FR3091770A1 - Method for deploying a task in a supercomputer, method for implementing a task in a supercomputer, corresponding computer program and supercomputer - Google Patents

Abstract

This method of deploying a task involves: allocating nodes (1 ... 16) to the task; the determination, in the network (110), of a sub-network, of interconnection of the allocated nodes, satisfying one or more predefined determination criteria comprising a first criterion according to which the determined sub-network is that, among at least two subnets satisfying the criteria other than the first criterion, using the most switches already allocated each to at least one task already deployed; the allocation of the subnetwork, and in particular the links belonging to this subnetwork, to the task; and the implementation of inter-node communication routes in the allocated subnet Figure for the abstract: Fig. 1

Description

Method for deploying a task in a supercomputer, method for implementing a task in a supercomputer, corresponding computer program and supercomputer

The present invention relates to a method for deploying a task in a supercomputer, a method for implementing a task in a supercomputer, a corresponding computer program and a supercomputer.

The invention applies more particularly to a method for deploying a task in a cluster of computers forming a supercomputer comprising:

- Computers each having at least one network interface forming a node; and

- an interconnection network of nodes, comprising:

-- switches each having network interfaces, and

-- links each connecting either a node and a network interface of a switch, or two network interfaces of two respective switches;

the process comprising:

- the allocation of nodes to the task.

The article by J. Domke and T. Hoefler entitled "Scheduling-Aware Routing for Supercomputers" and published in SC '16: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, Salt Lake City, UT, 2016, p. 142-153, doi:10.1109/SC.2016.12 describes a method using a database with the list of tasks running on the supercomputer, the routing recalculating the routes according to their use in order to reduce congestion on the network.

The article by Matthieu Perotin and Tom Cornebize entitled “Isolating Jobs for Security on High-Performance Fabrics” published in High-Performance Interconnection Networks in the Exascale and Big-Data Era (HiPINEB) 2017, IEEE 3rd International Workshop on, pp. 49-56, 2017, describes a method in which tasks are isolated from each other (in the sense that they cannot communicate with each other) by deleting rows in the routing tables, so that there is no There is no route between two nodes allocated respectively to two different tasks.

It may thus be desirable to provide a method for deploying a task in a cluster of computers forming a supercomputer, which makes it possible to reduce the energy consumption of the supercomputer.

The subject of the invention is therefore a method for deploying a task in a cluster of computers forming a supercomputer comprising:

- Computers each having at least one network interface forming a node; and

- an interconnection network of nodes, comprising:

-- switches each having network interfaces, and

-- links each connecting either a node and a network interface of a switch, or two network interfaces of two respective switches;

the process comprising:

- the allocation of nodes to the task;

the method being characterized in that it further comprises:

- the determination, in the network, of a sub-network, of interconnection of the allocated nodes, satisfying one or more predefined determination criteria comprising a first criterion according to which the determined sub-network is the one, among at least two sub- networks satisfying the criterion or criteria other than the first criterion, using the most switches already allocated each to at least one task already deployed;

- the allocation of the sub-network, and in particular of the links belonging to this sub-network, to the task; and

- the implementation of inter-node communication routes in the allocated subnet.

Thus, thanks to the invention, fewer switches are used, which makes it possible to turn off, or else place in a standby state, a large number of unused switches. Despite the overconsumption and the slight latency time due to switching on or waking up a switch, it has been found that, on average, the method of the invention could allow an energy saving without significantly delaying the task deployment.

Also optionally, the determination of the sub-network comprises the determination of the at least two sub-networks satisfying the criterion or criteria other than the second criterion, then the selection among these at least two sub-networks of the one using the most switches already allocated each to at least one task already deployed.

Also optionally, the criterion or criteria for determining the sub-network further comprise a second criterion according to which the sub-network uses only links which are not allocated to any other task already deployed or which are allocated to less than N other tasks already deployed, N being a predefined number of one or more.

Thus, the number of tasks sharing each link is limited, so that each link is unlikely to be congested. Thus, each task has the impression of being alone (or almost) using the interconnection network. Thus, in particular, when a task is executed several times, the execution environment varies little, since it is little affected by the other tasks being executed.

