessary network traffics expense and equipment expenditure, avoids the formation of network loop, realizes dynamic network status monitoring, realizes load balancing.
SDN controller has the switch of functional cohesion to manage by stream table set pair is multiple.Stream table collection is aimed at a upper-layer functionality request, the set of all switch stream tables that controller is formulated, and stream shows the processing mode of collection needs according to affairs, all unified loading or all unification unloadings.Safeguard stream table collection with affairs numbered sequence, the corresponding corresponding stream table collection of affairs numbering, upper layer request practical function, the stream table collection affairs numbering of generation, must be safeguarded by upper strata oneself, with support necessary newly-increased, change, delete function.On controller, every bar stream table needs the stream table collection numbering recording its correspondence.
Cluster resource manages, and during service node distribution calculation task, by resource management controlling planning computing cluster, determines list node, forms the double layer network of high performance computing system internal independence.After calculating completes, delete related streams table collection.As required, cluster resource management can be the independent storage of sub-dispensed, and builds independent storage networking.Elastomeric network resource management can be realized according to traffic monitoring, when network load high or certain intennediate switching node, intermediary switch load height, again plan forward-path, dispose forwarding flow table.
Distributed file system, when determining in access with customer in response end file server requests, dynamically arranges stream table by controller, realizes the connection in planning network path, the stream table collection affairs numbering that record is relevant.During the unloading of file client, delete relevant stream table collection.During cluster resource management Resources allocation, determine by Metadata Service analysis the node storing data place according to user's input information.Pre-determine, Metadata Service needs first to analyze user profile or store path, if these storage data exist, returns the node at this storage data place, if there is no then needs Metadata Service dynamic assignment memory node.
At the High Performance Cluster System based on application SDN technology, in network enabled resource QOS, load balancing, the application function expansion Design and implementation Part III such as energy-saving and cost-reducing, controller scheduler module is by the mode of resource reservation, prioritization, the calculating of different user, storage networking resource are distributed, even fine-grained resource division is realized to the calculating of same user, storage resources.When SDN controller issues stream table, quota restrictions of port, stream table being correlated with issue in the lump.
Resource management increases node newly, metadata access increases service, exchanges and reports statistics to occur port congestion.In above-mentioned situation, controller starts reschedule flow process immediately, analyze current network conditions, existing access focus is swapped out to centre and analyzes, by part via network access path move to other load relative free in the middle of exchange.
When all node loads are in free time or low load condition for a long time in resource management discovery system, start energy-saving and cost-reducing function, allow electricity under idle computing node, new planning of laying equal stress on changes the network path related to.Idiographic flow is, locks the node by lower electricity, not for it distributes new calculation task; Stream is shown timely processing node port and is deleted in advance, no longer distributes new stream table function for this node place switching port.After the calculation task waiting on it completes, start lower electricity or dormancy flow process.Affect the communication of original network path, report MISS event to trigger stream table by switch and regenerate, and upgrade its stream table collection record.
Resource management finds that system overload exceedes certain hour, and there is the node of idle lower electricity or dormancy, start revolution flow process, first determine the most suitable node powered on, add resource management after having powered on, rerun by the newly-increased resource management mode in HPCC.
This preferred embodiment uses SDN method decouples computation and stores flow, uses intelligent network adapter and SDN switch to substitute conventional switch.Network function to be no longer that legacy network function is prior like that determine forwarding route, but need the list node increased to be advertised to SDN controller by resource management and file metadata service by determining, dynamically determine forward-path by controller according to the actual connection state of current network and load intensity, avoid network local congestion and focus to occur.Adjust unbalanced network condition as required simultaneously.Object be the network traffics of the dissimilar calculating in cluster, storage resource request are separated, unified planning forward-path dynamic equalization, Internet resources peak use rate in raising system.
