Summary of the invention
Based on this, be necessary to provide a kind of method for scheduling task that can reduce the supercomputer of energy consumption.
In addition, also be necessary to provide a kind of task scheduling system that can reduce the supercomputer of energy consumption.
The method for scheduling task of described supercomputer comprises: the A. computing node is consulted the application load table, judges whether computing node can finish the parallel section of the task of waiting for operation, if then request moves the parallel section of described task; B. the dispatching center receives the operation request of computing node, obtains the energy consumption that parallel section increased that computing node is finished described task; C. the computing node of the energy consumption minimum of described increase is selected by described dispatching center, and the parallel section of described task is dispatched to this computing node.
Preferably, the content of described application load table record comprises: the computing power that the operation address of the running status of task, task, the parallel section of operation task are required and the required energy consumption of parallel section of operation task.
Judge in the described steps A computing node can finish the task of waiting for operation parallel section step specifically: computing node is consulted the application load table, obtain the required computing power of parallel section of finishing the task of waiting for operation, when the residue computing power of described computing node self during, judge that then described computing node can finish the parallel section of the task of described wait operation greater than the required computing power of described parallel section.
Also comprise after the described step C: D. moves the running status of the parallel section of described task by waiting for that operation is updated to, and the operation address of upgrading the parallel section of described task.
In addition, described method also comprises: the E. computing node is consulted the application load table, judges whether computing node can finish the parallel section of moving of task, if then request moves the parallel section of described moving of task; F. the dispatching center receives the operation request of computing node, obtains the parallel section of described moving of task is moved to the energy consumption that parallel section increased that computing node and this computing node move this task; G. the application load table is consulted by described dispatching center, obtain the required energy consumption of parallel section of the described task of operation, the required energy of the parallel section of the described task of operation that writes down in less than described application load table when the energy consumption of described increase is consuming time, then the parallel section of described moving of task is moved to described computing node.
Preferably, also comprise after the described step G: upgrade the operation address of parallel section of described task and the required energy consumption of parallel section of upgrading the described task of operation.
The task scheduling system of described supercomputer comprises: at least one computing node, described computing node is used to consult the application load table, judge whether computing node can finish the parallel section of the task of waiting for operation, if then request moves the parallel section of described task; The dispatching center, carry out alternately with described computing node, be used to receive the operation request of computing node, obtain computing node and finish the energy consumption that parallel section increased of described task, and the computing node of selecting the energy consumption minimum of described increase, and the parallel section of the task of described wait operation is dispatched to this computing node.
Preferably, described computing node also is used to consult the application load table, judges whether computing node can finish the parallel section of moving of task, if then request moves the parallel section of described task.
Further preferably, described dispatching center also is used to receive the operation request of computing node, the parallel section that obtains will moving of task is moved to the energy consumption that parallel section increased that computing node and this computing node move described task, and consult the application load table, obtain the required energy consumption of parallel section of the described task of operation, the required energy of the parallel section of the described task of operation that writes down in less than described application load table when the energy consumption of described increase is consuming time, then the parallel section of described task is moved to described computing node.
In the method for scheduling task and system of above-mentioned supercomputer, consult the application load table by computing node, when computing node self can be finished the parallel section of the task of waiting for operation, then active request is moved this parallel section, and obtaining each computing node, the dispatching center finishes the energy consumption that this parallel section increases, select the computing node of the energy consumption minimum increased, and with task scheduling to this computing node.Because computing node is according to self-ability active request operation task, and the computing node operation task of energy consumption minimum is selected by the dispatching center, like this, can effectively reduce the energy consumption of supercomputer.
Embodiment
Fig. 1 shows the method for scheduling task flow process of supercomputer among the embodiment, and detailed process is as follows:
In step S101, computing node is consulted the application load table, judges when computing node can be finished the parallel section of the task of waiting for operation, then the parallel section of the described task of request operation.
In step S102, the dispatching center receives the operation request of computing node, obtains the energy consumption that parallel section increased that computing node is finished described task.
In step S103, the computing node of the energy consumption minimum of described increase is selected by the dispatching center, and the parallel section of described task is dispatched to this computing node.
In one embodiment, the application load table can be stored on the computing node, consults for dispatching center and each computing node.The recordable content of application load table comprises: energy consumption that the computing power that the operation address of the running status of task, task, each parallel section of operation task are required and each parallel section of operation task are required etc.Wherein, the running status of task comprises: wait moves, moves and operation is finished; The computing node address of each parallel section of this task is promptly moved in the operation address of task.
