Energy efficient supercomputer job allocation

Patent metadata

Field	Value
Publication number	US-2016188375-A1
Application number	US-201414582297-A
Country	US
Kind code	A1
Filing date	Dec 24, 2014
Priority date	Dec 24, 2014
Publication date	Jun 30, 2016
Grant date	—

Abstract

A technique for defragmenting jobs on processor-based computing resources including: (i) determining a first defragmentation condition, which first defragmentation condition will be determined to exist when it is favorable under a first energy consideration to defragment the allocation of jobs as among a set of processor-based computing resources of a supercomputer (for example, a compute-card-based supercomputer); and (ii) on condition that the first defragmentation condition exists, defragmenting the jobs on the set of processor-based computing resources.

First claim

What is claimed is:

1 . A method comprising:
receiving, by job predicting machine logic, a plurality of input calendar entries from calendar(s) of a set of user(s) of a supercomputer, with the plurality of input calendar entries including expected run times for jobs included in a calendar(s) of a set of user(s); computing, by the job predicting machine logic, a first cost approximation associated with running a first job from a plurality of jobs on a supercomputer, with the computation of the first cost approximation being performed under an assumption that no defragmentation is performed prior to running the first job, wherein the computation of the first cost approximation is based, at least in part, upon the plurality of input calendar entries and power costs associated with running the first job without the defragmentation; computing, by the job predicting machine logic, a second cost approximation associated with running the first job on the supercomputer, with the computation of the second cost approximation being performed under an assumption that defragmentation of a first partition of the supercomputer is performed prior to running the first job, wherein the computation of the second cost approximation is based, at least in part, upon the plurality of input calendar entries and power costs associated with the defragmentation and with running the first job subsequent to the defragmentation of the first partition; and on condition that the first cost approximation is greater than the second cost approximation, performing defragmentation of the first partition; wherein the first partition is a partition large enough to run the first job.

2 . The method of claim 1 further comprising:
subsequent to performing the defragmentation of the first partition, performing the first job on the supercomputer.

3 . (canceled)

4 . (canceled)

5 . The method of claim 1 further comprising:
computing an expected time for a need for machine defragmentation, with the first expected time being based, at least in part, on a projected unavailability, due to the machine fragmentation;
computing a delay cost incurred for expected delay of the first job based, at least in part, on the expected time for the need for machine defragmentation; and
computing a maximum expected runtime interval based upon the following: (i) an assumption that the supercomputer is in a defragmented state with free partitions Pn of various sizes because defragmentation is performed prior to the expected time for the need for machine defragmentation, and (ii) a projected submission of the plurality of jobs prior to the expected time for the need for machine defragmentation;
wherein if the maximum expected runtime interval goes beyond the expected time for the need for machine defragmentation, then the maximum expected runtime interval is decreased so that the maximum expected runtime results in an interval that coincides with the expected time for the need for machine defragmentation.

6 . The method of claim 5 wherein:
the computation of the first cost approximation is based, at least in part, on a cost of power required for the defragmentation; and
the computation of the first cost approximation is further based, at least in part, upon a cost savings realized by an assumption that the free partitions of various sizes are powered down during the maximum expected runtime interval.

7 . A computer program product comprising a computer readable storage medium having stored thereon:
first program instructions programmed to receive, by job predicting machine logic, a plurality of input calendar entries from calendar(s) of a set of user(s) of a supercomputer, with the plurality of input calendar entries including expected run times for jobs included in a calendar(s) of a set of user(s); second program instructions programmed to compute, by job predicting machine logic, a first cost approximation associated with running a first job from a plurality of jobs on a supercomputer, with the computation of the first cost approximation being performed under an assumption that no defragmentation is performed prior to running the first job, wherein the computation of the first cost approximation is based, at least in part, upon the plurality of input calendar entries and power costs associated with running the first job without the defragmentation; third program instructions programmed to compute, by the job predicting machine logic, a second cost approximation associated with running the first job on the supercomputer, with the computation of the second cost approximation being performed under an assumption that defragmentation of a first partition of the supercomputer is performed prior to running the first job, wherein the computation of the second cost approximation is based, at least in part, upon the plurality of input calendar entries and power costs associated with the defragmentation and with running the first job subsequent to the defragmentation of the first partition; and fourth program instructions programmed to on condition that the first cost approximation is greater than the second cost approximation, perform defragmentation of the first partition; wherein: the first partition is a partition large enough to run the first job.

8 . The product of claim 7 wherein the medium has further stored thereon:
fifth program instructions programmed to subsequent to performing the defragmentation of the first partition, perform the first job on the supercomputer.

9 . (canceled)

10 . (canceled)

11 . The product of claim 7 wherein the medium has further stored thereon:
fifth program instructions programmed to compute an expected time for a need for machine defragmentation, with the first expected time being based, at least in part, on a projected unavailability, due to the machine fragmentation;
sixth program instructions programmed to compute a delay cost incurred for expected delay of the first job based, at least in part, on the expected time for the need for machine defragmentation; and
seventh program instructions programmed to compute a maximum expected runtime interval based upon the following: (i) an assumption that the supercomputer is in a defragmented state with free partitions Pn of various sizes because defragmentation is performed prior to the expected time for the need for machine defragmentation, and (ii) a projected submission of the plurality of jobs prior to the expected time for the need for machine defragmentation;
wherein if the maximum expected runtime interval goes beyond the expected time for the need for machine defragmentation, then the maximum expected runtime interval is decreased so that the maximum expected runtime results in an interval that coincides with the expected time for the need for machine defragmentation.

12 . The product of claim 11 wherein:
the second program instructions are further programmed to compute the first cost approximation based, at least in part, on a cost of power required for defragmentation; and
the second program instructions are further programmed to compute the first cost approximation further based, at least in part, upon a cost savings realized by an assumption that the free partitions of various sizes are powered down during the maximum expected runtime interval.

13 . A computer system comprising:
a processor(s) set; and a computer readable storage medium; wherein: the processor set is structured, located, connected and/or programmed to run program instructions stored on the computer readable storage medium; and the program instructions include:
first program instructions programmed to receive, b

Assignees

• IBM

Inventors

• Kozloski James R
• Lynar Timothy M
• Nelson Mark D
• Wagner John M

Classifications

• G06F9/5033 Primary

  considering data affinity · CPC title

• G06F9/5016

  the resource being the memory · CPC title

• G06F9/4893

  taking into account power or heat criteria (power management in computers in general G06F1/3203; thermal management in computers in general G06F1/206) · CPC title

• G06F15/825

  Dataflow computers · CPC title

• G06F9/5077 Primary

  Logical partitioning of resources; Management or configuration of virtualized resources (specific details on emulation or internal functioning of virtual machines G06F9/455) · CPC title