CPU binding: What are some appropriate uses?

Q&A
, , ,
1

Some colleagues were debating the merits of scheduler binding options, and we realized we lacked clarity as to when binding tasks to resources might be most beneficial. Below is a specific topic:

On most clusters, one can request a specific processor core binding (processor affinity). For example, with SLURM, one might include the --cpu-bind option with the srun command, to control task binding. On SGE-type of clusters, the option is -binding <binding_strategy>, where binding strategy could be linear, striding, explicit. (Based on the documentation, this latter option allows one to request a specific core binding.)

When would implementing cpu binding benefit a job? Are there specific software packages or examples of scripts for which the implementation of this option might improve the performance of a job? If so, can anyone offer descriptions of these scenarios? Any explanations regarding the mechanisms behind the increased performance would also be welcome!

Thank you!

Katia and Torey
BU/Mines

2

The only time I’ve seen binding matter was on hardware configurations that tightly coupled memory and CPU (e.g, an large SMP system). That particular system used a slower than memory speed bus to interconnect the “blades” that made up the system, so if you created a task that needed more memory than was on a single blade, you would see.major speed improvement from splitting it into threads, each bound to a cpu and only referencing on-blade memory. Without binding, cross-blade memory references would drag your process to a snail pace. On that same system, doing I/O from an on-blade controller or network card made a huge difference as well.

So, in short, my experience is you want binding when the hardware design makes it necessary for your application.

Cheers,
Ric