Optionally, the predefined number N is equal to one.

A method for implementing a task is also proposed, comprising:

- the deployment of the task in a supercomputer, in accordance with a deployment method according to the invention;

- the execution of the task; and

- at the end of the execution of the task:

-- the detection of at least one link which is not allocated to any task, and

the placement of each detected link in an inactive state, in which the link requires an energy consumption lower than an energy consumption required by a link associated with at least one task.

Thus, it is possible to reduce the energy consumption of the supercomputer, without the need to perform throughput measurements in the links to deactivate those whose throughput is zero.

Optionally, the method comprises, to place a link in the inactive state, the extinction of the two network interfaces that this link connects so that no current or light travels through this link.

Also optionally, the method comprises, to place a link in the inactive state, the reduction of an operating frequency of at least one of the two network interfaces that this link connects.

There is also proposed a computer program downloadable from a communication network and/or recorded on a computer-readable medium and/or executable by a processor, characterized in that it comprises instructions for the execution of the steps of a method according to the invention, when said program is executed on one or more computers.

It is also proposed a cluster of computers forming a supercomputer comprising:

- Computers each having at least one network interface forming a node;

- an interconnection network of nodes, comprising:

-- switches each having network interfaces, and

-- links each connecting either a node and a network interface of a switch, or two network interfaces of two respective switches;

the supercomputer further comprising software and/or hardware means for deploying a task, designed to implement the following step:

- the allocation of nodes to the task;

the supercomputer being characterized in that it further comprises software and/or hardware means for deploying a task, designed to implement the following steps:

- the determination, in the network, of a sub-network, of interconnection of the allocated nodes, satisfying one or more predefined determination criteria comprising a first criterion according to which the determined sub-network is the one, among at least two sub- networks satisfying the criterion or criteria other than the first criterion, using the most switches already allocated each to at least one task already deployed;

- the allocation of the sub-network, and in particular of the links belonging to this sub-network, to the task; and

- the implementation of inter-node communication routes in the allocated subnet.

The invention will be better understood with the aid of the following description, given solely by way of example and made with reference to the appended drawings in which:

FIG. 1 schematically represents the general structure of a cluster of computers forming a supercomputer, according to one embodiment of the invention,

FIG. 2 schematically represents a switch of the supercomputer of FIG. 1, as well as the environment of this switch,

FIG. 3 illustrates the successive steps of a method for implementing a task in the supercomputer of FIG. 1, according to one embodiment of the invention,

FIG. 4 illustrates an interconnection sub-network of nodes of the supercomputer of FIG. 1 allocated to a first task,

FIG. 5 illustrates a potential sub-network for interconnecting nodes of the supercomputer of FIG. 1 allocated to a second task,

FIG. 6 illustrates a potential sub-network for interconnecting nodes of the supercomputer of FIG. 1 allocated to a second task,

FIG. 7 illustrates a potential sub-network for interconnecting nodes of the supercomputer of FIG. 1 allocated to a second task,

FIG. 8 illustrates a potential sub-network for interconnecting nodes of the supercomputer of FIG. 1 allocated to a second task,

FIG. 9 illustrates the two interconnection sub-networks of nodes of the supercomputer of FIG. 1 allocated respectively to the first and second tasks,

FIG. 10 illustrates a potential sub-network for interconnecting nodes of the supercomputer of FIG. 1 allocated to a third task,

FIG. 12 illustrates a potential sub-network for interconnecting nodes of the supercomputer of FIG. 1 allocated to a third task,

FIG. 12 illustrates a potential sub-network for interconnecting nodes of the supercomputer of FIG. 1 allocated to a third task,

FIG. 13 illustrates the three interconnecting sub-networks of nodes of the supercomputer of FIG. 1 allocated respectively to the first, second and third tasks, and

figure 14 illustrates links released following the stopping of the execution of the second task, in the supercomputer of figure 1.

With reference to FIG. 1, a computer cluster 100 forming a supercomputer, implementing the invention will now be described.