Fig. 4 is the schematic network structure of the HPCC applying SDN technology according to the preferred embodiment of the invention, as shown in Figure 4, system node is divided into access node, service node, meter to deposit node, computing node by the High-Performance Computing Cluster computing system in this preferred embodiment.Specifically do not indicate switching node in Fig. 4, this switching node can be that independent node connects between the other nodes, also can be the function of exchange that other node contents realize.
Node deposited by meter is the node that Direct Attached Storage equipment provides stores service function, and can avoid expensive special purpose memory devices and storage networking, the distributed storage service of the isomery that provides support, certain conditions permit also directly can mount special purpose memory devices.Switching node comprises SDN switch, and special intelligent network adapter, by the directly actuated data transfer equipment of SDN controller.
Service node is responsible for the functions such as access user management, cluster resource management, calculation task scheduling, distributed storage and SDN controller.All SDN switch in SDN controller control system, comprise hardware-switch, virtual switch and intelligent network adapter and exchange, and to cluster resource management and storing metadata provisioning interface.
Meter is deposited node and is born calculating and memory function simultaneously, so need to access computing network and storage networking simultaneously.Therefore resident virtual switch, common virtual bridge, also can use intelligent network adapter as hardware-accelerated technology, for isolating at network edge or being communicated with corresponding storage networking and computing network.Service node and meter deposit resident distributed file system or the service of object stores service on node, and the cloud storage cluster of termination special purpose memory devices or self formation, forms storage networking thereafter.Input/output information, the information such as computational process and result of calculation of high-performance calculation all leave in this storage system.Storage networking can adopt deployment-specific network, so that isolate with supervising the network, Unified Network also can be adopted to realize.
Switching node is the intermediary switch node of application SDN technology, in the design architecture of SDN control and forward separating, realize data retransmission function specially.Data retransmission work born by the stream table issued according to SDN controller on aforementioned service node.
Calculation task is born in computing node sole duty, can dispose GPU subcard and accelerate function as calculating.
In this high performance computing system, cluster management and cluster-based storage service function adopt SDN technology to realize network.Existing SDN technology comprises controller and switch.Wherein SDN switch can adopt three kinds of forms, and physical exchange, virtual switch and intelligent network adapter exchange.Intelligent network adapter runs a processor on network interface card, message is submitted to each virtual machine in server again after directly carrying out procotol process after receiving carries out Business Processing or is forwarded directly to the equipment of other node without server process, and SDN controller is managed by the mode that stream table issues and controls the forwarding capability of respective switch.
SDN controller has the switch of functional cohesion to manage by stream table set pair is multiple.Be directed to a upper-layer functionality request, the set of all switch stream tables that controller is formulated is called stream table collection.Stream shows the processing mode of collection needs according to affairs, all unified loading or all unification unloadings.Safeguard stream table collection with affairs numbered sequence, the corresponding corresponding stream table collection of affairs numbering, upper layer request practical function, the stream table collection affairs numbering of generation, must be safeguarded by upper strata oneself, with support necessary newly-increased, change, delete function.On controller, every bar stream table needs the stream table collection numbering recording its correspondence.
Preferably, the application in cluster resource management, when referring to service node distribution calculation task, by resource management controlling planning computing cluster, determines list node, forms the double layer network of high performance computing system internal independence.After calculating completes, delete related streams table collection.As required, cluster resource management can be the independent storage of sub-dispensed, and builds independent storage networking.Elastomeric network resource management can be realized according to traffic monitoring, when network load high or certain intennediate switching node, intermediary switch load height, again plan forward-path, dispose forwarding flow table.
Preferably, the application in distributed file system, is mainly divided in access and determines and pre-determine two kinds of modes.Determine in access, during customer in response end file server requests, dynamically arrange stream table by controller, realize the connection in planning network path, the stream table collection affairs numbering that record is relevant.During the unloading of file client, delete relevant stream table collection.During cluster resource management Resources allocation, determine by Metadata Service analysis the node storing data place according to user's input information.Pre-determine, Metadata Service needs first to analyze user profile or store path, if these storage data exist, returns the node at this storage data place, if there is no then needs Metadata Service dynamic assignment memory node.
