WO2018109407A1

WO2018109407A1 - Method and device for updating

Info

Publication number: WO2018109407A1
Application number: PCT/FR2017/053585
Authority: WO
Inventors: Patrick Brown; Son Hai HA
Original assignee: Orange
Priority date: 2016-12-16
Filing date: 2017-12-14
Publication date: 2018-06-21
Also published as: FR3060791A1; FR3060791B1

Abstract

The invention envisages a method of updating for at least one server of a cluster a current value of a variable from among an estimation of a quantity already used of a computing resource on the server or a maximum capacity of the server for this resource, this current value being taken into account by a device for managing the cluster so as to select on receipt of a computing task execution request, a target server to execute this task, the updating being carried out at each detection (F10) of an event (EV) from among a predefined set of events and: ¾ if the detected event is the obtaining of a measurement of a current actual use of the resource on the server, the updating (F30) uses the measurement of the current actual use of the resource on the server; ¾ if the detected event is the reserving for a computing task of a quantity of the resource on the server, the updating (F40) uses the quantity of resource reserved; and ¾ if the detected event is the releasing by a computing task of a quantity of the resource on the server, the updating (F50) uses the quantity of resource released.

Description

Method and device for updating

Background rinenvention

The invention relates to the general field of computer systems, and more particularly to computer systems based on machine clusters (also more commonly referred to as "servers").

In a known manner, a cluster (or cluster) of servers groups together several independent computers (or nodes) placed in a network, so as to allow a global and optimized management of computing resources (eg memory, CPU (Central Processing Unit), disk space , network access, etc.) offered by these computers. Such a cluster of servers typically allows parallel distribution and launch, on several servers different from the cluster, of different computer tasks associated with the same program or the same computer application (also referred to as "work" or "job" in English). In the context of current data processing, which can involve very large amounts of data (eg so-called "Big Data" processes), it is not uncommon to encounter server clusters with thousands or even tens of thousands of machines that need to be administered effectively.

The distribution of tasks on different machines in a cluster is usually administered by a cluster management system. Known cluster management systems are for example the Borg system proposed by Google or the Hadoop Yarn system proposed by The Apache Software Foundation. These systems are capable of administering clusters of a very large number of machines, and running hundreds of thousands of tasks on these machines belonging to several thousand different software applications.

In this context, cluster management systems are often faced with tasks that need to be burst on clusters. It is therefore necessary for them to deal effectively with these tasks to make a decision very quickly as to their distribution on the different machines of the clusters. For this purpose, state-of-the-art cluster management systems allocate the resources of a cluster to the different tasks of a program (in other words, associate with these different tasks different servers in the cluster capable of providing the necessary resources at their execution and reserve on these servers the required resources) on the basis of the resources reserved by the user of the cluster which requests the execution of his program.

Thus, for example, the Hadoop Yarn cluster management system defines, for each server S of a Hadoop cluster and each computer resource R offered by the server, a limit capacity CL_R. For the sake of simplicity, it is considered here that all the servers S have the same capacity limit for the resource R. The computing resource R can be of various nature, for example memory, CPU (Central Processing Unit), space disk, network access, etc. To decide which resource to allocate on which server of the cluster to a task j of a program (or job), the Hadoop Yarn cluster management system quickly determines which servers are capable of performing this task, i. the servers that have the necessary resources for its execution in view of the resource requirements R declared by the user of the program. It uses for this purpose the limit capacity CL_R of each server, the R_RJ needs of the task j R resource declared by the user of the program, and the total amount T_R resource R already reserved on each server. A server is then considered to be able to execute a task j only if for this server:

R_RJ + T_R <CL_R.

A server is then selected by the Hadoop Yarn management system among the servers able to execute the task j, according to predetermined criteria, not detailed here.

This mode of operation is not without constraints for the users of the cluster. This requires users to specify the resource requirements of the different tasks in their programs in sufficient detail not to unnecessarily block resources at the cluster level or, on the contrary, risk seeing their tasks not be processed or treated less efficiently because they require more resources than those declared by the user. Such an estimation of the needs of the tasks of their programs is particularly difficult for the users, especially since these needs are likely to evolve over time. In this context, it is traditional for users to overestimate the resource requirements of the tasks in their programs, which leads to underutilization of cluster resources. Object and summary of the invention

The invention makes it possible to overcome the aforementioned drawbacks by proposing an updating method, for at least one server of a server cluster of at least one current value of a variable denoted V chosen from an estimate of one used amount of a computing resource on the server or a maximum server capacity for said computing resource, said at least one current value being intended to be taken into account by a server cluster management device to select, upon receipt of a request for execution of a computer task associated with a predetermined quantity of said computer resource initially required for its execution, a so-called target server of the cluster for executing this computer task, said method being such that said update is carried out at each detection of an event from a predefined set of events and:

If the detected event is to obtain a measure of a current actual use of the computing resource on the server, said updating of the current value of the variable V uses said measure of the current actual usage of the computer resource on the server;

If the detected event is a reservation by the management device for a computer task of a quantity of the computing resource on the server, said updating of the current value of the variable V uses said quantity of the reserved computer resource ; and If the detected event is a release by a computer task of a quantity of the computing resource on the server, said updating of the current value of the variable V uses said quantity of the released computing resource.

The invention also relates to a method for managing computer tasks comprising, on reception by a server cluster management device of a request to execute a computer task, this computer task being associated with a predetermined quantity of tasks. a computer resource initially required for its execution:

A step of determining, by the cluster management device, at least one cluster server capable of executing the computer task, this determining step using the predetermined quantity of the computer resource initially required for the execution of the task; computing, a maximum server capacity for the computing resource and an estimate of an already used amount of the computing resource on the server;

A step of selection, by the cluster management device, of a server for executing the computer task among said at least one server of the determined cluster.

This method is remarkable in that, for at least one server of the cluster, at least one current value of a variable denoted V chosen from the estimate of the quantity already used of the computing resource on the server and the maximum capacity of the server for the computer resource is updated according to an updating method according to the invention.

No limitation is attached to the nature of the computing resources of the cluster servers managed by the cluster management device and required for the execution of the computing tasks in the context of the invention. It can be any type of resource (i.e. hardware or software components) and / or networks such as memory, CPU, network access, disk space, etc.

In addition, the invention is described with reference to a single resource; however, the same method can be used to account for multiple resources, applying it to each resource individually. It is during the actual selection of the server on which to execute the task that the plurality of resources will be taken into account by choosing a server having each of the resources necessary for the execution of the task.

The invention therefore proposes a solution that makes it possible to take into account the current and actual use (i.e. effective) of the resources at the level of the servers of a cluster to manage the allocation of resources within this cluster.

This solution advantageously relies on the current mechanism of resource allocation applied by the state of the art cluster management systems, which makes it possible to ensure a rapid response of the management device, especially when bursting tasks must be started by it on the cluster. Indeed, without prior knowledge of the actual use of resources by a computer task, the management device reserves for this task at the server on which it launches, a quantity of resource equal to the need declared by the user at the origin of the computer task (predetermined quantity of the computer resource initially required for the execution of the computer task within the meaning of the invention).

The mechanism of the state of the art is however adapted so as to take into account, after the launching of a task by the management device on a server of the cluster, the actual use of the resources by this task on the server.

This allows the management device, if this actual use is less than the task requirements declared by its user, to allocate the resources not used by the task to other computer tasks presented later to the management device. This is taken into account by updating, from measurements made directly on the servers, the quantities used by the management device to decide on the allocation of resources, namely the estimate of the quantity already used. the IT resource required by the server task and / or the maximum server capacity for that resource. In a particular embodiment, the obtaining of the measures of current use of the resource can be periodic.

