(19)
(11)EP 3 502 890 A1

(12)EUROPEAN PATENT APPLICATION

(43)Date of publication:
26.06.2019 Bulletin 2019/26

(21)Application number: 17210405.1

(22)Date of filing:  22.12.2017
(51)International Patent Classification (IPC): 
G06F 9/50(2006.01)
(84)Designated Contracting States:
AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR
Designated Extension States:
BA ME
Designated Validation States:
MA MD TN

(71)Applicant: Bull SAS
78340 Les Clayes-sous-Bois (FR)

(72)Inventors:
  • CADEAU, Thomas
    91300 MASSY (FR)
  • RAJAGOPAL, Dineshkumar
    600126 Chennai - Tamilnadu (IN)

(74)Representative: LLR 
11, boulevard de Sébastopol
75001 Paris
75001 Paris (FR)

  


(54)METHOD FOR MANAGING RESOURCES OF A COMPUTER CLUSTER BY MEANS OF HISTORICAL DATA


(57) This method for managing resources of a computer cluster (1) involves automated means (2, 7, 9) to allocate to a job (4) at least one resource (3) among several resources (3) from the cluster (1), depending on at least one historical data (8) relative to at least one previous use of the resources (3).




Description


[0001] The invention deals with cluster computing. In particular, it deals with the management of resources in a computer cluster.

[0002] It is known to provide a computer cluster with a resource and job management system. The role of the system is to allocate the resources of the cluster to the jobs that it has to handle. To manage both resources and jobs in order to optimize the use of the cluster, the system obtains information concerning the nodes of the cluster like its topography and which nodes are free to be used, and the descriptions of the jobs that the cluster has to handle like the number of central processing units needed to perform the job.

[0003] Yet, it is observed that such a system leads to an increase of failures of the resources of the cluster over the time, such that the availability of the service provided by the cluster may be discarded. Furthermore, such failures generate loss of time and important costs to replace components of the cluster.

[0004] One object of the invention is to increase the lifetime of a computer cluster, in particular the lifetime of the resources provided by the cluster.

[0005] For this purpose, it is provided according to the invention a method for managing resources of a computer cluster, wherein automated means allocate to a job at least one resource among several resources from the cluster, depending on at least one historical data relative to at least one previous use of the resources.

[0006] Thus, by studying data relative to previous uses of resources, the automated means determine the wear of the resources in order to choose the resource with the lowest wear to perform the job. Thus, thanks to these historical data, automated means avoid the over-utilization of resources, which is the main cause of failures in a computer cluster. The lifetime of the cluster is thus increased.

[0007] Preferably, automated means generate the historical data during the previous use and record the historical data in a database of historical data.

[0008] Thus, when a resource is selected to perform a task, this use and/or data relative to this use is registered in the database. When the system has to allocate a resource to a task, it thus knows to which extent the resource has been utilized for previous tasks, and thus how is its wear.

[0009] Advantageously, the historical data is relative to a use rate of a resource or to a log concerning a use of a resource.

[0010] Thus, the automated means may compute a use rate in order to compare the use rates of free resources before allocating a resource to a job. Furthermore, a log concerning a use of a resource may be reported after using this resource and may provide information to the automated means.

[0011] Preferably, the allocating step is further performed depending on a cluster topology.

[0012] Thus, the system takes the topology of the cluster into account in addition to the wear of the resources. Indeed, some topologies involve different allocating strategies, depending for example on the redundancy or links between the nodes of the cluster.

[0013] Advantageously, the allocating step is further performed depending on a predetermined resource management strategy.

[0014] Thus, an external user may select a resource management strategy or policy, which is taken into account by the automated means.

[0015] Preferably, the allocating step is further performed depending on at least one data relative to the job.

[0016] Thus, the data relative to the job may deal with several features, for example with a number of tasks to perform the job, a number of central processing units or a quantity of memory or bandwidth needed, and/or a number of nodes needed.

[0017] Advantageously a sensor situated in the cluster obtains a data and transmits the data to the automated means, the allocating step being further performed depending on the data.

[0018] Preferably, the data relates to a temperature of the cluster.

[0019] Thus, if a component is too hot, the system may choose to allocate another component to the job in order to avoid overheating the component.

[0020] Advantageously, the allocating step is further performed depending on at least one data relative to a maintenance of the cluster.