The cluster of computers 100 brings together independent computers 102, 104 appearing from the outside as a single computer with very high computing power, called a high performance computing computer or HPC computer (from the English "High Performance Computing" ) or a supercomputer. Each computer 102, 104 comprises, as is known per se, a central processing unit, a main memory, in which instructions for the processing unit central are intended to be registered and at least one network interface.

The network interfaces respectively form nodes 1…16 of the cluster of computers 100. Several network interfaces (and therefore several nodes 1…16) can belong to the same computer 102, 104, or else, as in the example described, each computer 102, 104 may have only one network interface forming one of the nodes 1...16.

Still in the example described, the computers 102, 104 of the cluster of computers 100 comprise calculation computers 102 and storage computers 104, the latter comprising mass memories such as hard disks for recording data used and/or produced by the computing computers 102.

Still in the example described, the network interfaces of the storage computers 104 are optical network interfaces, while the network interfaces of the calculation computers 102 are electrical network interfaces having a lower bit rate than the optical network interfaces. Indeed, the storage computers 104 generally exchange more data than the calculation computers 102.

The cluster of computers 100 further comprises a network 110 for interconnecting the nodes 1...16.

The network 110 comprises, on the one hand, switches 1.1...4.8 each having at least one network interface and, on the other hand, communication links 114 each connecting either a node 1...16 and a network interface of a 1.1…1.8 switch, i.e. two network interfaces of two respective 1.1…4.8 switches.

The computer cluster 100 has a certain topology defining the relative arrangement of nodes 1...16, switches 1.1...4.8 and communication links 114. In the example described, the topology is that of a large generalized parallel tree (“Parallel Generalized Fat Tree” or PGFT in English). The topology of a PGFT is generally defined by the following nomenclature: PGFT(h; m_1…m_h; w_1…w_h; p_1…p_h), where h is the number of levels between which the switches are distributed, m_n is the number of level n-1 switches (or nodes for the first level) connected to each level n switch, w_n is the number of level n switches connected to each level n-1 switch (or to each node for the first level) and p_n is the number of parallel communication links used between levels n and n-1 (or else between level n and the nodes for the first level). In the example described, the topology of the computer cluster 100 is PGFT(4; 2, 2, 2, 2; 1, 2, 2, 2; 1, 1, 1, 1). Thus, in the example described the first level switches are switches 1.1…1.8, the second level switches are switches 2.1…2.8, etc.

Each switch 1.1…4.8 comprises connection ports present in the network interface(s) of this switch 1.1…4.8, to which the communication links 114 are respectively connected. In the example described, the communication links 114 are all bidirectional, so that all ports are both incoming and outgoing ports.

Furthermore, in the example described, each switch 1.1…4.8 comprises local routing tables respectively associated with its incoming ports. Each local routing table associates an outgoing port of the switch with each destination of a message arriving by the incoming port considered.

In addition, each switch 1.1…4.8 includes default local routing tables. In the example described, the default routing tables are for example previously determined by the d-mod-k algorithm or else by the s-mod-k algorithm.

For example, switch 1.3 is illustrated in Figure 2 with the four communication links 114 connecting it to its neighboring nodes 5 and 6 and to its neighboring switches 2.3 and 2.4. The switch 1.3 thus comprises the four ports: P_5, P_6, P_2.3 and P_2.4, which are both incoming and outgoing ports. An example of the default local routing table of switch 1.3 port P_6, obtained by the d-mod-k algorithm, is shown in Table 1 below.

Local routing table by default of port P_ 6 of the switch 1.3 Destination nodes output port 1 P_2.3 2 P_2.4 3 P_2.3 4 P_2.4 5 P_5 6 P_6 7 P_2.3 8 P_2.4 9 P_2.3 10 P_2.4 11 P_2.3 12 P_2.4 13 P_2.3 14 P_2.4 15 P_2.3 16 P_2.4

Returning to FIG. 1, the cluster of computers 100 further comprises an administration server 112. The administration server 112 is designed in particular to communicate with clients wishing to use the cluster of computers 100, the latter being then seen by customers as a unique machine.