In addition, according to the invention, other events affecting the resources available on the server trigger an update of the current value of the variable V considered. These events are the reservation on the server, by the management device, of a quantity of the resource considered for other tasks, as well as the release of a quantity of this resource on the server by tasks which have finished run on this one.

This makes it possible to take into account the computer tasks that are positioned asynchronously on the server by the management device between two uploads of measurements of actual use of the resource on the server (these updates are generally not continuous but periodic or at predetermined times), as well as those ending on the server. Taking into account the completed tasks is of particular interest when a computer task launched on a server ends very quickly, typically in a time less than the period of recovery of the actual usage measurements made on the server if necessary. This eliminates the resource remaining reserved for this task on the server and is no longer used by this server without being reflected in the measurement.

The invention therefore makes it possible to transmit in real time the various variations that may affect the quantity of resource available on the server in the quantity taken into account by the management device to respond to a request to execute a computer task. . The cluster management device has indeed a more accurate representation of the amount of resource actually used on the server at any time, or at least when it receives a request to run a task.

The cluster management device can thus simultaneously, thanks to the invention, learn how the resources are actually used on the servers while being able to make decisions quickly to allocate resources to the computer tasks presented to it. This quick decision making is done instantly on the basis of resources that are reserved by users for these tasks without waiting to know how these tasks will actually behave. The invention thus solves the dual problem of addressing bursts of tasks without having knowledge of their future use of the resources of the cluster while allowing to allocate to other tasks unused resources of the cluster.

The invention therefore proposes a solution compatible with large clusters and software applications comprising a large number of computer tasks to be executed on these clusters, and which makes it possible to progressively correct an erroneous or inaccurate estimate of the resource requirements of these tasks. by their users. Constraints imposed on cluster users can be relaxed.

Moreover, the resources of the clusters are better used while attempting to ensure the proper execution of the tasks on these clusters: it is indeed limited, thanks to the invention, the long waiting times for the execution of a task computer or eviction of a computer task especially in case of insufficient resources allocated to this task taking into account the actual use of resources on the servers of the cluster. It should be noted that an extended waiting time or the eviction of a task is always possible, however he / she is counterbalanced by the fact that more tasks can be started simultaneously on the job. cluster because of a better management of the resources of the latter. The performance of the cluster is thus globally improved.

In addition, the invention makes it possible to adapt to a possible variation over time in the resource requirements of the tasks launched on the servers of the clusters.

It is noted that updating the server's maximum capacity for the resource rather than estimating the amount already used of the computer resource required by the task on the server advantageously makes it possible to use as cluster management device the one of the management systems of the state of the art. Indeed, the maximum capacity of each server is information that is reported today periodically by this server cluster management system in the state of the art. Updating this capacity at the server level to reflect the actual usage of the resource ensures that the cluster management system takes into account this actual use for allocating cluster resources in a way that totally transparent for him. In other words, this makes it possible to envisage an implementation in which the operation of the centralized management system of the cluster remains the same (and therefore in which it is possible to use management systems of the state of the art, such as previously mentioned Borg or Hadoop Yarn systems), and where only the servers in the cluster must be adapted to take into account in the maximum capacity that they go back to the cluster management system, the actual use of their resources. Such an implementation is particularly interesting because it allows a progressive installation of the invention on all servers in the cluster while keeping the cluster management system unchanged. But updating servers is in practice a much simpler task than an update of the cluster management system which in itself is often an already very complex module. In addition, in this variant, the invention can rely on the signaling (ie communication protocol) already defined between the servers of the cluster and the cluster management system.

In a particular embodiment, if the detected event is to obtain the measurement of the current real use of the computing resource on the server, updating the current value of the variable V also uses at least a value of the variable V noted Vpast updated at a previous time.

In this embodiment, during the update, the measurement of the current actual use of the computing resource can be weighted in particular by a real factor a between 0 and 1, and said at least one Vpast value can be weighted by a factor equal to (1-a).

This embodiment, which implements an exponential moving average, makes it possible to smooth over time the taking into account of the actual use at the server of the resource with respect to the quantity of resource initially reserved. This smoothing makes it possible to progressively take into account the actual use of resources at the level of each server while avoiding abrupt transient fluctuations that could lead to a less efficient allocation of resources by the cluster management device. Indeed, the more the factor a is chosen close to 0, the longer the amount of reserved resources is taken into account and the use of resources available in the cluster by the management device is delayed. On the contrary, by choosing a factor a close to 1, the management device can use the available resources earlier in time. It is noted that the factor a may vary over time.

Of course, other ways of taking into account the measurement of the actual use of the computer resource and, if appropriate, said at least one value updated at a previous instant may be envisaged. For example, it is possible to rely on the principle of an arithmetic or weighted moving average of measurements obtained at different times and evaluated on a sliding window of predetermined size whose end is positioned on the current instant, etc. In the case of a weighted average, one can consider applying decreasing weights as the age of the measures increases.

By way of example, in a variant embodiment of the invention in which the updated variable V is the estimate EU_R of the quantity already used of the computing resource on the server, if the detected event is obtaining of measuring the actual current usage of the computing resource on the server, said current value denoted EU_R (n) of the variable V is updated according to the equation:

EU_R (n) = max ((l - a) EU_R (n - 1) + aU _R (n), U _R (n)) where n denotes an integer greater than 1, EU_R (n - 1) is the value of estimating the previously used amount of the computer resource on the server updated at a previous time, and U_R (n) denotes the metric obtained from the actual current usage of the computing resource on the server. This variant has a preferred application when the computing resource in question is memory. Indeed, the constraints in terms of memory resources are relatively strong constraints in the sense that a task can be evicted from a machine on which it is executed if it requires at a given moment more memory than the machine has at that time. moment. Forcing the updating of the value EU_R (n) to a value greater than or equal to the memory resource U_R (n) actually used on the server when the value of (l - a) EU_R (n - 1) + U_R ( n) does not sufficiently reflect the actual memory usage on the server, it is ensured that the management device does not unduly allocate to new memory tasks on the server which might then fail to the computer tasks already running on the server and prove to be detrimental to the processing of these.

In another embodiment of the invention, if the detected event is to obtain the measurement of the actual current use of the computer resource on the server, this measurement corresponds to a use of the computer resource on the server. by one or more computer tasks running on the server and all managed by the cluster management device.

On a server can in fact run in addition to the computer tasks launched on this server by the cluster management device, other processes such as typically internal server processes launched for example by its operating system. In this embodiment, it is ensured that only the tasks managed and positioned strictly speaking by the cluster management device are taken into account by the latter when allocating resources to the tasks presented to it. In other words, this embodiment allows a strong intervention of the management device with respect to the rights (i.e. resource quantities) reserved for it on the server.

One way to implement this embodiment is for example the following:

The measurement of the current real use of the computing resource on the server comprises a plurality of values denoted U_R (n, f 1, j = 1, ... J, associated respectively with a plurality of computer tasks Tj, j = l , ... J, set by the cluster management device on the server, where J and n denote integers greater than or equal to 1, each value U_R (n, fi corresponding to a measure of the actual use of the computing resource on the server by the task Tj;

Said update (F30) comprises:

o a first evaluation step for each task Tj, of a current value EU_R (n, j) of an estimate of a quantity already used of the computer resource on the server by the computer task Tj according to the equation:

EU_R (n, j) = max ((l - a) EU_R (- 1,;) + U_R (n, j), U_R (n, j)) or according to the equation: EU_R (n, j) = (1 - a) EU_R (n - 1,;) + U_R (n, j)

or :

EU_R (n - lj) denotes a value evaluated at a previous instant of the estimate of the quantity already used of the computing resource on the server by this computing task Tj; and

■ is a real factor between 0 and 1;

a second step of evaluating the current value denoted EU_R (n) of the estimate of the quantity already used of the computing resource on the server according to the equation:

or according to the equ

where J denotes the number of computer tasks set on the server by the cluster management device at the time of the second evaluation step.