[0021] Thus, a user may register the date of the last maintenance of a resource. The system may take this date into account to limit the number of uses of this resource if the date is ancient, and may allocate recently maintained resources to the job as the risk of failure is lower when a maintenance is recent.

[0022] Preferably, automated means determine relative ages of at least two components of the cluster depending on at least one of the data recited above.

[0023] Thus, with data obtained by the system, this latter may determine the wear and thus the relative ages of components of the cluster when compared to each other. By "relative ages", we mean that the goal is to compare the ages of components in respect with each other, not necessary the absolute age of a component. An age could be related to several values like a number of cycles of a component, a duration since implementation in the cluster, or a duration of active utilization of a resource. Furthermore, it is also possible to compute an absolute age, that is to say the age of a component without comparing it to other components.

[0024] Advantageously, automated means determine a conclusion concerning the relative ages of the components and emit a message comprising this conclusion.

[0025] Thus, a message to an external user may be sent.

[0026] Preferably, the conclusion is relative to a workload of the cluster, to a maintenance date of the cluster, to a temperature of the cluster, to a power consumption of the cluster, and/or to a bandwidth of the cluster.

[0027] Thus, the message may warm a user that the computed age of the resource/component is due to one of these indications.

[0028] Advantageously, by means of at least one of the data recited above, automated means determine at least one action to be taken in order to improve the managment of the resources of the cluster.

[0029] Thus, the system suggests an action to an external user who may decide to take it or not. Or the system may be configured to take it itself automatically.

[0030] Preferably, the automated means further determine the action by means of at least one data relative to at least one effect of at least one action previously taken.

[0031] Thus, taken actions generate effects which may be registered in a database in order to be studied by automated means or by an external user. The efficiency of the action may thus be measured in order to select it later again or not.

[0032] Advantageously, the action to be taken is relative to shutting down a component of the cluster, substituting a component of the cluster and/or sending a message.

[0033] Preferably, the method according to the invention is implemented as part of a high performance computing method.

[0034] Indeed, high performance computing needs one or several computer clusters with an optimized management of the resources and a maximal lifetime in order to be efficient.

[0035] According to the invention, it is also provided a computer program, comprising code instructions able to control the implementation of steps of a method recited above when executed on a computer.

[0036] According to the invention, it is also provided a method for making the program of the preceding claim available for downloading over a telecommunications network.

[0037] According to the invention, it is also provided a device for managing resources of a computer cluster, configured to allocate to a job at least one resource among several resources of the cluster depending on at least one historical data relative to at least one previous use of the resources.

[0038] Preferably, the resource is a node, a server, a bandwidth, a power, a memory such as a ram or a flash memory, or a component such as a processor, a connector or a data storage medium.

[0039] According to the invention, it is also provided a system comprising a device according to the invention and a computer cluster.

[0040] According to the invention, it is also provided a database including an electronical support comprising, in registered form, at least one historical data relative to at least one use of at least one resource of a computer cluster.

[0041] We are now going to introduce two embodiments of the invention provided as non-limiting examples and in support of the appended figures on which :
  • figure 1 illustrates a first embodiment of the invention ; and
  • figure 2 illustrates a second embodiment of the invention.


[0042] As illustrated on figure 1, a system according to the invention comprises at least a computer cluster 1 and a device 2 for managing resources of this cluster. A computer cluster is, as generally defined, an assembly of several servers whose resources are combined to behave like one single installation, in order to increase the speed of computing by parallel computing and allocating tasks to adapted resources. The computer cluster 1 includes several servers 3 arranged according to a predetermined topology. Here, four servers are illustrated, some of them being directly linked, other being indirectly linked.

[0043] In a computer server, one server corresponds to one node, which is a term known from a skilled person. But one node may also be considered as a part of a server or as several servers arranged together. In this description, for a better understanding, one node corresponds to one server 3.

[0044] A server 3 comprises at least one resource which may be used by the cluster 1. It could be any resource that a computer includes, like any type of memory, such that a flash memory or a volatile memory via any data storage medium such as a hard disk, a bandwidth, a power of computing via processors such as central processor units, connectors to connect components between them. Even a server or node itself may be considered as a resource. To be more general, in this invention, any component or capacity, known by a skilled person as a resource of any computer cluster, is considered as a resource of the computer cluster 1. Thus, all servers 3 are or comprise one or several of these resources, and each server 3 could comprise resources different from each other.