The administration server 112 is designed to implement a task implementation method, as will be described later. For this, the administration server 112 comprises software and/or hardware means, such as for example a resource manager (such as the “slurm” software) and a routing establishment module, to carry out the steps which will be described by the following. In the example described, the administration server 112 comprises, as is known per se, a central processing unit and a main memory in which instructions for the central processing unit are intended to be recorded. Thus, in the example described, the means are software means, in the form of a computer program downloadable from a communication network and/or recorded on a computer-readable medium and/or executable by a processor, comprising instructions for the execution of the steps of the method when said computer program is executed on the administration server. Alternatively, all or part of these means could be micro-programmed or micro-wired hardware means in dedicated integrated circuits. Thus, as a variant, the administration server 112 could be an electronic device composed solely of digital circuits (without a computer program) for carrying out the same actions.

In other embodiments of the invention, the various functions of the administration server 112 could be distributed among several devices, for example between several computers. For example, the resource manager could be implemented on one computer and the routing establishment module on another computer.

The cluster of computers 100 allows the distribution of complex processing and/or parallel calculations over at least part of the calculation computers 102.

As will be explained in more detail below, each deployed task is allocated, on the one hand, nodes and, on the other hand, switches and links between these switches, forming a sub-network for interconnecting the allocated nodes.

In the following description the following terminology will be used.

A link is said to be unused when it is not allocated to any task. Thus, no task uses this link, so that no data related to the execution of a task passes through this link. The link is said to be used in the opposite case, i.e. when it is allocated to at least one task, so that data related to the execution of this or these tasks passes through this link.

A link is said to be available when it is unused (i.e., as explained previously, allocated to no task) or when it is allocated to less than N other tasks, N being a predefined number equal to at least a. The predefined number N therefore represents the maximum number of task(s) that can use the same link. The number N can be equal to one, in which case the two branches of the preceding alternative are equivalent. Alternatively, the number N can be two or more. The number N is for example defined before tasks are deployed in the cluster of computers 100, and applies for all the tasks deployed subsequently, until a redefinition of this number N. This redefinition takes place for example when no tasks are deployed in computer cluster 100.

A switch is said to be unused when it is not allocated to any task. It is said to be used when it is allocated to at least one task.

Referring to Figure 3, a method 300 of implementing a task, hereinafter referred to as the current task, in the computer cluster 100 will now be described.

The method 300 comprises a step 302 of deploying the current task in the cluster of computers 100, a step 304 of executing the current task and a step 306 of cleaning up after execution.

Step 302 deployment of the current task comprises the steps 308 to 316 following. Step 308 is for example performed by the resource manager, while steps 310 to 316 are performed by the routing establishment module.

During a step 308, the administration server 112 receives the task in progress, as well as a number representing the number of nodes that the task in progress needs for its execution.

During a step 310, the administration server 112 allocates nodes to the task in progress, connected to the minimum possible number of unused switches. For example, the administration server 112 allocates in priority the nodes connected to the switches already in use, then consecutive nodes.

During a step 312, the administration server 112 determines a subnet of the network 110, satisfying predefined determination criteria.

The determination criteria include in particular a first criterion according to which the sub-network interconnects the allocated nodes.

In the example described, the determination criteria also includes a second criterion according to which the subnet uses only available links (that is to say, allocated to no other task or allocated to less than N other tasks already deployed) and a third criterion according to which the determined sub-network is that, among at least two sub-networks satisfying the other criteria, using the most switches already used (that is to say allocated to at least one task already deployed).

In the example described, the at least two subnets include all the subnets satisfying the other criteria.

In one embodiment, the searched sub-networks are trees. Thus, the administration server 112 traverses the switches, for example in ascending order of level, to find each tree (i) originating from this switch, (ii) using only available links and (iii) serving all the nodes allocated to the current task. Thus, in the example described, the administration server 112 determines all the possible trees having the same minimum level to serve all the allocated nodes (indeed, the trees whose top is at a higher level will necessarily require more unused switches) and selects, among these possible trees, the one using the most switches already used.