In this variant, we therefore have access to more precise information on the actual use of the server, since these are measures of the actual use of the server for each computer task positioned on the server (that is to say executing on it or for which a quantity of the resource has been reserved by the management device on the server) that are available and used when updating the value EU_R (n). In this way, when a task disappears and frees the resource on the server, this is taken into account instantaneously in its contribution to the quantity EU_R (n) (the sum realized to evaluate EU_R (n) only takes into account that tasks set on the server at this time, ie at the time of evaluation of EU_R (n)).

Note, however, that this variant requires additional calculations to update the quantity EU_R (n), related to the individual management tasks Tj, j = l, ..., J. These calculations can be implemented preferentially locally at the level of each server and then be sent back to the cluster management device by the nodes.

In another embodiment of the invention in which the updated variable V is the estimate EU_R of the amount already used of the computing resource on the server, and if the detected event is obtaining the measurement of the actual current use of the computing resource on the server, said current value denoted EU_R (n) of the variable V is updated according to the equation:

EU_R (n) = U_R (n) + DR_R (ri) where U_R (n) denotes a current measure of the actual use of the computing resource on the server, and DR_R (n) denotes a current value of a quantity of the computing resource reserved for at least one computing task by the computing device. management on the server.

In addition, in this embodiment, the method may further comprise:

A step of updating the current value DR_R (n) of the quantity of the computer resource reserved for at least one computer task by the management device on the server is obtained according to the equation:

DR_R (n) = (1 -) DR_R (n - 1)

where β denotes a real factor of between 0 and 1 and DR_R (n - 1) denotes an amount of the computer resource reserved by the management device on the server obtained at a previous instant;

A step of detecting the updating of the current value DR_R (n); and

A step of updating the current value of the variable EU_R according to the equation:

EU_R (n) = U_R (n) + DR_R (n).

In this embodiment, the weighting (by a factor (l - β) here) is no longer applied to the measurement of the actual use of resources on the server but on the resources that are reserved by the cluster management device. on this server. In other words, the amounts of reserved resources are smoothed rather than the actual use of the resource. This embodiment ensures that the cluster management device takes into account, when allocating resource to a new task, a quantity of this resource used on the server always at least equal to its actual use (ie in this embodiment, it is not possible to resort to a "max" function as in the examples previously described).

In another embodiment, the updating method further comprises:

If the detected event is a reservation by the management device for a computer task of a quantity R_R_Tnew of the computer resource on the server, a step of updating the current value DR_R (n) of the quantity of the computer resource reserved for at least one computer task by the management device on the server using said quantity R_R_Tnew and

a quantity DR_R (n - 1) of the computer resource reserved by the management device on the server obtained at a previous instant,

If the detected event is a release by a computer task of a quantity R_R_Told of the computing resource on the server, a step of updating the current value DR_R (n) of the quantity of the computer resource reserved for the at least one computer task by the management device on the server using said quantity R_R_Told and the quantity DR_R (n - 1);

and wherein said update of the current value of the variable EU_R uses said amount of the reserved computer resource according to the equation: EU_R (n) = U_R (n) + DR_R (ri).

The choice of one or other of the embodiments described above may depend on various parameters, such as in particular the context in which the invention is implemented or a compromise between implementation complexity and performance of the invention. resource allocation made by the management device.

In a particular embodiment, the update method is implemented by the cluster management device.

Correlatively, in this embodiment, the invention also provides a device for managing a server cluster comprising:

A reception module;

A determination module, activated on reception by the reception module of a request for execution of a computer task, said computer task being associated with a predetermined quantity of a computer resource initially required for its execution, the module of wherein the determination is configured to determine at least one server in the cluster capable of performing the computing task using the predetermined amount of the computing resource initially required for the execution of the computing task, a maximum server capacity for the computing resource and an estimate of an already used amount of the computing resource on the server;

A selection module configured to select a server to execute the computer task among said at least one server of the cluster determined by the determination module;

The management device according to the invention is remarkable in that it furthermore comprises:

A detection module configured to detect a predefined set of events; and

An update module activated for at least one server of the cluster, and configured to update at least one current value of a variable rated V chosen from the estimate of the quantity already used of the computing resource on the server and the maximum capacity of the server for the computing resource, said update module being activated on each detection of an event among the predefined set of events by the detection module, and configured to:

o if the detected event is to obtain a measurement (U_R (n)) of a current real use of the computer resource on the server, update the current value of the variable V using said measure of the actual actual use of the computing resource on the server;

o if the detected event is a reservation by the management device for a computer task of a quantity of the computing resource on the server, updating the current value of the variable V by using said quantity of the reserved computer resource; and o if the detected event is a release by a computer task of a quantity of the computing resource on the server, updating the current value of the variable V using said amount of the released computing resource.

In another embodiment, the updating method is implemented by the server. The management method then further comprises a step of transmitting said at least one value updated by the server to the cluster management device.

Correlatively, in this embodiment, the invention also relates to a server belonging to a server cluster managed by a management device, said server being able to execute at least one computer task managed by said management device, said at least one computer task using during its execution at least one computer resource of the server, this server comprising modules activated for at least one time interval and comprising:

A measurement module configured to measure a real use of the computer resource on the server;

A detection module configured to detect a predefined set of events; and an update module configured to update at least one current value of a variable rated V selected from an estimate of an already used amount of the computing resource on the server and a maximum capacity of the server for the resource computer, said update module being activated on each detection by the detection module of an event among the predefined set of events and configured to:

o if the detected event is to obtain a measure of a current actual use of the computing resource on the server, update the current value of the variable V using said measure of the current actual usage of the resource computer on the server;

o if the detected event is a reservation by the management device for a computer task of a quantity of the computing resource on the server, updating the current value of the variable V by using said quantity of the reserved computer resource; and

o if the detected event is a release by a computer task of a quantity of the computing resource on the server, updating the current value of the variable V by using said quantity of the released computing resource; and

A transmission module, configured to transmit to the management device said current value of the updated variable V.

As mentioned above, this embodiment can facilitate the progressive implementation of the invention by limiting the developments necessary for its implementation on the cluster management device.

In a particular embodiment, the various steps of the update method and / or the different steps of the management method are determined by computer program instructions. Accordingly, the invention also relates to a computer program on an information carrier, this program being capable of being implemented in a management device, in a server or more generally in a computer, this program comprising instructions adapted to the implementation of the steps of an updating method as described above.

The invention also relates to a computer program on an information medium, this program being capable of being implemented in a management device or more generally in a computer, this program comprising instructions adapted to the implementation steps of a management method as described above.

Each of these programs can use any programming language, and be in the form of source code, object code, or intermediate code between source code and object code, such as in a partially compiled form, or in any form what other form is desirable.

The invention also relates to a computer readable information or recording medium, and comprising instructions of a computer program as mentioned above.

The information or recording medium may be any entity or device capable of storing the program. For example, the medium may comprise storage means, such as a ROM, for example a CD ROM or a microelectronic circuit ROM, or a magnetic recording medium, for example a floppy disk or a disk. hard.

On the other hand, the information or recording medium may be a transmissible medium such as an electrical or optical signal, which may be conveyed via an electrical or optical cable, by radio or by other means. The program according to the invention can be downloaded in particular on an Internet type network.

Alternatively, the information or recording medium may be an integrated circuit in which the program is incorporated, the circuit being adapted to execute or to be used in the execution of the method in question.

The invention also relates to a computer system comprising:

A server cluster comprising a plurality of servers, each server comprising at least one computer resource enabling the execution of computer tasks;

A server cluster management device;

wherein the cluster management device is in accordance with the invention and / or at least one server of the cluster is in accordance with the invention.

