Mechanism for reducing remote memory accesses to shared data in a multi-nodal computer system

A program product, said program product comprising,

• recordable-type-media;
• and
• said program product disposed on said recordable-type-media, said program product comprising a grouping mechanism that groups threads executable on a multiprocessor computer system containing a plurality of processors including a plurality of subsets of said plurality of processors, each said processor subset containing a respective at least one processor, each said processor subset having a respective memory associated therewith, said processor subsets being disjoint, said multiprocessor computer system supporting concurrent execution of a plurality of jobs, a first job of said plurality of jobs having at least one respective thread, a second job of said plurality of jobs having plurality respective threads;
• wherein said grouping mechanism groups said at least one respective thread of said first job and at least one thread of said plurality respective threads of said second job thread into a first group of a plurality of groups, said at least one respective thread of said first job and said at least one thread of said plurality respective threads of said second job being grouped independently of any association with said first job and said second job, each said first group of a plurality of groups having an affinity with a respective said processor subset and said first group being used to provide said processor subset affinity for said at least one respective thread of said first job and said at least one thread of said plurality respective threads of second job contained therein, wherein said first group of said plurality of groups contains said at least one respective thread of said first job and said at least one thread of said plurality respective threads of said second job, and said first group contains fewer than all said threads of said first job and said second job;
• and
• said program product further comprising an affinity changing mechanism which alters said first groups from said processor subset affinity to a respective altered processor subset, wherein said at least one respective thread of said first job and said at least thread of said plurality respective threads of said second job of said first group are preferentially dispatched to processors of said processor subset affinity before said altering said first group, and said at least one respective thread of said first job and said at least one thread of said plurality respective threads of said second job of said first group are preferentially dispatched to processors of said respective altered processor subset after altering said first group is performed.

A computer-implemented method, said method comprising the steps of:

• creating a first thread, said first thread being capable of executing on any of a plurality of processors of a computer system, said plurality of processors including a plurality of processor subsets of said plurality of processors, each said processor subset containing a respective at least one processor, each said processor subset having a respective memory of said computer system associated therewith, said processor subsets being disjoint, said computer system supporting concurrent execution of a plurality of jobs, said first thread being associated with a first job of said plurality of jobs;
• creating a second thread, said second thread being capable of executing on any of said plurality of processors, said second thread being associated with said first job;
• creating a third thread, said third thread being capable of executing on any of said plurality of processors, said third thread being associated with said first job;
• creating a fourth thread, said fourth thread being capable of executing on any of said plurality of processors, said fourth thread being associated with a second job of said plurality of jobs;
• grouping said first thread, said second thread and said fourth thread, but not said third thread, such that they are contained in a group, said group having an affinity with a first processor subset of said plurality of processor subsets and said group being used to provide said processor subset affinity for said first thread, said second thread and said fourth thread contained therein, and contains fewer than all said threads of said first job and said second job;
• and
• altering of said group from a said first processor subset affinity to a second processor subset, wherein said first thread, said second thread and said fourth thread of said group are preferentially dispatched to said processors of said first processor subset affinity before said step of altering said group is performed, and said first thread, said second thread and said fourth thread of said group are preferentially dispatched to processors of said second processor subset after said step of altering said group is performed.