Messages in this thread |  | | Subject | Re: [RESEND RFC PATCH V3] sched: Improve scalability of select_idle_sibling using SMT balance | From | Steven Sistare <> | Date | Fri, 2 Feb 2018 16:06:32 -0500 |
| |
On 2/2/2018 2:59 PM, Peter Zijlstra wrote: > On Fri, Feb 02, 2018 at 11:53:40AM -0500, Steven Sistare wrote: >> On 2/1/2018 7:33 AM, Peter Zijlstra wrote: >>> On Mon, Jan 29, 2018 at 03:31:02PM -0800, subhra mazumdar wrote: >>>> + rcu_read_lock(); >>>> + sd = rcu_dereference(per_cpu(sd_llc, this_cpu)); >>>> + if (util) { >>>> + for_each_lower_domain(sd) { >>>> + if (sd->level == 0) >>>> + break; >>> >>> afaict you really only need this for the core, and here you're assuming >>> everything below the LLC is cores. Would it not be much clearer if you >>> introduce sd_core. >>> >>> As is, for_each_lower_domain includes the starting domain, sd->group >>> then is the first core group for this cpu. But then you continue to the >>> smt domain (on Intel, on other architectures there could be a cluster >>> domain in between) and then you bail using that sd->level == 0 hack >>> because otherwise things would go *bang*. >> >> Hi Peter, >> >> The code here and in smt_balance intentionally visits each level between >> the llc and smt, including core-cluster on architectures that define it. >> smt_balance thus has the chance to randomly pick a better cluster, >> and then within that cluster randomly pick a better core. It makes sense, >> as resources are shared within a cluster, and choosing a less loaded cluster >> should give better performance. As you suggest in a few other places, >> it would be nice to see performance results for this case. We have >> SPARC processors with core clusters. >> > > But then you get that atomic crud to contend on the cluster level, which > is even worse than it contending on the core level.
True, but it can still be a net win if we make better scheduling decisions. A saving grace is that the atomic counter is only updated if the cpu makes a transition from idle to busy or vice versa.
We need data for this type of system, showing improvements for normal workloads, and showing little downside for a high context switch rate torture test.
- Steve
|  |