It may also be envisaged, in other embodiments, that the updating method, the management method, the management device, the server and the computer system according to the invention present in combination all or some of the aforementioned characteristics. . Brief description of the drawings

Other features and advantages of the present invention will emerge from the description given below, with reference to the accompanying drawings which illustrate an embodiment having no limiting character. In the figures:

- Figure 1 shows, schematically, a computer system according to the invention, in a particular embodiment, the computer system comprising a server cluster and a cluster management device;

FIG. 2 represents the hardware architecture of the cluster management device of FIG. 1;

FIG. 3A represents, in the form of a flow chart, the various steps of a management method as implemented in a particular embodiment by the management device of FIG. 1;

FIG. 3B represents, in the form of a flow chart, the various steps of an updating method as implemented in a particular embodiment by the management device of FIG. 1; and

FIG. 3C represents, in the form of a flow chart, the various steps of an updating method in an alternative embodiment.

Detailed description of the invention

FIG. 1 represents, in its environment, a computer system 1, according to the invention, in a first embodiment.

According to the invention, the computer system 1 comprises

A cluster 2 of servers comprising a plurality of servers S1, S2,..., SK, K designating an integer greater than 1, connected in a network. Each server Sk, k = 1,..., K comprises various computer resources (hardware and / or software) enabling the execution of computer tasks. These resources are, for example, memory, disk space, CPU, network access, etc. No limitation is attached to the nature of the computer resources proposed by the servers S1,..., SK of the cluster 2. However, to better illustrate the invention, here is focused on a particular computer resource R which consists of memory; and

A server cluster management device 2 according to the invention.

The management device 3 of the cluster 2 is configured to receive a REQ request from a client device 4 (eg a computer) of a user of the cluster 2 requesting the execution of at least one computer task T, allocate cluster 2 resources to this client for the execution of this task. More specifically, upon receiving the REQ request, the management device 3 of the cluster 2 identifies a server Sk of the cluster 2 capable of executing the computer task T, that is to say a server Sk having the necessary resources for its execution (memory, CPU, etc.), then starts the computer task T on this server Sk for it to be executed. No limitation is attached to the nature of the computer task T to be performed on the cluster 2. It may be any type of task associated with an APP software application installed on the client device 4 or any program in general. In a known manner, such a computer program comprises a plurality of computer tasks to be executed, and which can be submitted in bursts to the management device 3 of the cluster 2 to be executed in parallel on different servers of the cluster 2. We are interested here more particularly the processing performed by the management device 3 on one of these tasks, namely on the task T. A similar treatment is applied to all the tasks contained in the REQ request.

Each of the tasks sent by the APP software application to the cluster management device 2 has IT resource requirements to enable its execution on the cluster 2. These requirements are specified here by the user of the client device 4 for each of the tasks. and for each resource. For the computing resource R and the task T, the user of the client device 4 specifies a quantity R_R_T of memory required "initially" for the execution of the task T. The quantity R_R_T can be determined roughly by the user of the client device 4, on the basis of its knowledge of the operations implemented by the computer task T and the consumption in terms of memory required by these operations. Note that although the user of the client device 4 tries as far as possible to provide a quantity R_R_T close to the actual use of the resource R by the computer task T, this quantity R_R_T is difficult in practice to estimate precisely for the user: it is generally overestimated by the user to ensure that the computer task T is allocated a sufficient amount of memory by the management device 3 cluster 2 to be executed in good conditions. The quantity R_R_T thus predetermined is transmitted here in the REQ request for execution of the computer task T addressed to the management device 3 of the cluster 2.

In the embodiment described here, the management device 3 of the cluster 2 has the hardware architecture of a computer, as illustrated in FIG. 2. It notably comprises a processor 5, a random access memory 6, a read-only memory 7, a nonvolatile flash memory 8 as well as communication means 9 enabling it to communicate in particular on the one hand with the client device 4 and on the other hand with the servers Sk, k = 1,..., K of the cluster 2. These communication means 9 comprise for example a network card, well known per se.

The ROM 7 of the management device 3 of the cluster 2 constitutes a recording medium according to the invention, readable by the processor 8 and on which is recorded here a computer program PROG according to the invention.

The computer program PROG defines functional modules (and software here), configured to implement the steps of the updating process and the steps of the computer task management method according to the invention. These functional modules rely on and / or control the hardware elements 5-9 of the management device 3 mentioned above. They include in particular here, as illustrated in FIG. A module 3A for receiving requests from client devices such as the client device 4, this module 3A based on the communication means 9;

A determination module 3B, activated on reception by the module 3A for receiving a request for execution of a computer task, the determination module being configured to determine at least one server of the cluster 2 capable of executing the computer task ;

A selection module 3C configured to select a server to execute the computer task among the cluster servers determined by the determination module 3B;

A 3D detection module, configured to detect a predefined set of events relating to the resource R and further detailed; and

A module 3E update, configured to update each detection event by the detection module 3D, one or more values used by the determination module 3B to determine the servers that can perform the computer task in using the actual usage measurements of the computing resource R obtained by the obtaining 3D module.

The functions of these various modules are described in more detail now, with reference to FIG. 3 (consisting of FIGS. 3A and 3B) and to the steps of the updating method and the method of managing computer tasks according to the invention, such as that they are implemented, in the first particular embodiment, by the device 3 for managing the cluster 2.

As illustrated in FIG. 3, the method of managing computer tasks according to the invention is based, in the first embodiment described here, on two processes that take place in parallel:

An asynchronous processing process PI, illustrated in FIG. 3A, which is triggered by the cluster management device 2 on receipt by its receiving module 3A of each request for execution of the computer task (s) on cluster 2. This IP process aims to assign each IT task computing resources on a server of cluster 2 to allow its execution; and

A P2 update process, illustrated in FIG. 3B, which is implemented here continuously by the cluster management device 3. More particularly here, the updates are aimed at values used by the device 3 of FIG. managing cluster 2 to select a server capable of executing a computer task entrusted to it. These updates are performed in the first embodiment on detection of different events belonging to a predetermined set of events. This set includes here three events, namely, the (periodic) escalation by the servers of the cluster 2 of measures of use of their resources, the positioning of tasks and the reservation of resulting resources on servers, and the release of resources on the servers. The values updated during the P2 process are preferably taken into account as soon as they are available by the PI process for processing requests received after each update.

We will now describe each of these processes in more detail. With reference to FIG. 3A, it is assumed that the management device 3 of the cluster 2 receives, via its communication means 9 and its reception module 3A, a REQ request for execution of a computer task T of the client device 4 (step E10). This request REQ comprises, as previously mentioned, a quantity R_R_T of resource R initially required so that the computer task T can execute properly, this quantity R_R_T having been predetermined by the user of the client device 4.

On receiving the request REQ (T, R_R_T), the management device 3 of the cluster 2, through its determination module 3B, determines here among the servers S1,..., SK of the cluster 2 which ones which are able to execute the task T, taking into account the needs R_R_T in resource R specified by the user. In other words, the determination module 3B determines which server (s) has a sufficient amount of resource R to execute the computer task T. It is noted that this determination does not need to be exhaustive, and a server list reduced to a single server may be sufficient.

For this purpose, for each server Sk of cluster 2, k = 1,..., K (initialization step E20, test step E50 then incrementation step E60), the determination module 3B determines whether the server Sk checks the following inequality (eql) (test step E30):

R_R_T + EU_R (Sk) <CL_R (Sk) (eql)

in which :

- CL_R (Sk) denotes the maximum capacity of the server Sk for the computing resource R (i.e. maximum quantity of resource R available on the server Sk). This maximum capacity is raised here periodically to the management device 3 by each server Sk; and

- EU_R (Sk) denotes an estimate of an already used quantity of the computing resource R on the server Sk.