SDN exchanges three kinds of forms, and physical exchange, virtual switch and intelligent network adapter exchange, and physical aspect is different, but function is substantially identical.Describe for intelligent network adapter, the control of intelligent network adapter and forwarding capability, Fig. 5 is the schematic diagram of intelligent network adapter data retransmission according to the preferred embodiment of the invention, as shown in Figure 5, intelligent network adapter possesses multiple physical internet ports, under stream table (Flow Table) effect that SDN controller is issued by escape way, complete data retransmission function.Business datum carries out the stream table forward process of SDN exchange on intelligent network adapter after receiving, be forwarded to this locality or other network node after completing according to result.Such as file access request, concrete node location and the switching port corresponding relation of this file access request is comprised in the stream list processing strategy that Metadata Service is issued by controller, according to node location and path planning after the list processing of local exchange stream, by this file request forward to local or other node port.
Based on this high performance computing system, abundanter application function can be provided.Under the effect of SDN controller, following function support can be provided to system:
One, network enabled resource QOS accesses.Controller, by the mode of resource reservation, prioritization, distributes the calculating of different user, storage networking resource, even realizes fine-grained resource division to the calculating of same user, storage resources.When SDN controller issues stream table, port restriction is issued in the lump.By rational network scheduling mode, reduce the unbalanced and access exception of access.Restriction and reduce malicious user or abnormal user for the impact of whole Internet resources, makes the mistake unique user and is limited within rational scope.
Fig. 6 is the flow chart of Internet resources QOS access according to the preferred embodiment of the invention, and as shown in Figure 6, this flow chart comprises the steps.
Step S602, user applies for calculating, storage resources time option and installment Internet resources, reserve resource bandwidth sum priority;
Step S604, controller completes network-in-dialing path planning, after issuing stream table, by resource QOS configuration distributing to the middle crosspoint of each on forward-path (comprise physical switches and virtual switch, intelligent network adapter exchanges); Realize controller in step S604 and complete network-in-dialing path planning, with above-mentioned Fig. 1 step 102, plan that the function and efficacy of the communication path of the node in described cluster is identical.
Step S606, each crosspoint (crosspoint also can be called switching node) for resource reservation port or stream list item, carry out corresponding real-time traffic monitoring;
Step S608, crosspoint exceedes local transfer capability for network load, preferentially carries out Flow Control for low priority with without quota or above-quota port or stream list item, and reports controller to make subsequent treatment.
Its two, controller switching at runtime access path realizes load balancing.When occurring in system that resource management increases node newly, metadata access increases service, when exchange reports in statistics and occurs continuing the situations such as port congestion, controller starts reschedule flow process immediately.Analyze current network conditions, existing access focus is swapped out to centre and analyzes, by part via network access path move to other load relative free in the middle of exchange.
Fig. 7 is the flow chart realizing load balancing according to the preferred embodiment of the invention, and as shown in Figure 7, this flow chart comprises the steps.
Step S702, controller starts Internet resources reschedule flow process;
Step S704, controller, according to the network condition information of collecting, according to the situation of current residual Internet resources, to the link of access focus occurring busy inequality, adds the weight of real network load, recalculates most suitable forward-path;
Step S702 and step S704, recalculates most suitable forward-path, with above-mentioned Fig. 1 step 102, plans that the function and efficacy of the communication path of the node in described cluster is identical.
Step S706, determines the forward-path stream table collection needing change, with transactional manner, first cancels original stream table collection, then re-issue new stream table collection;
Step S708, crosspoint, for the stream table cancelled in change, needs to carry out local cache to former stream table matching message, has new stream table to issue, reopen buffer memory message repeating if follow-up, otherwise wait for a period of time and report controller or directly delete packet buffer, dropping packets.