Alternatively, one could use an algorithm constructing the sub-network satisfying the three determination criteria. Thus, there would be no need to explicitly determine all subnets satisfying the criteria other than the third, to select one. The build itself would ensure that the subnet built would be the one using the most switches already in use.

For example, a subnet construction algorithm is presented in the appendix.

This algorithm uses the notion of sub-topology.

A sub-topology is, for example, a set of one or more switches, none of which is connected to a lower level switch. Thus, the complete topology is an example of a sub-topology. Each top-level switch is also an example subtopology.

Another definition of sub-topology is given in the article by Jean-Noël QUINTIN and Pierre VIGNERAS entitled "Transitively Deadlock-Free Routing Algorithms" published in 2016 in the Conference: HiPINEB, HPCA and accessible at the following address: https ://www.researchgate.net/publication/301345528_Transitively_Deadlock-Free_Routing_Algorithms

This article defines the concept of sub-topology (“enclosing sub-topology”), represented by a tuple (d ₁ ,…d _l ), l < l _max -1 : the set of level k switches, k < l _max -l with topological addresses in the form of (k, d ₁ , …, d _l , e _i …, e _lmax-1 ): they all share the same infix given by the subtopology notation.

During a step 314, the administration server 112 allocates the subnet determined in step 312 to the current task. The links and switches of this subnet are thus allocated to the current task. Thus, each link is dedicated to a maximum of N tasks, which makes it possible to reduce the impact of the tasks between them. Thus, it is possible to reduce the variance of successive execution times of the same task. Indeed, each time this task runs, the links it will use will be shared with a limited number of other tasks. At best, when the number N is one, the links used by this task will be completely dedicated to it. Thus, the execution environment of this task varies little with each new execution, which leads to the reduction of the variance of the execution times.

During a step 316, the administration server 112 implements, in the interconnection network 110, inter-node communication routes in the allocated subnet. In the example described, this implementation is carried out by modifying, if necessary, the local routing tables. More precisely, in the example described where a PGFT network is used, it is only necessary to modify the uplink routing tables, i.e. to go to a higher level. Furthermore, the upstream routing tables should only be modified in the case where the nodes allocated to the task do not form a complete sub-topology. Thus, the possible modifications to be made are very limited in number, which does not significantly slow down the deployment of the task.

After deployment step 302, the current task runs in step 304.

Stage 306 cleanup after execution includes steps 318 to 324 following.

During a step 318, the administration server 112 cancels the modifications made in step 316 in the local routing tables. This return to the default routing tables means that the deployment of a new task will require only a few modifications to the default routing tables, as explained above in connection with step 316.

During a step 320, the administration server 112 releases (that is to say, deallocates) the nodes and the subnet allocated to the task in progress.

During a step 322, the administration server 112 searches for the unused links of the computer cluster 100 (that is to say which are not allocated to any task). Preferably, the search is performed only among the links released at step 320.

During a step 324, the administration server places each link found in an inactive state in which the link requires an energy consumption lower than an energy consumption required by an active link associated with at least one task (and therefore used by this or these tasks).

For example, in one embodiment, administration server 112 shuts down both network interfaces that this link connects so that no power or light travels through this link. In another embodiment, the administration server 112 decreases an operating frequency of the two network interfaces that this link connects.

With reference to FIGS. 4 to 12, a non-limiting example of implementation of the method 300 will now be described.

It is assumed that computer cluster 100 is configured such that the predefined number N is one. Thus, each link in the computer cluster 100 can only be allocated to at most one task. Also, it is assumed that initially no tasks are deployed in computer cluster 100. Thus, all links are available. As a result, the local default routing tables are used and the links are placed in the inactive state in order to reduce the power consumption of the cluster of computers 100.

Referring to Figure 4, the administration server 112 implements step 308 and receives a task t1 requiring n1 nodes of the computer cluster 100. In the example described, n1 equals five.

The administration server 112 implements the step 310 and allocates to the task t1 the following n1 nodes: 1, 2, 3, 4, 5.