If the server Sk checks the inequality (eql) (yes answer to the test step E30), it is identified by the determination module 3B as a server capable of performing the task T (step E40).

If not (answer no to the test step E30 and test step E50), the determination module 3B tests a new server among the servers that have not yet been tested in cluster 2 (answer no to test step E50 and step of incrementation E60).

In the first embodiment described here, once all the servers Sk, k = 1,..., K of the cluster 2 have been tested (yes answer to the test step E50), the selection module 3C selects a server SkO among the servers identified as being able to execute the computer task T taking into account its needs R_R_T R resource initially provided by the user (step E70). No limitation is attached to the criteria used by the cluster management device 2 to select the SkO server. This choice can be made for example compared to the other resources required for the execution of the task and available on the server SkO, with respect to geographical constraints, etc. Following this selection, the management device 3 of the cluster 2 orders (or triggers) the launch of the task T on the selected server SkO (step E80). In other words, it positions the task T on the server SkO and reserves on this server the quantity R_R_T of resource R for execution.

Cluster 2 management device 3 thus proceeds for each computer task submitted to it.

Note that the asynchronous process PI which has just been described with reference to FIG. 3A and which is implemented by the management device 3 of the cluster 2 approaches the processes implemented by the cluster management systems of the cluster. state of the art, and in particular by the Borg and Hadoop Yarn systems mentioned above. In the processing implemented by these systems of the state of the art, the estimate EU_R (Sk) of the quantity of resource R already used on the server Sk is taken equal, in the equation (eq l), to the quantity Q_R (Sk) of resource R initially reserved (ie allocated) by the management system on the server Sk for the execution of the tasks received previously and still being processed by the server Sk, that is to say always positioned on the server Sk (this number of tasks positioned is supposed equal to JO, JO designating an integer higher than 0). This quantity Q_R (Sk) is based on the declarations of needs R_R_Tj in resource R provided by the users for these previous tasks Tj, j = l, ..., J0, in other words, it is equal to the sum of the needs R_R_Tj, j = l, ..., J0 in resource R declared by the users at the origin of these J0 tasks Tj, j = l, ..., J0, again:

Q_R (Sk) = Σ ^J ° _{= 1} R_R_Tj.

In other words again, the management systems of the state of the art such as those mentioned above apply, for the allocation of a resource R to an incoming computer task T that is submitted to them, the following inequality (eq. l:

R_R_T + Q_R (Sk) <CL_R (Sk) (eql ')

According to the invention, the management device 3 of the cluster 2 differs from the management systems of the state of the art in that certain values used in the equation (eq l), and in particular the value EU_R (Sk and / or the limit capacitance value CL_R (Sk), are updated on detection of three predetermined events (defining a predetermined set of events in the sense of the invention) (step F10). These three events include:

Obtaining real (current) usage measurements of the computing resource R on the servers Sk, k = 1,..., K of the cluster 2 server;

The reservation by the cluster management device 3 for a computer task of a quantity of the computing resource R on a server Sk, k = 1,..., K of the cluster 2; and

The release by a computer task of a quantity of the computing resource R on a server Sk, k = 1,..., K of the cluster 2.

This update is illustrated in FIG. 3B described now (continuous process P2) for the quantity EU_R (Sk) (variable V within the meaning of the invention). Taking into account the three aforementioned events allows the management device 3 of cluster 2 to have more precise information on the amount of resource R available on each server when assigning computer tasks via the process Pl. Each designated event by EV detected among these three events causes an update of the current value of the variable V considered (here EU_R (Sk)). In the remainder of the description V (n), the current updated value of the variable V is denoted by n denoting an integer greater than or equal to 0 indexing the update (in other words, each detected event increments the integer n (initialization steps F00 then incrementing F20)). Thus, when this variable is EU_R (Sk), the current value of this variable is denoted EU_R (n, Sk). It is the same for the other variables considered in the description. It is noted that a so-called current value designates a value that is available or that is modified at a considered instant called current (that is to say at a moment in progress). In the case of a measurement, it may have been carried out at a previous time (especially in case of periodicity of the measurement); if so, the current measure is the measure that is available at the current time when it is used.

According to the invention, according to the detected event, a different update of the current value of the variable V is envisaged. We will detail now these three updates.

As indicated above, in the first embodiment described here, it is assumed that the feedback of the measurements carried out on the servers Sk, k = 1,..., K to the management device 3 of the cluster 2 is carried out periodically, according to a Tupd period. For the sake of simplicity, here we consider the same Tupd period for all servers in the cluster, but the invention also applies to different periods from one server to another. In addition, the invention also applies when the measurements are reported at predetermined times, not necessarily spaced regularly, or completely asynchronously.

This periodic obtaining of the actual utilization measurements of the resource R by the management device 3 is detected by the cluster management device 3 (yes response to step F10 with EV = UPD-U). Such a measurement can be performed by each server Sk using tools known per se, for example by means of a "top" command launched by the operating system of each server Sk periodically. U_R (n, Sk) is a measure of the actual use of the computing resource R carried out on the server Sk and available at the current time at the level of the management device 3 of the cluster 2.

Each of the measurements U_R (n, Sk), k = 1,..., K obtained by the cluster management device 2 2 is used by the cluster management device 3 as a measure of the actual current use of the cluster. computer resource R on the server Sk.

From this measurement, the management device 3 of the cluster 2, via its update module 3E, updates, in the first embodiment described here, the current value of the estimate EU_R (Sk) of the used amount of the R computing resource on the Sk server intended to be taken into account by the management device 3 during step E30 of the process PI (step F30).

In the first embodiment described here, the update module 3E calculates the current value EU_R (n, Sk) according to the following equation (eq2) (step F30):

EU_R (n, Sk) = (1 - a) EU_R (n-1 Sk) + ai / _i? (N, 5fe) (eq2) where a is a real weighting factor between 0 and 1. It is noted that in this equation, EU_R (0, Sk) is initialized to the quantity of the resource R reserved on the server Sk at a time t = 0 for all the tasks set on this server by the management device 3 of the cluster 2.

Note that the equation (eq2) makes it possible to update the estimate of the quantity already used of the computing resource R on the server Sk by calculating the value EU_R (n, Sk) at the current instant from a value Vpast (n) = EU_R (n - 1, Sk) computed (updated) at a previous instant (in this case here, during its previous update). It includes an exponential moving average parameterized by the factor. The closer the factor is chosen to 0, the longer the cluster management device 3 takes into account the amount of resources reserved and delays the time when it uses free resources of the servers that seem "unduly" reserved. On the contrary, the closer the factor is chosen to 1, the more quickly the cluster management device 3 forgets these resources reserved by the tasks, taking the risk that these resources are in fact necessary for the tasks for which they were reserved but later used. . The choice of the factor therefore results from a compromise between these different considerations.

The equation (eq2) advantageously makes it possible to adjust (ie update) the level EU_R (n, Sk) of use of the resource R used by the management device 3 of cluster 2 for the allocation of resources to IT tasks that are entrusted to it taking into account the actual use of resources on cluster 2 servers. This allows, if a task running on a Sk server in cluster 2 requires less R resource to run than the needs in this resource initially announced by the user in the request to execute this task, to allocate to other tasks of the resource R not used on the server Sk, or conversely to stop the allocation of resource R to incoming tasks if the Sk server no longer has an available R resource. This results in a better allocation by the resource management device 3 of the cluster 2, while ensuring a good performance of the tasks on the cluster.