Its three, support energy-saving and cost-reducing function.When all node loads are in free time or low load condition for a long time in resource management discovery system, start energy-saving and cost-reducing function, allow electricity under idle computing node, new planning of laying equal stress on changes the network path related to.Idiographic flow is, locks the node by lower electricity, not for it distributes new calculation task; Stream is shown timely processing node port and is deleted in advance, no longer distributes new stream table function for this node place switching port.After the calculation task waiting on it completes, start lower electricity or dormancy flow process.Affect the communication of original network path, report MISS event to trigger stream table by switch and regenerate, and upgrade its stream table collection record.
Fig. 8 is the flow chart realizing according to the preferred embodiment of the invention saving energy and reduce the cost, and as shown in Figure 8, this flow chart comprises the steps.
Step S802, system starts energy-saving and cost-reducing scheduling flow;
Step S804, system resource management calculates current system load, determines the list node needing lower electricity, gives controller and performs.Be pre-deletion by these vertex ticks in resource management simultaneously, new calculation task of not reallocating;
Step S806, controller determines that this lower film rings the forward-path stream table collection needing change, deletes equally first in advance, no longer bears new forwarding flow table, also needs the network condition after according to pre-deletion to recalculate forward-path stream table collection;
Step S808, controller, with transactional manner, to the all-network link rung by this lower film, is first cancelled original stream table collection, is then re-issued new stream table collection;
Step S810, controller advertisement resource manages, electricity under start node.Note depositing the order of node and switching node according to computing node, meter, guarantee that a category node completes and now bear function, lower electricity and after completing, restart electricity under Equations of The Second Kind node, prevent network interruption from forming system isolated island.
Resource management finds that system overload exceedes certain hour, and there is the node of idle lower electricity or dormancy, start revolution flow process, first determine the most suitable node powered on, add resource management after having powered on, rerun by the increase resource management mode in HPCC.
This preferred embodiment uses SDN method decouples computation and storage networking, realizes user network Dynamic Programming, network and the function such as memory access load balancing, automatic energy saving consumption reduction.Improve the utilization ratio of system resource, reduce unnecessary resource occupation, in resolution system, network and storage resources access divide uneven, and the problem that during access conflict, time delay is larger.
Obviously, those skilled in the art should be understood that, above-mentioned of the present invention each module or each step can realize with general calculation element, they can concentrate on single calculation element, or be distributed on network that multiple calculation element forms, alternatively, they can realize with the executable program code of calculation element, thus, they can be stored and be performed by calculation element in the storage device, or they are made into each integrated circuit modules respectively, or the multiple module in them or step are made into single integrated circuit module to realize.Like this, the present invention is not restricted to any specific hardware and software combination.
The foregoing is only the preferred embodiments of the present invention, be not limited to the present invention, for a person skilled in the art, the present invention can have various modifications and variations.Within the spirit and principles in the present invention all, any amendment done, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.
Claims (20)
1. a task processing method, is characterized in that comprising:
Plan the communication path of the node in described cluster, wherein, described cluster is that described cluster comprises multiple node at least according to the cluster for performing described task that task is determined;
The information of described communication path sent to the node bearing function of exchange in described cluster to build described cluster network, wherein, the node bearing function of exchange described in is the node for the information forwarding data according to the described communication path received in described multiple node.
2. method according to claim 1, is characterized in that, at least determines that the cluster performing described task comprises according to task:
At least determine according to the load of described task and network the cluster performing described task.
3. method according to claim 1, is characterized in that, after the described cluster network of structure, described method also comprises:
Again the communication path of the node in described cluster is planned according to the situation change of described cluster;
The information of the communication path again planned is sent to the node bearing function of exchange in described cluster.
4. method according to claim 3, is characterized in that, after again planning the communication path of the node in described cluster according to the situation change of described cluster, described method also comprises:
Send the instruction cancelling the information of former communication path, wherein, the described node bearing function of exchange carries out buffer memory according to described instruction for the message of the information matches of the former communication path of cancelling, and after the information receiving the communication path again planned, reopens the forwarding of buffer memory message; The described node bearing function of exchange, when not receiving the information of the communication path again planned, is deleted the message of buffer memory or reports the information not receiving the communication path again planned.