The administration server 112 implements step 312 and determines an interconnecting subnet of the allocated nodes. In the example described, the administration server 112 traverses the switches in increasing order of level and, in the example described, by increasing number on the same level (from left to right in FIG. 4). Thus, the administration server 112 traverses the first level switches, then those of the second level. However, for the switches 1.1 to 1.8 and 2.1 to 2.8, there is no tree connecting all the allocated nodes 1, 2, 3, 4, 5. Thus, the administration server 112 arrives at the switch 3.1, d 'where the A1 tree shown in bold in Figure 4 starts. The A1 tree only uses available links and interconnects all allocated nodes 1, 2, 3, 4, 5.

There is also another tree, from Switch 3.3, which interconnects all allocated nodes and only uses available links. These two trees have the same number of unused switches (six), so the administration server can retain any one. The administration server 112 retains for example the tree A1.

Administration server 112 implements step 314 and then allocates tree A1 to task t1.

The administration server 112 implements step 316 and implements, in the tree A1, communication routes between the allocated nodes. For this, in the example described, the administration server 112 modifies, if necessary, the local routing tables.

At this point, the task t1 is deployed and is executed (step 304) by the cluster of computers 100, on the allocated nodes which communicate with each other following the routes of the tree A1.

Referring to Figure 5, the administration server 112 implements step 308 and then receives a task t2 requiring n2 nodes of the computer cluster 100. In the example described, n2 equals five.

The administration server 112 implements the step 310 and allocates the task t2 nodes by allocating in priority the nodes connected to the switches already used, that is to say the node 6. Thus, in the example described, the administration server 112 allocates to the task t2 the following n2 nodes: 6, 7, 8, 9, 10.

The administration server 112 implements step 312 and determines an interconnecting subnet of the allocated nodes. As before, the administration server 112 traverses the switches by increasing level. The first switch from which a tree serving the allocated nodes 6, 7, 8, 9, 10 starts is switch 4.1. This tree nevertheless includes the link between the switches 2.3 and 3.1, which is already allocated to task t1 and therefore not available. Thus, the administration server 112 does not retain this tree and passes to the switch 4.2, from which the tree A2 starts. This tree only uses available links and serves all the allocated nodes 6, 7, 8. In the same way, the administration server 112 determines the trees coming from the switches 4.4, 4.6, and 4.8 respectively illustrated in FIGS. 6 , 7 and 8. These four trees have the same number of unused switches. Thus, the administration server 112 can select any tree, for example the tree A2 coming from the switch 4.2 (FIG. 9).

Administration server 112 implements step 314 and allocates tree A2 to task t2.

The administration server 112 implements step 316 and implements, in the tree A2, communication routes between the allocated nodes.

For this, in the example described, the administration server 112 modifies, if necessary, the local routing tables. For example, the local routing table for port P_6 of switch 1.3 is modified as shown in Table 2 below (modification shown in bold and italics).

Local routing table modified from port P_6 of the switch 1.3 Destination nodes output port 1 P_2.3 2 P_2.4 3 P_2.3 4 P_2.4 5 P_5 6 P_6 7 P_2.4 8 P_2.4 9 P_2.3 10 P_2.4 11 P_2.3 12 P_2.4 13 P_2.3 14 P_2.4 15 P_2.3 16 P_2.4

Thus, the messages originating from node 6 and destined for nodes 7 to 10 all pass through the link connecting switches 1.3 and 2.4 and therefore avoid the link connecting switches 1.3 and 2.3 allocated to task t1, which would not have not be the case if the default routing table had been kept.

At this stage, the tasks t1, t2 are deployed and executed (step 304) by the cluster of computers 100, on the respectively allocated nodes which communicate with each other by respectively following the routes of the trees A1 and A2.

Referring to Figure 10, the administration server 112 implements step 308 and then receives a task t3 requiring n3 nodes of the computer cluster 100. In the example described, n3 equals three.

The administration server 112 implements step 310 and allocates nodes to task t3 by first allocating nodes connected to switches already in use (none at this stage), then consecutive nodes. Thus, in the example described, the administration server 112 allocates to the task t3 the following n3 nodes: 11, 12, 13.