[0045] The computer server 1 is connected to the resource and job management device 2. Alternatively, this device 2 could also be considered as a part of the computer cluster 1.

[0046] This device 2 comprises automated means to receive external data and to provide decisions. The type of these automated means will be deduced by the skilled person and could be for examples server, connections, databases or processors. A member of this device is the decisional unit 9 that comprises processors and emits the decisions depending on all received data. In other terms, this entity makes the resource selection. The decisions are to allocate the resources of the cluster 1 to the jobs to be executed. Indeed, when a cluster 1 has to perform an important computing, this computing is divided between jobs, wherein jobs may be themselves divided between tasks. The device 2 obtains here the knowledge of the resources that are needed to perform the job 4. How these information are obtained is not concerned by the invention and is fully known by a skilled person. The description 5 of the job 4 includes several factors like the number of processors needed, the extent of the bandwidth needed, the total of nodes 3 needed, and could include any type of any other resource needed to perform the job 4. Then, the device 2 has to allocate the resources of the cluster 1 to the job 4. This part will be described later.

[0047] The device 2 also obtains data of configuration from a table 6. These data include the type of management strategy determined for example by an external user. Indeed, such a user may decide to prioritize some resources with respect to other resources and thus could select a specific strategy among several suggested strategies. The table 6 also includes data relative to the topography of the cluster 1. Indeed, depending on the topography, the resources are more or less closed to each other, redundancy of components is more or less present and thus allocating the resources to the job 4 may vary depending on these factors. This table 6 may also contain other data relative to the configuration of the cluster 1, like the type of queue and all data relative to the configuration that are obvious for a skilled person.

[0048] The device 2 also obtains data from a predictive engine 7. This engine is a device that gathers historical data and deduces from them factors to be taken into account by the system 2 when allocating the resources to the jobs. In this embodiment, the predictive engine 7 thus obtains historical data from a database 8, these data being related to the use of resources of the cluster 1. Thus, the predictive engine gathers these historical data relative to previous use and computes data relative to the wear of the resources concerned by the historical data. Indeed, by knowing which resource has been used, how many time, to which extent and during which duration, the predictive engine 7 may compute the wear and thus an age of the resource, and compare the ages of all the resources of the cluster. These ages, also called relative ages, or wear, are thus utilized by the device 2 as a supplementary factor to be taken into account when allocating the resources to jobs and tasks. It helps avoiding overutilization of some resources and thus it allows to decrease the number of failures of resources.

[0049] Finally, in this embodiment, the nodes 3 of the cluster 1 are linked to the database 8 of historical database. Thus, when the nodes comprise resources that are used to perform a job/task, the nodes or the resources may make reports, or logs concerning their use. These reports or logs could comprise a variety of data, from a single indication that the resource has been used to a detailed report of the use with problems encountered, number of cycles, durations, etc. These data could also refer to statistics like a use rate of a resource. These historical data are then computed by the predictive engine 7 and used by the device 2 to allocate the resources in order to optimise the selection of resources and thus to increase the lifetime of the resources and of the cluster 1.

[0050] To finish, it has to be noted that the cluster 1 and its related devices such as device 2 are advantageously a part of a supercomputer to make high performance computing. Indeed, such type of computing particularly needs a cluster with a good lifetime and an efficient management of its resources as high performance computing is very consumptive and wearing.

[0051] We are now going to introduce a description of a method according to the invention with reference to the embodiment illustrated on figure 1.

[0052] The cluster 1 is tasked with computing high calculation. To perform this calculation, the job 4 needs some nodes 3 to be executed, these nodes providing resources asked by the descriptions 5 of the job 4. To select the resources and thus the nodes to perform the job 4, device 2 takes into account the table 6 with configuration data, and the historical data of the database 8.

[0053] Thus, the device 2 determines which are the nodes 3 that are free. Indeed, some nodes may be already taken to perform other calculations. Then, among these free nodes, the predictive engine 7 reports the ages of each nodes that have been computed thanks to the historical data. Device 2, thanks to its decisional unit 9, then allocates the youngest nodes 3 to perform the jobs 4.

[0054] To conclude, nodes 3 perform the jobs or tasks and report data relative to the use of their resources into the database 8.