5. method according to claim 3, is characterized in that, the communication path of the node again planned in described cluster according to the change of the situation of described cluster comprise following one of at least:
Determine the load of the node in described cluster, and the load of the node of described cluster is planned the communication path of the node in described cluster again as weights;
One or more node needing lower electricity or power on is determined according to the load of the node in described cluster; Again communication path is planned according to the node in lower electricity or the described cluster after powering on.
6. method according to claim 1, is characterized in that, described method also comprises:
The described node bearing function of exchange determines that the network load of self exceedes threshold value, then carry out flow control according to the priority of task.
7. method according to any one of claim 1 to 6, is characterized in that, the node in described cluster comprise in Types Below one of at least: multiple computing node, multiple memory node, multiplely deposit node, multiple switching node, wherein,
Described computing node for performing calculation function, or, for performing calculation function and bear function of exchange;
Described memory node is for performing data storage function;
Described meter deposits node for performing calculation function and data storage function, or, for performing calculation function, data storage function and bearing function of exchange.
8. according to the method described in claim 7, it is characterized in that, node realizes described switching node function by intelligent network adapter and/or virtual switch deposited by described meter.
9. method according to any one of claim 1 to 8, is characterized in that, the mode that the information of described communication path is shown by stream sends.
10. according to the method described in claim 9, it is characterized in that, described method also comprises: after task completes, and reclaims the list-directed Internet resources of described stream.
11. 1 kinds of Task Processing Units, is characterized in that comprising:
Planning module, for planning the communication path of the node in described cluster, wherein, described cluster is that described cluster comprises multiple node at least according to the cluster for performing described task that task is determined;
Sending module, for the information of described communication path being sent to the node bearing function of exchange in described cluster to build described cluster network, wherein, the node bearing function of exchange described in is the node for the information forwarding data according to the described communication path received in described multiple node.
12. devices according to claim 11, is characterized in that, described planning module is used at least determining according to the load of described task and network the cluster performing described task.
13. devices according to claim 11, is characterized in that,
Described planning module, also for again planning the communication path of the node in described cluster according to the situation change of described cluster;
Described sending module, also for the information of the communication path again planned being sent to the node bearing function of exchange in described cluster.
14. devices according to claim 13, it is characterized in that, described sending module is also for sending the instruction of the information cancelling former communication path, wherein, the described node bearing function of exchange carries out buffer memory according to described instruction for the message of the information matches of the former communication path of cancelling, after the information receiving the communication path again planned, reopen the forwarding of buffer memory message; And described in bear the node of function of exchange when not receiving the information of the communication path again planned, delete buffer memory message or report the information not receiving the communication path again planned.
15. devices according to claim 13, is characterized in that,
The load of the node of described cluster also for determining the load of the node in described cluster, and is planned the communication path of the node in described cluster by described planning module again as weights;
Described planning module is also for determining one or more node needing lower electricity or power on according to the load of the node in described cluster; Again communication path is planned according to the node in lower electricity or the described cluster after powering on.
16. devices according to claim 11, is characterized in that, comprising:
The described node bearing function of exchange for when determining that the network load of self exceedes threshold value, then carries out flow control according to the priority of task.
17., according to claim 11 to the device according to any one of 16, is characterized in that, the node in described cluster comprise in Types Below one of at least: multiple computing node, multiple memory node, multiplely deposit node, multiple switching node, wherein,
Described computing node for performing calculation function, or, for performing calculation function and bear function of exchange;
Described memory node is for performing data storage function;
Described meter deposits node for performing calculation function and data storage function, or, for performing calculation function, data storage function and bearing function of exchange.
18., according to the device described in claim 17, is characterized in that, node realizes described switching node function by intelligent network adapter and/or virtual switch deposited by described meter.