In the first embodiment described here, the updating of the current value of the variable EU_R (Sk) by means of the equation (eq2) is carried out periodically, at each measurement feedback received by the device 3 of In other words, the event of obtaining measurements (EV = UPD-U) by the management device 3 is reproduced periodically.

In addition to this periodic update, according to the invention, other events (EV = RESA and EV = FREE in FIG. 3B) are capable of triggering updates. asynchronous day of the current value of the variable V = EU_R (Sk) for the different servers Sk, k = l, ..., K. Specifically, these asynchronous updates are triggered by the update module 3E each time it detects one and / or the other of the following two events for one of the servers Sk:

The reservation by the management device 3 of a quantity of the resource R on the server Sk, for the execution of a computer task Tnew (following the application of the process PI represented in FIG. 3A by the device 3 management on the task Tnew) (branch EV = RESA in Figure 3B). The amount of resource reserved for this task Tnew by the management device 3 on the server Sk corresponds to the needs R_R_Tnew R resource initially required by the user in the request for execution of this task transmitted to the device 3 management. In the remainder of the description, this event is designated by EV = RESA;

The release by a computer task Told of a quantity of the resource R on the server Sk, in particular due to the completion of this task Told or its branch interruption EV = FREE in FIG. 3B). In the remainder of the description, EV = FREE denotes this event.

For each concerned server Sk for which an event EV = RESA is detected by the management device 3 (via its 3D detection module) in association with a task Tnew (yes response to the detection step F10, branch EV = RESA) , the current value of the variable EU_R (Sk) is updated by the update module 3E according to the following equation (eq3-RESA) (step F40):

EU_R (n, Sk) = EU_R (n - l.Sk) + R_R_Tnew (eq3-RESA) Cluster 2 management device 3 is responsible for the positioning (assignment) of tasks on the various servers Sk of the cluster, it can inform without delay via its selection module 3C, the 3D detection module of the reservation of new resources on a server Sk.

For each concerned server Sk for which an event EV = FREE is detected by the management device 3 in association with a task Told (answer yes to the detection step F10, branch EV = FREE), the current value of the variable EU_R (Sk) is updated by the update module 3E according to the following equation (eq3-FREE) (step F50):

EU_R (n, Sk) = EU_R (n. - the Sk) - (1 - a) ^nold R_R_Told (eq3-FREE) where indexes nold the number of updates to EU_R performed on detection of UPD-U event ( new measurement returned to the device 3 for managing the actual use of the resource) from the moment when the quantity of resource R_R_Told was reserved for the task Told on the server Sk until its release, the quantity R_R_Told corresponding to the needs initially required R resource declared by the user in the execution request of the Told task addressed to the management device 3.

In equation (eq3) -FREE, the factor (1 -) ^nold R_R_Told takes account of the balance of resource R that is released by the Told task at the server Sk due to end its execution by the server Sk. It is noted that the taking into account of this balance is of special interest for tasks whose execution is of short duration (for example less than the period of Tupd update).

The asynchronous updates of the variable EU_R (Sk) thus carried out each time an event EV = RESA and / or EV = FREE are detected, are intended to take into account the tasks allocated to each server Sk by the device 3 for managing the cluster 2 and tasks that have been completed on each Sk server between two uploads of the server Sk (and therefore between two periodic updates of the variable EU_R (Sk)).

Note that, as mentioned above, the events EV = RESA can be easily detected by the management device 3 of the cluster 2, the latter being in charge of the allocation of resources to the computer tasks entrusted to it. The events EV = FREE are detected here by the servers Sk on which the Told tasks had been distributed (positioned) and sent back by these servers to the management device 3 of the cluster 2 (accompanied by the corresponding nold indices). The reception by the management device 3 of a notification from a server Sk of the completion of a task Told constitutes a detection of an event EV = FREE for this server Sk within the meaning of the invention. 3 The management device then has the information necessary to evaluate the factor (1 - a) ^nold R_R_Told.

The current value EU_R (n, Sk) updating the variable EU_R (Sk) is used as soon as it is available (ie as soon as it has been updated) by the determination module 3B of the device 3 for managing the cluster 2 as part of the PI process, and more precisely in the inequality (eql) for selecting a cluster 2 server capable of executing a new incoming task received as of this update.

The steps F10 to F50 of the process P2 are then reiterated each time an event EV = UPD-U, RESA or FREE is detected by the detection module 3D.

In a first variant of this first embodiment, the update module 3E uses in step F30 to calculate the current value EU_R (n, Sk) the equation (eq2 following instead of the equation ( eq2):

EU_R (n, Sk) = max ((l - a) EU_R (n - 1. Sfc) + aU_R (n, Sk ^~ ), U_R (n, Sk)) (eq2 ') where max (x, y) designates the function that provides the maximum value among x and y values. The update processes described in steps F40 and F50 remain otherwise identical to that described above.

This first embodiment makes it possible to take account more quickly of the real use of the resources on the server Sk.

In a second variant of the first embodiment illustrated in FIG. 3C, the updating module 3E uses to calculate the current value EU_R (n, Sk) the following equation (eq2 ") in place of the equation ( eq2) (step G30):

EU_R (n, Sk) = DR_R (n, Sk) + U_R (n, Sk) (eq2 ") where DR_R (n, Sk) denotes a smoothed current value of the amount of the computer resource reserved for at least one computer task by the management device 3 on the server Sk at the current time. The smoothed value DR_R (n, Sk) is calculated here by the updating module 3E during step G30, in the second variant, according to the following equation (eq4):

DR_R (n, Sk) = (1 -) DR_R (n - 1, Sk) (eq4)

where β is a real weighting factor between 0 and 1. It is noted that the choice of the weighting factor /? responds to a compromise similar to the one mentioned above for the choice of the weighting factor a (the speed of forgetting the reserved resources versus the actual observed use of these resources). In this equation, DR_R (0, Sk) is initialized to the quantity of the resource R reserved on the server Sk at time t = 0 for all the tasks set on this server.

It is assumed here that the value of the variable DR_R is periodically updated according to the same period Tupd as the feedback of the measurements to the management device 3.

However this assumption is not limiting in itself, and the update of the value of DR_R according to the equation (eq4) can be carried out according to another period, asynchronous with respect to the ascent of U_R (n, Sk). Each update of the value of DR_R (Sk) then causes an update of the value of EU-R (Sk). This is therefore another event (event EV = UPD-D shown in dashed lines in FIG. 3C), in the same way as the events EV = UPD-U, EV = RESA and EV = FREE described previously, susceptible in the second variant described here to be detected by the 3D detection module and to trigger an update of the current value of the variable V (= EU-R (Sk) here) (step G40).

In addition, in this second variant, during the detection of an event EV = RESA or EV = FREE, the value of DR_R is updated according to the equations (eq5-RESA) for the detection of an event EV = RESA (step G50) and (eq5-FREE) for detecting an EV event EV = FREE (step G60):

D_R (n, Sk) = D_R (n-1, Sk) + R_R_Tnew (eq5-RESA) and

D_R n, Sk) = D_R (n - l, Sk) - (1 -) ^nold R_R_Told (EQ5-FREE) where indexes nold the number of updates to D_R performed on detection of UPD-U event (new measure returned to the device 3 for managing the actual use of the resource) since the quantity of resource R_R_Told was reserved for the task Told on the server Sk until its release, the quantity R_R_Told corresponding to the initially required needs in resource R declared by the user in the execution request of the task Told addressed to the management device 3.

The current value of EU-R is then updated during the steps G50 and G60 according to the equation (eq2 ") where U_R (n, Sk) denotes the current value available at the device 3 for managing the measurement of actual use of resource R. In a third variant of the first embodiment, each measurement of the actual use of the computing resource R is performed on each server Sk of cluster 2, k = 1, ..., K, for each separate task Tj executed by this Sk server using the computer resource R, j = 0, ..., Jl, J l designating an integer greater than 0 counting the tasks positioned on the server Sk during the measurement. Here we limit ourselves to the tasks Tj executed by the server Sk and managed by the management device 3 of the cluster 2 (that is to say those which were launched on the server Sk on command 3 of the management of the cluster 2 after execution of the PI process).