[0055] Predictive engine 7 uses these historical data to update the age of the resources of cluster 1, in order to allow the device 2 to allocate the updated youngest nodes for the next calculation.

[0056] It has to be noted that, depending on the configuration and choices of a user, the device 2 may chose resources that are not necessarily the youngest. Indeed, the configuration in table 6 could ask the device 2 to weight the ages of the predictive engine 7 in order to take them into account less or more than other factors.

[0057] Furthermore, the configuration may allow the device 2 to deallocate resources to make them available to new jobs. Thus, all the ages of the resources may be compared, whether they are free or not.

[0058] To sum-up, the use of historical data thus allows to compute ages of the resources in order to avoid overutilization of the resources when allocating them to jobs.

[0059] We now are going to introduce a second embodiment of the invention, with reference to figure 2.

[0060] This embodiment comprises the same component as the embodiment of figure 1. In particular, the topology of the cluster 1 is the same, the predictive engine 7 plays the same role, and the historical data of the database 8 are updated as described above.

[0061] The difference in respect with the first embodiment is the presence of two new entities: the external interface 10 and the table of messages 11.

[0062] The external interface 10 allows a user 12 to enter other data, like the date of maintenance of the cluster 1, or the date of maintenance of one of the components or resources of the cluster 1. Connected to the device 2, this latter may thus determine the last date of maintenance of each resource and use this factor when allocating the resources to the jobs to be executed. This interface 10 may also allow the user to enter other types of administrative data, which he may consult or which could be used by the device 2 when allocating the resources. This interface 10 may be viewed and implemented simply as a table 6 with supplementary data registered.

[0063] The table of messages 11 allows the device 2 to send messages to an external user 12. It may send all messages or information that the user wishes to know about the system, and it particularly allows to explain to the user the ages of the resources that have been computed. Thus, device 2 could determine a main reason for the age of the resource, like an important workload of some resources of the cluster, a maintenance date of the cluster that would be ancient, a temperature of a component that would be hot, or an important power consumption of a component. It may also relate to a bandwidth of a component of the cluster, like a lack of bandwidth or debit. The user 12 is thus aware of the state of each resource of its cluster 1. It may allow him to efficiently decide of a different configuration data to enter or to use another cluster if necessary. For this purpose, it is clear that in this embodiment, the cluster 1 comprises sensors and/or analyse tools, like a temperature sensor situated in a specific component or part of a server 3 or of the cluster 1, or in each server 3 or several servers if necessary.

[0064] The table of messages 11 also allows to suggest actions to the user. Indeed, in this embodiment, the device 2 may deduce from the computed relative ages actions that could be taken to improve the efficiency of the cluster 1. For example, these actions may be to shut down a component of the cluster 1, substitute a component of the cluster and/or sending a message to the system, for example to change a parameter of the configuration via the table 6 or the interface 10. This message may also be an e-mail or a sms alert to be sent to other systems.

[0065] If the external user 12 decides to execute an action that the device 2 suggested via the table 11, the effect of its action is analyzed by the device 2. For example, if the user decides to replace a processor unit of a server 3 by another, the effect of this change may be considered, for example via the historical data when the resource is used. If the historical data concerning this resource makes it clear, via a log or report, that less problems are encountered when the resource is used in a job, the device 2 deduces that the replacement of the resource was efficient.

[0066] This effect may be reported to the user. Above all, this effect is registered by the device 2, in any database, such that if a problem similar to the problem encountered and solved by this action appears, the device 2 will suggest the same action to the user 12.

[0067] Furthermore, the device may be configured to execute itself any action that it determined, such that shutting down a component or change a parameter in a configuration of a component of the cluster 1.

[0068] The steps of a method to be performed on this embodiment are the same as the steps described above in respect with the first embodiment. But for this embodiment, as explained, the decision unit 9 decides the allocation of resources with other factors, in addition to the age of the resources computed by predictive engine 7.

[0069] To sum up, we introduced here a method for managing resources of a computer cluster 1, wherein automated means, in this case a device 2 comprising a unit 9 and with the help of others devices, allocate to a job 4 at least one resource among several resources from the cluster 1, depending on at least one historical data 8 relative to at least one previous use of the resources. This method allows to increase the lifetime of the cluster by avoiding the overutilization of resources.