19., according to claim 11 to the device according to any one of 18, is characterized in that, the mode that the information of described communication path is shown by stream sends.
20., according to the device described in claim 19, is characterized in that, described device also comprises: recycling module, for after task completes, reclaim the list-directed Internet resources of described stream.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410006581.XA CN104767778A (en) | 2014-01-07 | 2014-01-07 | Task processing method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410006581.XA CN104767778A (en) | 2014-01-07 | 2014-01-07 | Task processing method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN104767778A true CN104767778A (en) | 2015-07-08 |
Family
ID=53649381
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410006581.XA Withdrawn CN104767778A (en) | 2014-01-07 | 2014-01-07 | Task processing method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104767778A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106209478A (en) * | 2016-08-31 | 2016-12-07 | 上海地面通信息网络有限公司 | A kind of publicly-owned cloud network architecture system based on OpenStack |
| WO2017113929A1 (en) * | 2015-12-29 | 2017-07-06 | 中兴通讯股份有限公司 | Method and controller for controlling forwarding device cluster in software-defined networking |
| CN108287669A (en) * | 2018-01-26 | 2018-07-17 | 平安科技（深圳）有限公司 | Date storage method, device and storage medium |
| CN109656911A (en) * | 2018-12-11 | 2019-04-19 | 江苏瑞中数据股份有限公司 | Distributed variable-frequencypump Database Systems and its data processing method |
| CN109803024A (en) * | 2019-01-28 | 2019-05-24 | 北京中科晶上科技股份有限公司 | A method of for cluster node network |
| CN112398905A (en) * | 2020-09-28 | 2021-02-23 | 联想(北京)有限公司 | Node and information synchronization method |
| CN112418599A (en) * | 2020-10-15 | 2021-02-26 | 重庆市科学技术研究院 | Enterprise growth path planning method and system based on index set |
| CN115904738A (en) * | 2023-01-05 | 2023-04-04 | 摩尔线程智能科技(北京)有限责任公司 | Management system and control method for data processing device cluster |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102104542A (en) * | 2011-01-14 | 2011-06-22 | 中国人民解放军信息工程大学 | Method for realizing service group router under forwarding and controlling separated network architecture |
| CN102404385A (en) * | 2011-10-25 | 2012-04-04 | 华中科技大学 | Virtual cluster deployment system and deployment method for high performance computing |
| CN102571568A (en) * | 2011-12-27 | 2012-07-11 | 北京奇虎科技有限公司 | Method and device for processing task |
| CN103368864A (en) * | 2013-07-31 | 2013-10-23 | 北京华易互动科技有限公司 | Intelligent load balancing method based on c/s (Client/Server) architecture |
| US8949410B2 (en) * | 2010-09-10 | 2015-02-03 | Cisco Technology, Inc. | Server load balancer scaling for virtual servers |
-
2014
- 2014-01-07 CN CN201410006581.XA patent/CN104767778A/en not_active Withdrawn
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8949410B2 (en) * | 2010-09-10 | 2015-02-03 | Cisco Technology, Inc. | Server load balancer scaling for virtual servers |
| CN102104542A (en) * | 2011-01-14 | 2011-06-22 | 中国人民解放军信息工程大学 | Method for realizing service group router under forwarding and controlling separated network architecture |
| CN102404385A (en) * | 2011-10-25 | 2012-04-04 | 华中科技大学 | Virtual cluster deployment system and deployment method for high performance computing |
| CN102571568A (en) * | 2011-12-27 | 2012-07-11 | 北京奇虎科技有限公司 | Method and device for processing task |
| CN103368864A (en) * | 2013-07-31 | 2013-10-23 | 北京华易互动科技有限公司 | Intelligent load balancing method based on c/s (Client/Server) architecture |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017113929A1 (en) * | 2015-12-29 | 2017-07-06 | 中兴通讯股份有限公司 | Method and controller for controlling forwarding device cluster in software-defined networking |
| CN106209478A (en) * | 2016-08-31 | 2016-12-07 | 上海地面通信息网络有限公司 | A kind of publicly-owned cloud network architecture system based on OpenStack |