U_R (n, Sk, j) is a measure of the actual use of the computing resource R carried out on the server Sk for the task Tj at the end of the time interval t (n-1). Each measurement U_R (n, Sk, j) carried out on a server Sk is sent at the end of the time interval t (n) to the management device 3 of the cluster 2, and obtained during the step F20 by the 3D module for obtaining the cluster management device 3. Each U_R measurement (n, Sk, j) is used by the cluster management device 3 as a measure of the actual use of the computer resource R on the Sk server by the task Tj for the time interval t (n).

From these measurements, the management device 3 of the cluster 2, via its update module 3E, evaluates in the third variant, during the step F30, an estimate EU_R (n, Sk, j) of the amount already used of the computing resource R on each server Sk by each computer task Tj managed by the management device 3 of the cluster 2 and executed on the server Sk using the following equation (eq6):

EU_R (n, Sk, j) = (1 - a) EU_R (n - 1. Sk.j) + U_R (n, Sk, j) (eq 6) or the following equation (eqô:

EU_R (n, Sk, j) = max ((l - a) EU_R (n - l.Sk.j) + aU_R (n, Sk, j), U_R (n, Sk, jj) (eq 6 ') in which is a real weighting factor between 0 and 1, chosen as described previously in the other variant embodiments It should be noted that in this third variant, EU_R (n0j, j) is initialized to the quantity R_R_Tj of computing resource R initially required. for the execution of the task Tj, nOj denoting the updating index corresponding to the allocation of the computer task Tj to the server Sk by the management device 3.

Then, in the third variant described here, during the step F30, the update module 3E evaluates from the values EU_R (n, Sk, fi obtained for each task Tj the quantity EU_R (n, Sk) according to the following equation (eq2 '"):

Jlact

EU_R (n, Sk) = ^ EU_R (n, Sk, j) (eq2 "')

7 = 0

or according to the following equation (eq2 ⁽⁴⁾ ):

EU_R (n, Sk) = max (Σ ^J _{j =} ^1a ₀ ^ct EU_R (n, S k, j), Σ ^ 'U_r (n, S k, j)) (eq2 ⁽⁴⁾⁾ wherein Jlact denotes the number of tasks set on the server Sk (that is, allocated to the server Sk by the management device 3) among the Jl tasks for which measurements have been reported to the management device 3. In the first embodiment and in its three variants previously described, it is the management device 3 of the cluster 2 which, from the measurements sent by the servers Sk, k = 1,..., K of the cluster 2, updates the value of the quantity EU_R (Sk) used by the module 3B for determining the cluster management device 3 to select a server of the cluster to execute an incoming task.

In a second embodiment, each server Sk of cluster 2 is a server according to the invention. He understands :

A measurement module configured to periodically measure the actual use of its computing resource R;

A detection module configured to detect a predefined set of events (UPD-U, RESA and FREE); and

An update module configured to update at least one current value of a variable rated V selected from an estimate of an already used quantity of the computing resource on the server and a maximum capacity of the server for the computing resource , said update module being activated on each detection by the detection module of an event among the predefined set of events and configured to:

More specifically, in the second embodiment, on each detection of an EV event (EV = UPD-U, RESA, FREE or UPD-D as in the first embodiment) each Sk server estimates via its update module. day the current amount EU_R (n, Sk) already used of its computer resource R. It can proceed for this purpose for example similarly or identically to what has been described previously in the first embodiment and implemented by the management device 3 with reference to FIGS. 3B and 3C.

Then, from the quantity EU_R (n, Sk) thus estimated by the server Sk, the update module of the server Sk updates the value of its maximum capacity CL_R (Sk) for the resource R and for the time interval t (n) according to the following equation (eq7): CL_R (n, Sk) = CL_R (Sk) + Q_R (n, Sk) - EU (n, Sk) (eq7) where Q_R (n, Sk) denotes the quantity of resource R initially reserved by the management device 3 on the server Sk for the execution of the tasks positioned on the server Sk at the current time of update. This quantity Q_R (n, Sk) is given by, taking into account the needs R_R_Tj, in resource R declared by the users at the origin of these tasks:

Q_R (n Sk) = R_R_Tj.

where J2 (n) denotes the number of tasks Tl, ..., TJ2 (n) positioned on the server Sk at the current update time.

The current value of the maximum capacity CL_R (n, Sk) of the server Sk for the resource R is then raised (periodically for example here) by the server Sk to the management device 3 of the cluster 2, for example using the communication protocol already used in computer systems based on a Hadoop Yarn management system.

The management device 3 of the cluster 2 then uses the updated value CL_R (n, Sk) of the maximum capacity CL_R (Sk) of the server Sk in the equation (eql ') mentioned previously to determine, for an incoming task T it is subject to the servers Sk, k = l, ..., K of the cluster 2 capable of performing this task.

It should be noted that this second embodiment does not require any modification of the processing performed today by the management systems of the state of the art (ie they use the equation (eql). In this embodiment, the cluster management device 2 can be a Hadoop Yarn management system or a Borg management system already known, only the cluster 2 Sk servers must be modified to implement the invention. Cluster 2 servers.

In the two embodiments described here, as well as in their variants, it has been considered to update the quantity EU_R (n, Sk) an exponential moving average. However, this assumption is not limiting in itself, and other types of smoothing or sliding averages can be considered to calculate the quantity EU_R (n, Sk). For example, in an alternative embodiment, we can consider an arithmetic average of actual usage measurements sent back to a sliding window comprising N time intervals preceding the time interval t (n).

In yet another variant, a weighted moving average MMP (n, Sk) may be considered of actual usage measurements sent back to a sliding window comprising N time intervals preceding the current update time, such as for example:

R (n, Sk) + (N'-1) u R (n-1, Sk) + --- + U R (n-N '+ 1, Sk)

MMP (n, Sk) = - ^J - ^ -, - - -

J N '+ QV'-1) + --- + 1

where N 'is the maximum of the number of measurements U_R (n, Sk) available preceding measurement U_R (n, Sk) and of N. This average MMP (n, Sk) is written in an equivalent way, after obtaining N measurements. U_R (m, Sk), in the form: MMP (n, Sk) = MMP (n-1, 57c) +; - -

J ^J QV + l) / 2

It should be noted that in this weighted average, the U_R (n, Sk) measures are all the more important as they are recent.

The value EU_R (n, Sk) can then be calculated according to the following equation, on receipt of a new measurement U_R (n, Sk) (detection of an event EV = UPD-U):

"(K + N - n) * R_R_Tj

EU_R (n, Sk) = MMP (n, Sk) +>> - ' _T

^ - ', N

k = nN +1 ^to l- ^k

dj> n

where aj indexes the beginning of the task Tj and already the end of the task Tj on the server Sk.

The value EU_R (n, Sk) may further be updated as follows when detecting an event EV = RESA:

EU_R (n, Sk) = EU_R (n, Sk) + R_R_Tnew

These examples are given for illustrative purposes only and other examples may be considered to update the EU_R (n, Sk) value.