In one embodiment, judge computing node can finish the task of waiting for operation parallel section process specifically: computing node is consulted the application load table, obtain the required computing power of parallel section of finishing the task of waiting for operation, when the residue computing power of computing node self during greater than the required computing power of described parallel section, then this computing node of Pan Duaning can be finished the parallel section of the task of described wait operation.The computing power here is meant central processing power (specifically can be flops), memory power (specifically can be memory size) and the network capabilities (specifically can be the network bandwidth) etc. of computing node.
In one embodiment, task scheduling behind the computing node that increases the energy consumption minimum, is upgraded the application load table, specifically: the running status of the parallel section of described task is moved by waiting for that operation is updated to; Upgrade the operation address (promptly moving the address of the computing node of this parallel section) of the parallel section of described task.Because all computing nodes can both initiatively participate in the process of dispatching, when computing node is judged a certain parallel section that self can finish the work, then active request is moved this parallel section, calculating it simultaneously finishes the energy consumption that this parallel section increases and offers the dispatching center, the minimum computing node of energy consumption is then selected by the dispatching center, make task scheduling more reasonable, fair and quick like this, can effectively reduce the energy consumption of supercomputer, also can improve the operational efficiency of supercomputer.
Fig. 2 shows the method for scheduling task flow process of supercomputer among another embodiment, and this flow process has been described the process of the parallel section of the task that computing node request operation moving, and detailed process is as follows:
In step S201, computing node is consulted the application load table, judges when computing node can be finished the parallel section of moving of task, then the parallel section of the described task of request operation.
In step S202, the dispatching center receives the operation request of computing node, obtains the energy consumption that parallel section increased of the parallel section of this moving of task being moved to computing node and this computing node operation task.
In step S203, the application load table is consulted by the dispatching center, obtains the required energy consumption of parallel section of the described task of operation.This energy consumption is promptly just at the energy that computing node consumed of the parallel section of operation task.
In step S204, whether the energy consumption of judging described increase is less than the required energy consumption of parallel section of the described task of operation that writes down in the application load table, if then enter step S205, otherwise finish.
In step S205, the parallel section of described moving of task is moved to described computing node.
In step S206, upgrade the operation address of parallel section of described task and the required energy consumption of parallel section of upgrading the described task of operation.In one embodiment, finish the energy consumption that a certain parallel section increased of moving of task and this parallel section is moved to the summation of the energy consumption that this computing node increases consuming time less than the energy of the computing node that moves this parallel section when computing node, then computing node can be moved to this parallel section of task in the dispatching center.And the operation address of the parallel section of this task that writes down in the application load table also will be updated to the address of new computing node by the address of former computing node, and the required energy consumption of parallel section of this task of operation of record also will be changed to the energy consumption that new computing node moves this parallel section by the energy consumption of former computing node.
Fig. 3 shows the task scheduling system of supercomputer among the embodiment, this system call center 100 and carry out at least one mutual computing node (computing node 200,300......N) with it.Wherein:
Each computing node (computing node 200,300......N) is used to consult the application load table, judges whether computing node finishes the parallel section of the task of waiting for operation, if then request moves the parallel section of described task.
Dispatching center 100 is used to receive the operation request of each computing node, obtain computing node and finish the energy consumption that parallel section increased of described task, and the computing node of selecting the energy consumption minimum of described increase, and the parallel section of the task of described wait operation is dispatched to this computing node.
In one embodiment, the application load table can be stored on the computing node, consults for dispatching center and each computing node.The recordable content of application load table comprises: energy consumption that the computing power that the operation address of the running status of task, task, each parallel section of operation task are required and each parallel section of operation task are required etc.Wherein, the running status of task comprises: wait moves, moves and operation is finished; The address at computing node place of the parallel section of this task is promptly moved in the operation address of task.
In one embodiment, each computing node (computing node 200,300......N) also can be used for consulting the application load table, judge whether it can finish the parallel section of moving of task, if then request moves the parallel section of this moving of task.The 100 operation requests that receive computing node in dispatching center, the parallel section that obtains will moving of task is moved to the energy consumption that parallel section increased that computing node and this computing node move described task, simultaneously, the application load table is consulted by dispatching center 100, obtain the required energy consumption of parallel section of this task of operation that writes down in the application load table, the required energy of the parallel section of the operation task that writes down in less than the application load table when the energy consumption of above-mentioned increase is consuming time, and then the parallel section of the task that this is being moved is moved to this computing node.
The above only is preferred embodiment of the present invention, not in order to restriction the present invention, all any modifications of being done within the spirit and principles in the present invention, is equal to and replaces and improvement etc., all should be included within protection scope of the present invention.