| CN108287669A (en) * | 2018-01-26 | 2018-07-17 | 平安科技（深圳）有限公司 | Date storage method, device and storage medium |
| CN108287669B (en) * | 2018-01-26 | 2019-11-12 | 平安科技（深圳）有限公司 | Date storage method, device and storage medium |
| CN109656911A (en) * | 2018-12-11 | 2019-04-19 | 江苏瑞中数据股份有限公司 | Distributed variable-frequencypump Database Systems and its data processing method |
| CN109803024A (en) * | 2019-01-28 | 2019-05-24 | 北京中科晶上科技股份有限公司 | A method of for cluster node network |
| CN112398905A (en) * | 2020-09-28 | 2021-02-23 | 联想(北京)有限公司 | Node and information synchronization method |
| CN112398905B (en) * | 2020-09-28 | 2022-05-31 | 联想(北京)有限公司 | Node and information synchronization method |
| CN112418599A (en) * | 2020-10-15 | 2021-02-26 | 重庆市科学技术研究院 | Enterprise growth path planning method and system based on index set |
| CN112418599B (en) * | 2020-10-15 | 2023-02-10 | 重庆市科学技术研究院 | Enterprise growth path planning method and system based on index set |
| CN115904738A (en) * | 2023-01-05 | 2023-04-04 | 摩尔线程智能科技(北京)有限责任公司 | Management system and control method for data processing device cluster |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104767778A (en) | Task processing method and device | |
| CN111953526B (en) | Hierarchical computational power network arrangement method, device and storage medium | |
| Dong et al. | Energy-saving virtual machine placement in cloud data centers | |
| Jin et al. | Joint host-network optimization for energy-efficient data center networking | |
| CN105721300B (en) | The method and apparatus of management is searched for network equipment stream | |
| CN106464530B (en) | The method and system for the dynamic dispatching that network updates | |
| CN103491129B (en) | A kind of service node collocation method, pool of service nodes Register and system | |
| CN101207550B (en) | Load balancing system and method for multi business to implement load balancing | |
| CN107819662B (en) | A kind of service function chain real-time migration method calculating network applied to cloud and mist | |
| CN108540876A (en) | Service path choosing method, SDN controllers, storage medium and electronic equipment | |
| CN104503826B (en) | The virtual machine mapping method and device of a kind of cloud computation data center | |
| CN104601680B (en) | A kind of method for managing resource and device | |
| CN110087250B (en) | Network slice arranging scheme and method based on multi-objective joint optimization model | |
| CN104202183B (en) | The method and apparatus that a kind of solution SDN stream ranks configuration conformance updates | |
| CN105610715B (en) | A kind of cloud data center multi-dummy machine migration scheduling method of planning based on SDN | |
| CN103997513A (en) | Programmable virtual network service system | |
| CN102164184A (en) | Computer entity access and management method for cloud computing network and cloud computing network | |
| CN106936705B (en) | Software defined network routing method | |
| Dong et al. | Virtual machine placement for improving energy efficiency and network performance in iaas cloud | |
| CN101753405A (en) | Cluster server memory management method and system | |
| CN105391651A (en) | Virtual optical network multilayer resource convergence method and system | |
| CN102104542A (en) | Method for realizing service group router under forwarding and controlling separated network architecture | |
| CN102917287A (en) | Intelligent optical network exchange device and edge cashing method facing content center | |
| Liu et al. | PSNet: Reconfigurable network topology design for accelerating parameter server architecture based distributed machine learning | |
| Yi et al. | Energy‐aware disaster backup among cloud datacenters using multiobjective reinforcement learning in software defined network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20150708 |
|
| WW01 | Invention patent application withdrawn after publication |