The administration server 112 implements step 312 and determines an interconnecting subnet of the allocated nodes. The administration server 112 therefore finds the three trees originating respectively from the switches 3.5, 3.7, 3.8, respectively illustrated in FIGS. 10, 11 and 12, satisfying the first and second determination criteria. Of these three trees, the one from switch 3.8 (hereinafter called A3) has four unused switches (numbered in bold in Figure 12), while the other two trees each require five unused switches (numbered in bold on the figures 10 and 11). Thus, the tree from switch 3.8 (tree A3) satisfies the third determination criterion and is therefore retained by the administration server 112.

Administration server 112 implements step 314 and allocates tree A3 to task t3.

The administration server 112 implements step 316 and implements, in the tree A3, communication routes between the allocated nodes, modifying, if necessary, the local routing tables.

At this stage, the tasks t1, t2 and t3 are deployed and executed (step 304) by the cluster of computers 100, on the respectively allocated nodes which communicate with each other by respectively following the routes of the trees A1, A2 and A3 (figure 13).

Referring to Figure 14, at a later time, the task t2 ends its execution.

Thus, the administration server 112 implements step 318 and cancels the modifications made in the local routing tables to deploy the task t2.

Administration server 112 implements step 320 and releases (i.e. deallocates) the nodes and subnet A2 allocated to task t2. Released links are shown in dotted and bold.

The administration server 112 implements step 322 and searches for the links of the cluster of computers 100 which are not allocated to any task. Preferably, the search is carried out only among the links of task t2. In the example described, as a link can only be allocated to at most one task (N is equal to one), all the links that were allocated to task t2 become unused.

The administration server 112 implements step 324 and places each link found (that is to say, in the example described, all the links which were allocated to the task t2, represented in dotted and bold) in the inactive state.

Thus, it is possible to reduce the energy use by deactivating the unused links, in particular, those connecting switches to the computing nodes. Generally, on a PGFT, half of the links are dedicated to the compute nodes and the other half to the inter-switch links. If for example ten nodes connected to the same switch are used for a particular task, these ten nodes will communicate outside only towards the input/output nodes (for example, the nodes of the storage computers 104). Thus it is possible to turn off up to ten links during the execution of this task (depending on the number of input/output nodes in the supercomputer as well as the input/output node requirements requested by the user , it is possible to keep these links active).

It is clear that a process such as the one described above makes it possible to reduce the energy consumption of the supercomputer.

It will also be noted that the invention is not limited to the embodiments described above. It will indeed appear to those skilled in the art that various modifications can be made to the embodiments described above, in the light of the teaching which has just been disclosed to them. In the detailed presentation of the invention which is made above, the terms used must not be interpreted as limiting the invention to the embodiments set out in the present description, but must be interpreted to include therein all the equivalents whose provision is within the reach of those skilled in the art by applying their general knowledge to the implementation of the teaching which has just been disclosed to them.

ANNEX

– Count the number of nodes allocated per sub-topology.

– For each sub-topology with a non-zero number of allocated nodes:

-- (switches respectively bearing increasing numbers s starting from 0) for s = 0; s < S_n; s++:

--- check if there is a tree of links allowing access to all the nodes allocated in the sub-topology,

--- if yes, increment the number of accessible top switches of these switches,

--- if not, go to the next one.

– Take one of the most important sub-topologies (called T_n_c of level n: the highest level switches are of level n).

– Consider that there are S_n switches of level n.

– For each sub-topology included in T_n_c of levels n-1, we have a number of nodes allocated x_n-1 (To simplify, it is considered that each sub-topology has the same number of nodes allocated. This particular case is easily generalizable in case two sub-topologies have different numbers of allocated nodes).

– As long as nb_switch_allocated_level_n at level n is less than x_n-1:

-- take an unvisited l1 switch:

--- there are X level l1 switches in the smallest enclosing subtopology,

--- select a higher level switch with a counter of value T as close as possible (nb_switch_allocated - x_n-1)

--- allocate all downlinks from this switch to the X switches with allocated nodes (including the selected one)

--- select in the twin sub-topologies (“twin” in the article by QUENTIN and VIGNERAS) the same switch (it is necessarily allocable due to the way the counter is calculated),

--- do the same again at most X times considering a switch l2 which has just been allocated as long as nb_switch_allocated_level_n is less than x_n-1,

---- at each top switch allocation, increment nb_switch_allocated_level_n,

--- then at the following levels,

--- once all the levels are done, decrement the counters of the allocated switches by X.

