Techniques for dynamically assigning jobs to processors in a cluster based on inter-thread communications

• US 9,384,042 B2
• Filed: 12/16/2008
• Issued: 07/05/2016
• Est. Priority Date: 12/16/2008
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Arimilli, Lakshminarayana Baba, Arimilli, Ravi Kumar, Basso, Claude, Calvignac, Jean L.
• Primary Examiner(s): ZHE, MENG YAO
• Application Number:
  US12/336,302
• Publication Number:
  US 20100153965A1
• Time in Patent Office: 2,758 Days
• Field of Search: None
• US Class Current: 1/1
• CPC Class Codes: G06F 9/4856 resumption being on a different machine, e.g. task migration, virtual machine migration G06F9/5088 takes precedence
• CPC Class Codes: G06F 9/5066 Algorithms for mapping a plurality of inter-dependent sub-tasks onto a plurality of physical CPUs mappping at compile time, see G06F8/451
• CPC Class Codes: G06F 9/5088 involving task migration

Abstract

A technique for operating a high performance computing (HPC) cluster includes monitoring communication between threads assigned to multiple processors included in the HPC cluster. The HPC cluster includes multiple nodes that each include two or more of the multiple processors. One or more of the threads are moved to a different one of the multiple processors based on the communication between the threads.

Claims

1. A method of operating a high performance computing cluster that includes multiple nodes that each include multiple processors and monitoring hardware, the method comprising:

monitoring, by the monitoring hardware, communication between a plurality of threads assigned to the multiple processors;

periodically broadcasting, by the monitoring hardware, information, related to a level of processor utilization and network utilization at each of the multiple nodes, from each of the multiple nodes to remaining ones of the multiple nodes;

updating, by the monitoring hardware, respective local job tables maintained in each of the multiple nodes based on the broadcast information; and

reassigning, by the monitoring hardware, based on the broadcast information in the respective local job tables, at least one of the one or more threads to a different one of the multiple processors such that threads of a job complete at substantially the same time, wherein the reassigning includes;

cracking one of the threads that is executing on a first processor, included in the multiple processors, into at least two secondary threads based on a first workload of the first processor;

moving one of the at least two secondary threads to a second processor, included in the multiple processors, based on a second workload of the second processor; and

moving identified threads of the plurality of threads that communicate above a threshold level to identified processors of the multiple processors that are located physically closer to each other than processors of the multiple processors that previously executed the identified threads.

2. The method of claim 1, further comprising:

moving one or more of the threads to a different one of the multiple processors based on workloads of the multiple processors.

3. The method of claim 2, wherein the workloads of the multiple processors are based on utilization of one or more floating-point units included within each of the multiple processors.

4. The method of claim 1, wherein the multiple nodes are arranged in a three-dimensional Torus topology.

5. The method of claim 1, wherein the information is broadcast using a message passing interface.

6. The method of claim 1, wherein the multiple processors are included within multiple chip-level multiprocessors.

7. The method of claim 1, wherein at least some of the multiple processors included within different ones of the multiple chip-level multiprocessors are coupled together via host channel adapters and switch fabrics including one or more switches.

8. A high performance computing cluster, comprising:

multiple nodes that each include multiple processors; and

monitoring hardware included in each of the multiple nodes, wherein the monitoring hardware is configured to monitor communication between a plurality of threads assigned to the multiple processors, periodically broadcast information, related to a level of processor utilization and network utilization at each of the multiple nodes, from each of the multiple nodes to remaining ones of the multiple nodes, update respective local job tables maintained in each of the multiple nodes based on the broadcast information, and move, based on the broadcast information in the respective local job tables, at least one of the one or more threads to a different one of the multiple processors such that threads of a job complete at substantially the same time, and wherein the move includes cracking one of the threads that is executing on a first processor, included in the multiple processors, into at least two secondary threads based on a first workload of the first processor, moving one of the at least two secondary threads to a second processor, included in the multiple processors, based on a second workload of the second processor, and moving identified threads of the plurality of threads that communicate above a threshold level to identified processors of the multiple processors that are located physically closer to each other than processors of the multiple processors that previously executed the identified threads.

9. The high performance computing cluster of claim 8, wherein the high performance computing cluster is further configured to move one or more of the threads to a different one of the multiple processors based on workloads of the multiple processors.

10. The high performance computing cluster of claim 9, wherein the workloads of the multiple processors are based on utilization of one or more floating-point units included within each of the multiple processors.

11. The high performance computing cluster of claim 8, wherein the multiple nodes are arranged in a three-dimensional Torus topology.

12. The high performance computing cluster of claim 8, where the information is broadcast using a message passing interface.

13. The high performance computing cluster of claim 8, wherein the multiple processors are included within multiple chip-level multiprocessors and at least some of the multiple processors included within different ones of the multiple chip-level multiprocessors are coupled together via host channel adapters and switch fabrics including one or more switches.

14. A method of operating a high performance computing cluster that includes multiple nodes that each include multiple processors and monitoring hardware, the method comprising:

monitoring, by the monitoring hardware, communication between a plurality of threads assigned to the multiple processors

periodically broadcasting, by the monitoring hardware, information, related to a level of processor utilization and network utilization at each of the multiple nodes, from each of the multiple nodes to remaining ones of the multiple nodes;

updating, by the monitoring hardware, respective local job tables maintained in each of the multiple nodes based on the broadcast information; and

moving, by the monitoring hardware, based on the broadcast information in the respective local job tables, at least one of the one or more threads to a different one of the multiple processors such that threads of a job complete at substantially the same time, wherein the information is broadcast using a message passing interface and the multiple nodes are arranged in a three-dimensional Torus topology, and wherein the moving includes;

cracking one of the threads that is executing on a first processor, included in the multiple processors, into at least two secondary threads based on a first workload of the first processor;

moving one of the at least two secondary threads to a second processor, included in the multiple processors, based on a second workload of the second processor; and

moving identified threads of the plurality of threads that communicate above a threshold level to identified processors of the multiple processors that are located physically closer to each other than processors of the multiple processors that previously executed the identified threads.