Proportional resizing of a logical partition based on a degree of performance difference between threads for high-performance computing on non-dedicated clusters

Patent metadata

Field	Value
Publication number	US-9600344-B2
Application number	US-35660609-A
Country	US
Kind code	B2
Filing date	Jan 21, 2009
Priority date	Jan 21, 2009
Publication date	Mar 21, 2017
Grant date	Mar 21, 2017

Abstract

Techniques for enabling high-performance computing are provided. The techniques include resizing a logical partition in a non-dedicated compute cluster server to enable high-performance computing, wherein a high performance computing application is executed such that the high performance computing application is configured to complete execution of each of one or more application threads at a similar time as a slowest thread in the cluster, and wherein the non-dedicated compute cluster comprises one or more servers and the logical partition is created by partitioning one or more server resources.

First claim

What is claimed is:

1. A method for enabling high-performance computing, the method comprising:
resizing one or more logical partitions in a non-dedicated compute cluster to enable high-performance computing, wherein a high-performance computing application is executed such that the high-performance computing application is configured to execute each of multiple high-performance computing application threads on a corresponding individual logical partition on a separate corresponding server, and to reduce differences in execution completion time for each of the multiple high-performance computing application threads, and wherein the non-dedicated compute cluster comprises multiple servers and a logical partition is created by partitioning one or more server resources;
wherein resizing the one or more logical partitions comprises:
determining a difference in performance of each of said high-performance computing application threads of the high-performance computing application in the non-dedicated compute cluster with respect to the slowest thread of the high-performance computing application; and
resizing a current allocation of capacity of each logical partition in which a corresponding high-performance computing application thread has a difference in performance with respect to the slowest thread of the high-performance computing application via a percentage reduction of the current allocation of capacity, wherein said percentage reduction of the current allocation of capacity is:
computed for the logical partition with the highest product of (i) time taken to complete a phase and (ii) capacity in proportion to the associated difference in performance in terms of a percentage computed via [max i (T i *C i )−(T i *C i )]/(T i *C i )*100, wherein, T i is the time taken to complete a phase on a given logical partition with capacity C i ; and
set for the remaining logical partitions, excluding the logical partition associated with the slowest thread of the high-performance computing application, in the same proportion as computed for the logical partition with the highest product of (i) time taken to complete a phase and (ii) capacity.

2. The method of claim 1 , wherein logical partition creation comprises creating one or more separate logical partitions for each one of one or more high-performance computing cluster workloads and for a non-high-performance computing cluster workload.

3. The method of claim 1 , comprising determining unutilized compute resources in a logical partition in a high-performance computing cluster via monitoring performance imbalance among each of one or more high-performance computing application threads.

4. The method of claim 3 , wherein monitoring performance imbalance among each of the one or more high-performance computing application threads comprises monitoring excess capacity at each of one or more nodes.

5. The method of claim 4 , wherein monitoring excess capacity comprises:
measuring one or more cycles used by each high-performance computing application thread on each node in a given time window; and
identifying the one or more cycles that are in excess at each node with respect to a slowest high-performance computing application thread.

6. The method of claim 1 , wherein resizing the one or more logical partitions in the non-dedicated compute cluster comprises allocating unutilized compute resources to one or more non-high-performance computing cluster workloads.

7. The method of claim 1 , further comprising ensuring that one or more logical partitions across one or more different servers that run one or more threads belonging to a same high-performance computing workload are scheduled concurrently.

8. The method of claim 1 , wherein partitioning one or more server resources comprises using historical resource usage data of a non-high-performance computing cluster workload to predict one or more resource requirements of the non-high-performance computing cluster workload such that one or more service level agreement (SLA) requirements of the non-high-performance computing cluster workload are guaranteed.

9. The method of claim 1 , further comprising implementing a gang-scheduling scheme at a hypervisor layer such that one or more logical partitions across different servers that run one or more threads belonging to a same high-performance computing workload are scheduled concurrently.

10. The method of claim 1 , wherein resizing the one or more logical partitions comprises resizing compute resources at each of one or more physical nodes.

11. The method of claim 1 , wherein resizing the one or more logical partitions comprises dynamically resizing the one or more logical partitions in proportion of an excess capacity.

12. The method of claim 1 , further comprising creating a separate high-performance computing partition for each high-performance computing application.

13. The method of claim 1 , wherein resizing the one or more logical partitions comprises taking into account one or more resource requirements of a non-high-performance computing workload such that one or more service level agreement (SLA) requirements of a non-high-performance computing workload are guaranteed.

14. A computer program product comprising a tangible non-transitory computer readable storage medium having computer readable program code for enabling high-performance computing, said computer program product including:
computer readable program code for resizing one or more logical partitions in a non-dedicated compute cluster to enable high-performance computing, wherein a high-performance computing application is executed such that the high-performance computing application is configured to execute each of multiple high-performance computing application threads on a corresponding individual logical partition on a separate corresponding server, and to reduce differences in execution completion time for each of the multiple high-performance computing application threads, and wherein the non-dedicated compute cluster comprises multiple servers and a logical partition is created by partitioning one or more server resources;
wherein resizing the one or more logical partitions comprises:
determining a difference in performance of each of said high-performance computing application threads of the high-performance computing application in the non-dedicated compute cluster with respect to the slowest thread of the high-performance computing application; and
resizing a current allocation of capacity of each logical partition in which a corresponding high-performance computing application thread has a difference in performance with respect to the slowest thread of the high-performance computing application via a percentage reduction of the current allocation of capacity, wherein said percentage reduction of the current allocation of capacity is:
computed for the logical partition with the highest product of (i) time taken to complete a phase and (ii) capacity in proportion to the associated difference in performance in terms of a percentage computed via [max i (T i *C i )−(T i *C i )]/(T i *C i )*100, wherein, T i is the time taken to complete a phase on a given logical partition with capacity C i ; and
set for the remaining logical partitions, excluding the logical partition associated with the slowest thread of the high-performance computing application, in the same proportion as computed for the logical partition with the highest product of (i) time taken to complete a phase and (ii) capacity.

15. The computer program product of claim 14 , comprising computer readable program code for creating one or more separ

Assignees

• De Pradipta
• Kothari Ravi
• Mann Vijay
• Sachdev Rajiv
• IBM

Inventors

• De Pradipta
• Kothari Ravi
• Mann Vijay
• Sachdev Rajiv

Classifications

• 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

• G06F9/52

  Program synchronisation; Mutual exclusion, e.g. by means of semaphores · CPC title