Note: In case of unbalanced allocation, it is possible to redo it for certain sub-topologies (in order to have the necessary bandwidth for the nodes).

Claims (9)

Method (302) for deploying a task in a cluster of computers forming a supercomputer comprising:

• computers (102, 104) each having at least one network interface forming a node (1…16); and
• a network (110) for interconnecting the nodes (1...16), comprising:
  • switches (1.1…4.8) each having network interfaces, and
  • links (114) each connecting either a node (1…16) and a network interface of a switch (1.1…1.8), or two network interfaces of two switches (2.1…4.8) respectively;

the method (302) comprising:

• allocating (310) nodes (1…16) to the task;

the method (302) being characterized in that it further comprises:

• the determination (312), in the network (110), of a sub-network, of interconnection of the allocated nodes, satisfying one or more predefined determination criteria comprising a first criterion according to which the determined sub-network is the one, among at least two sub-networks satisfying the criterion or criteria other than the first criterion, using the most switches already allocated each to at least one task already deployed;
• the allocation (314) of the sub-network, and in particular of the links belonging to this sub-network, to the task; and
• implementing (316) inter-node communication routes in the allocated subnet.

The method (302) of claim 1, wherein determining the subarray includes determining the at least two subarrays satisfying the one or more criteria other than the second criterion, then selecting from among those at least two subarrays of the one using the most switches already allocated each to at least one task already deployed. Method (302) according to claim 1 or 2, in which the criterion or criteria for determining the subnet further comprise a second criterion according to which the subnet only uses links which are not allocated to any other task already deployed or good which are allocated to less than N other tasks already deployed, N being a predefined number equal to one or more. A method (302) according to claim 3, wherein the predefined number N is one. Method (300) of implementing a task, comprising:

• deploying the task in a supercomputer, in accordance with a deployment method (302) according to any one of claims 1 to 4;
• performing (304) the task; and
• at the end of the task execution:
  • the detection (322) of at least one link which is not allocated to any task, and
  • placing (324) each detected link in an inactive state, in which the link requires less power consumption than a power consumption required by a link associated with at least one task.

Method (300) according to claim 5, comprising, to place a link in the inactive state, the extinction of the two network interfaces that this link connects so that no current or light traverses this link. Method (300) according to claim 5, comprising, to place a link in the inactive state, the reduction of an operating frequency of at least one of the two network interfaces that this link connects. Computer program downloadable from a communication network and/or recorded on a computer-readable medium and/or executable by a processor, characterized in that it comprises instructions for the execution of the steps of a method according to any of claims 1 to 7, when said program is executed on one or more computers. Cluster of computers forming a supercomputer comprising:

• computers (102, 104) each having at least one network interface forming a node (1...16);
• a network (110) for interconnecting the nodes (1...16), comprising:
  • switches (1.1…4.8) each having network interfaces, and
  • links (114) each connecting either a node (1…16) and a network interface of a switch (1.1…1.8), or two network interfaces of two respective switches (2.1…4.8);

the supercomputer further comprising software and/or hardware means for deploying a task, designed to implement the following step:

• allocating (310) nodes (1…16) to the task;

the supercomputer being characterized in that it further comprises software and/or hardware means for deploying a task, designed to implement the following steps:

• the determination (312), in the network (110), of a sub-network, of interconnection of the allocated nodes, satisfying one or more predefined determination criteria comprising a first criterion according to which the determined sub-network is the one, among at least two sub-networks satisfying the criterion or criteria other than the first criterion, using the most switches already allocated each to at least one task already deployed;
• the allocation (314) of the sub-network, and in particular of the links belonging to this sub-network, to the task; and
• implementing (316) inter-node communication routes in the allocated subnet.