Claims

1. Method for updating, for at least one server (Sk) of a server cluster, at least one current value of a variable V denoted from an estimate of an already used quantity of a resource computing on the server or a maximum capacity of the server for said computing resource, said at least one current value being intended to be taken into account by a server cluster management device to select, upon receipt of a execution request d a computer task associated with a predetermined quantity of said computer resource (R) initially required for its execution, a server said target of the cluster to perform this computer task, said method being such that said update is performed at each detection (F10 ) an event (EV) from a predefined set of events and:

If the detected event (UPD-U) is obtaining a measurement (U_R (n)) of a current real use of the computing resource on the server, said updating (F30, G30) of the value current variable V uses said measure of the actual actual usage of the computing resource on the server;

If the detected event (RESA) is a reservation by the management device for a computer task of a quantity of the computer resource on the server, said update (F40, G50) of the current value of the variable V uses said amount of the reserved computing resource; and

If the detected event (FREE) is a release by a computer task of a quantity of the computing resource on the server, said updating (F50, G60) of the current value of the variable V uses said quantity of the released computer resource.

The updating method according to claim 1, wherein if the detected event (UPD-U) is obtaining the measurement of the actual current usage of the computing resource on the server, said update of the The current value of the variable V also uses at least one value of the variable V noted Vpast updated at a previous time.

3. An updating method according to claim 2 wherein if the detected event (UPD-U) is obtaining the measurement of the actual current use of the computer resource on the server, during said update. the actual current usage metric of the computing resource is weighted by a real factor a between 0 and 1, and said at least one Vpast value is weighted by a factor equal to (1-and).

The updating method according to claim 3 wherein the updated variable V is the EU_R estimate of the amount already used of the computing resource on the server, and if the detected event is obtaining the measure. the actual actual use of the resource on the server, said current value denoted EU_R (n) of variable V is updated according to the equation:

EU_R (n) = max ((l - a) EU_R (n - 1) + aU_R (n), U_R (n)) where n denotes an integer greater than 1, EU_R (n - 1) is the value of estimating the previously used amount of the computer resource on the server updated at a previous time, and U_R (n) is the measurement obtained from the actual current usage of the computing resource on the server.

The updating method of claim 1, wherein said measuring the actual actual usage of the computing resource is a measure of the actual use of the computing resource on the server by one or more computing tasks executing. on the server and all managed by the cluster management appliance.

The updating method according to claim 5 wherein the updated variable V is the EU_R estimate of the amount already used of the computing resource on the server, and if the detected event is obtaining the measure the actual current usage of the IT resource on the server:

The measurement of the current real use of the computing resource on the server comprises a plurality of values denoted U_R (n, f 1, j = 1, ... J, associated respectively with a plurality of computer tasks Tj, j = l , ... J, set by the cluster management device on the server, where J and n denote integers greater than or equal to 1, each value

U_R (n, fi corresponding to a measurement of the actual use of the computer resource on the server by the task Tj;

Said update (F30) comprises:

EU_R (n, j) = max ((l - a) EU_R (- 1,;) + U_R (n, j), U_R (n, j)) or according to the equation:

EU_R (n, j) = (1 - a) EU_R (n - 1,;) + U_R (n, j)

or :

EU_R (- lj) designates a value evaluated at a previous instant of the estimation of the quantity already used of the computing resource on the server by this computing task Tj; and

■ is a real factor between 0 and 1;

a second step of evaluating the current value denoted EU_R (n) of the estimate of the quantity already used of the computing resource on the server according to the equation: EU_R (n) = ^ EU_R (n, j) or according to the equation: EU_R (n) = max U_R (n, j)

The updating method according to claim 1, wherein the updated variable V is the EU_R estimate of the amount already used of the computing resource on the server, and if the detected event is obtaining the measure. of the actual current use of the computing resource on the server, said current value denoted EU_R (n) of the variable V is updated according to the equation:

EU_R (n) = U_R (n) + DR_R (n)

where U_R (n) denotes a current measure of the actual use of the computing resource on the server, and DR_R (n) denotes a current value of a quantity of the computing resource reserved for at least one computing task by the computing device. management on the server.

8. Update method according to claim 7, further comprising: a step of updating the current value DR_R (n) of the quantity of the computer resource reserved for at least one computer task by the management device; the server is obtained according to the equation:

DR_R (n) = (1 -) DR_R (n - 1)

A step of detecting (G10) the update (UPD-D) of the current value DR_R (n); and

A step of updating (G40) the current value of the variable EU_R according to the equation:

EU_R (n) = U_R (n) + DR_R (n).

The updating method of claim 7 or 8 further comprising:

If the detected event is a reservation by the management device for a computer task of a quantity R_R_Tnew of the computer resource on the server, a step of updating the current value DR_R (n) of the quantity of the computer resource reserved for at least one computer task by the management device on the server using said quantity R_R_Tnew and a quantity DR_R (n - 1) of the computer resource reserved by the management device on the server obtained at a previous instant,

and wherein said update of the current value of the variable EU_R uses said amount of the reserved computer resource according to the equation:

EU_R (n) = U_R (n) + DR_R (n).

10. A method for managing computer tasks comprising, on reception (E10) by a device (3) for managing a server cluster (2) (S1, S2, ..., SK) of a request (REQ ( T, R_R_T)) execution of a computer task (T), this computer task being associated with a predetermined quantity (R_R_T) of a computer resource (R) initially required for its execution:

A step of determining (E40), by the cluster management device, at least one server of the cluster capable of executing the computer task, this determining step using the predetermined quantity of the computer resource initially required for the execution the computing task, a maximum server capacity for the computing resource and an estimate of an already used amount of the computing resource on the server;

A step of selecting (E70), by the cluster management device, a server for executing the computer task among said at least one server of the determined cluster;

said method being characterized in that, for at least one server of the cluster, at least one current value of a variable rated V selected from the estimate of the amount already used of the computing resource on the server and the maximum capacity of the server for the computer resource is updated according to an updating method according to any one of claims 1 to 9.

11. Management method according to claim 10 wherein the update process is implemented by the device (3) for managing the cluster.

12. Management method according to claim 10 wherein the updating method is implemented by said at least one server (Sk), said management method further comprising a step of transmitting said current value of the variable V updated by said at least one server to the cluster management device.

Computer program (PROG) comprising instructions for performing the steps of the updating method according to any one of claims 1 to 9 or instructions for carrying out the steps of the management method according to the claim When said program is executed by a computer.

14. A computer-readable recording medium on which is recorded a computer program comprising instructions for performing the steps of the updating method according to any one of claims 1 to 9 or instructions for the execution of the steps of the management method according to claim 10.

15. Device for managing (3) a server cluster (2) (S1, ..., SK) comprising:

A receiving module (3A);

A module (3B) for determining, activated on reception by the reception module of a request for execution of a computer task, said computer task being associated with a predetermined quantity of a computer resource initially required for its execution, the determination module being configured to determine at least one cluster server capable of performing the computing task using the predetermined amount of the IT resource initially required for the execution of the computing task, a maximum server capacity for the computing resource and an estimate of an already used amount of the computing resource on the server;

A selection module (3C) configured to select a server to execute the computer task among said at least one server of the cluster determined by the determination module;

said management device being characterized in that it further comprises:

A detection module configured to detect a predefined set of events; and

o if the detected event is a reservation by the management device for a computer task of a quantity of the computing resource on the server, update the current value of the variable V using said amount of the reserved computer resource; and

o if the detected event is a release by a computer task of a quantity of the computing resource on the server, updating the current value of the variable V using said amount of the released computing resource.

16. Server (Sk) belonging to a cluster (2) of servers managed by a management device (3), said server being able to execute at least one computer task managed by said management device, said at least one computer task using when executing at least one computer resource of the server, said server comprising:

A detection module configured to detect a predefined set of events; and

Computer system (1) comprising:

A cluster (2) of servers comprising a plurality of servers, each server comprising at least one computer resource allowing the execution of computer tasks;

A device (3) for managing the server cluster;

wherein said cluster management device is in accordance with claim 15 and / or at least one server of the cluster is in accordance with claim 16.