sched-domains
Folks, I remain all in a dither over the sched-domains NUMA/SMT work which has been in the -mm kernels for a while.
I have no concerns about its stability, but I do want to see general agreement that this is the way to proceed, although of course it may need enhancements/tuning later on.
In terms of performace testing results I believe everything has been positive except for some testing which Andi did on the STREAM benchmark.
My question is: to what extent have the ia64 NUMA architectural people reviewed this code, and what is the risk that someone will come along in six months time saying we need to rip it all out?
Thanks.
From: "Andrew Morton" <[email protected]>
...
Quoted 3 lines
My question is: to what extent have the ia64 NUMA architectural people reviewed this code, and what is the risk that someone will come along in six months time saying we need to rip it all out?
I am generally happy with the architecture of sched-domains. The big advantage I see for large Altix-like NUMA platforms is that much of the scheduler policy decisions are parameterized in sched_domain fields and are more easily tuned on a per-architecture and per-platform basis. In particular, it is easy to define additional scheduler domains that are larger than a single node and smaller than all the CPUs in the system, each domain having its own parameterized characteristics of load-balance frequency, etc.
John Hawkes