[0070] Among other possible data, in addition to historical data, to be considered by the automated means when executed, we may refer to :
  • a cluster topology,
  • a determined resource management strategy for example selected by an user,
  • a data relative to the job 4 to be executed, like the type of resources that the job needs,
  • a data obtained from at least one sensor situated in the cluster 1, for example a temperature sensor in a processor of a server 3, the data being transmitted to the automated means,
  • a data relative to a maintenance of the cluster.


[0071] At least with the historical data, the device is able to compute an age of a resource and to select resources to be allocated depending on their age.

[0072] The invention is not limited to the embodiments presented and other embodiments will become apparent to those skilled in the art.

[0073] It is especially possible to combine the two described and illustrated embodiments. Thus, the device 2 could combine the factors of the historical data and of the interface 10 only to allocating the resources, or to add other data, depending on the presence or not of different sensors in the cluster or on data registered by the user, for example.

[0074] Furthermore, it is possible to imagine other types of data to help the device 2 with allocating the resources to the jobs, in order to improve again the lifetime of the cluster.


Claims

1. A method for managing resources of a computer cluster (1), wherein automated means (2, 7, 9) allocate to a job (4) at least one resource (3) among several resources (3) from the cluster (1), depending on at least one historical data (8) relative to at least one previous use of the resources (3).
 
2. A method according to the previous claim, wherein automated means (2, 7, 9) generate the historical data (8) during the previous use and record the historical data in a database of historical data.
 
3. A method according to any one of the previous claims, wherein the historical data (8) is relative to a use rate of a resource (3) or to a log concerning a use of a resource (3).
 
4. A method according to any one of the previous claims, wherein the allocating step is further performed depending on a cluster topology.
 
5. A method according to any one of the previous claims, wherein the allocating step is further performed depending on a predetermined resource management strategy (6).
 
6. A method according to any one of the previous claims, wherein the allocating step is further performed depending on at least one data (5) relative to the job (4).
 
7. A method according to any one of the previous claims, wherein a sensor situated in the cluster obtains a data and transmits the data to the automated means, the allocating step being further performed depending on the data.
 
8. A method according to the previous claim, wherein the data relates to a temperature of the cluster (1).
 
9. A method according to any one of the previous claims, wherein the allocating step is further performed depending on at least one data (10) relative to a maintenance of the cluster (1).
 
10. A method according to any one of the previous claims, wherein automated means (2, 7, 9) determine relative ages of at least two components (3) of the cluster (1) depending on at least one of the data (5, 6, 8,10) recited in anyone of the previous claims.
 
11. A method according to the previous claim, wherein automated means (2, 7, 9) determine a conclusion concerning the relative ages of the components (3) and emit a message comprising this conclusion.
 
12. A method according to the previous claim, wherein the conclusion is relative to a workload of the cluster (1), to a maintenance date of the cluster, to a temperature of the cluster, to a power consumption of the cluster, and/or to a bandwidth of the cluster.
 
13. A method according to any one of the previous claims wherein, by means of at least one of the data (5, 6, 8,10) recited in any one of the previous claims, automated means (2, 7, 9, 11) determine at least one action to be taken in order to improve the managment of the resources of the cluster.
 
14. A method according to the previous claim, wherein the automated means (2, 7, 9, 11) further determine the action by means of at least one data relative to at least one effect of at least one action previously taken.
 
15. A method according to any one of claims 13 to 14, wherein the action to be taken is relative to shutting down a component of the cluster (1), substituting a component (3) of the cluster and/or sending a message.
 
16. A method according to any one of the previous claims, implemented as part of a high performance computing method.
 
17. A computer program, comprising code instructions able to control the implementation of steps of a method according to any one of the previous claims when executed on a computer.
 
18. A method for making the program of the preceding claim available for downloading over a telecommunications network.
 
19. A device (2) for managing resources (3) of a computer cluster (1), configured to allocate to a job (4) at least one resource among several resources of the cluster depending on at least one historical data (8) relative to at least one previous use of the resources.
 
20. A device (2) according to the previous claim, wherein the resource is a node, a server, a bandwidth, a power, a memory such as a ram or a flash memory, or a component such as a processor, a connector or a data storage medium.
 
21. A system comprising a device (2) according to any one of claims 19 to 20 and a computer cluster (1).
 
22. A database (8) including an electronical support comprising, in registered form, at least one historical data relative to at least one use of at least one resource of a computer cluster.
 




Drawing










Search report















Search report