| Message ID | 1395246165-31150-7-git-send-email-vincent.guittot@linaro.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Vincent, On 03/19/2014 09:52 PM, Vincent Guittot wrote: > Set the power domain dependency at SMT level of Power8 but keep the flag > clear at CPU level. The goal is to consolidate tasks on the threads of a > core up to a level as explained by Preeti: > "On powerpc we would want to clear the SD_SHARE_POWERDOMAIN flag at the CPU > domain. On Power8, considering we have 8 threads per core, we would want to > consolidate tasks atleast upto 4 threads without significant performance > impact before spilling over to the other cores. By doing so, besides making > use of the higher power of the core we could do cpuidle management at the > core level for the remaining idle cores as a result of this consolidation." > > Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> > Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> We were discussing the impact of this consolidation and we are not too sure if it will yield us good power efficiency. So we would want to experiment with the power aware scheduler to find the "sweet spot" for the number of threads to consolidate to and more importantly if there is one such number at all. Else we would not want to go this way at all. Hence it looks best if this patch is dropped until we validate it. We don't want the code getting in and then out if we find out later there are no benefits to it. I am sorry that I suggested this patch a bit pre-mature in the experimentation and validation stage. When you release the load balancing patchset for power aware scheduler I shall validate this patch. But until then its best if it does not get merged. Thanks Regards Preeti U Murthy > --- > arch/powerpc/kernel/smp.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/arch/powerpc/kernel/smp.c b/arch/powerpc/kernel/smp.c > index c9cade5..fbbac3c 100644 > --- a/arch/powerpc/kernel/smp.c > +++ b/arch/powerpc/kernel/smp.c > @@ -759,7 +759,7 @@ int setup_profiling_timer(unsigned int multiplier) > /* cpumask of CPUs with asymetric SMT dependancy */ > static const int powerpc_smt_flags(void) > { > - int flags = SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES; > + int flags = SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES | SD_SHARE_POWERDOMAIN; > > if (cpu_has_feature(CPU_FTR_ASYM_SMT)) { > printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n"); >
On Sun, 2014-03-23 at 07:19 +0530, Preeti U Murthy wrote: > We were discussing the impact of this consolidation and we are not too > sure if it will yield us good power efficiency. So we would want to > experiment with the power aware scheduler to find the "sweet spot" for > the number of threads to consolidate to and more importantly if there > is > one such number at all. Else we would not want to go this way at all. > Hence it looks best if this patch is dropped until we validate it. We > don't want the code getting in and then out if we find out later there > are no benefits to it. > > I am sorry that I suggested this patch a bit pre-mature in the > experimentation and validation stage. When you release the load > balancing patchset for power aware scheduler I shall validate this > patch. But until then its best if it does not get merged. It's quite possible that we never find a correct "sweet spot" for all workloads. Ideally, the "target" number of used threads per core should be a tunable so that the user / distro can "tune" based on a given workload whether to pack cores and how much to pack them, vs. spreading the workload. Akin to scheduling for performance vs. power in a way (though lower perf usually means higher power due to longer running jobs of course). In any case, we need to experiment. Cheers, Ben.
On 23 March 2014 02:49, Preeti U Murthy <preeti@linux.vnet.ibm.com> wrote: > Hi Vincent, > > On 03/19/2014 09:52 PM, Vincent Guittot wrote: >> Set the power domain dependency at SMT level of Power8 but keep the flag >> clear at CPU level. The goal is to consolidate tasks on the threads of a >> core up to a level as explained by Preeti: >> "On powerpc we would want to clear the SD_SHARE_POWERDOMAIN flag at the CPU >> domain. On Power8, considering we have 8 threads per core, we would want to >> consolidate tasks atleast upto 4 threads without significant performance >> impact before spilling over to the other cores. By doing so, besides making >> use of the higher power of the core we could do cpuidle management at the >> core level for the remaining idle cores as a result of this consolidation." >> >> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> >> Reviewed-by: Preeti U Murthy <preeti@linux.vnet.ibm.com> > > We were discussing the impact of this consolidation and we are not too > sure if it will yield us good power efficiency. So we would want to > experiment with the power aware scheduler to find the "sweet spot" for > the number of threads to consolidate to and more importantly if there is > one such number at all. Else we would not want to go this way at all. > Hence it looks best if this patch is dropped until we validate it. We > don't want the code getting in and then out if we find out later there > are no benefits to it. > > I am sorry that I suggested this patch a bit pre-mature in the > experimentation and validation stage. When you release the load > balancing patchset for power aware scheduler I shall validate this > patch. But until then its best if it does not get merged. Hi Preeti, ok. I will drop this patch. Nevertheless, this patch only adds details about the power domain topology of Powerpc arch but it doesn't make any assumption about the scheduler policy or how it can be used by a power aware scheduler. Regards, Vincent > > Thanks > > Regards > Preeti U Murthy >> --- >> arch/powerpc/kernel/smp.c | 2 +- >> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> diff --git a/arch/powerpc/kernel/smp.c b/arch/powerpc/kernel/smp.c >> index c9cade5..fbbac3c 100644 >> --- a/arch/powerpc/kernel/smp.c >> +++ b/arch/powerpc/kernel/smp.c >> @@ -759,7 +759,7 @@ int setup_profiling_timer(unsigned int multiplier) >> /* cpumask of CPUs with asymetric SMT dependancy */ >> static const int powerpc_smt_flags(void) >> { >> - int flags = SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES; >> + int flags = SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES | SD_SHARE_POWERDOMAIN; >> >> if (cpu_has_feature(CPU_FTR_ASYM_SMT)) { >> printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n"); >> >
diff --git a/arch/powerpc/kernel/smp.c b/arch/powerpc/kernel/smp.c index c9cade5..fbbac3c 100644 --- a/arch/powerpc/kernel/smp.c +++ b/arch/powerpc/kernel/smp.c @@ -759,7 +759,7 @@ int setup_profiling_timer(unsigned int multiplier) /* cpumask of CPUs with asymetric SMT dependancy */ static const int powerpc_smt_flags(void) { - int flags = SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES; + int flags = SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES | SD_SHARE_POWERDOMAIN; if (cpu_has_feature(CPU_FTR_ASYM_SMT)) { printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");