Messages in this thread | | | Subject | Re: [PATCH] cpufreq: intel_pstate: Enable HWP IO boost for all servers | From | srinivas pandruvada <> | Date | Mon, 20 Mar 2023 11:33:43 -0700 |
| |
On Mon, 2023-03-20 at 18:03 +0100, Giovanni Gherdovich wrote: > On Thu, 2023-03-02 at 20:14 -0800, Srinivas Pandruvada wrote: >
...
Hi Giovanni,
> Hello Srinivas, > > Good catch. We've had HWP IO Boost in the kernel for a while now but > we > weren't enabling on most of the modern CPUs... This fixes it. > > One thing though: I've the impression that HWP IO Boost depends on > having > per-core p-states -- otherwise you'll be boosting up and down the > entire machine > instead of just the one core waking up from IO. > Enabling the feature on all machines with the ACPI PM server profile > would > force it also where per-core p-states aren't available.
This feature only exists on HWP systems. There are no HWP systems without per core P-states. Here we are enabling for performance and enterprise servers only.
> > Would you agree with this assessment? > Do I correctly understand the reason why per-core p-states are needed > here? This problem with IO will be more pronounced in per-core P-states systems as it will not be influenced by other cores. But even if non per-core systems if other cores are idle or lightly loaded, the same problem can occur. But I don't have data on such systems.
> > I remember you mentioned the the dependency on per-core p-states in > the cover > letter from the HWP IO Boost submission in 2018 > https://lore.kernel.org/lkml/20180605214242.62156-1-srinivas.pandruvada@linux.intel.com/ > > I think there's a tradeoff here. Up until this patch, we forgot to > enable the > feature on four generations of Intel CPUs. That's a lot of lost > performance; > thanks to this patch it won't happen ever again, because nobody will > have to > update the model list at every new CPU release. > CPU model updates always gets missed and also misses testing opportunity.
> On the other side, there may be some penalty for machines that: > - have HWP > - don't have per-core p-states > I don't know how large that interesection is, or how big the penalty. >
Thanks, Srinivas
> > Giovanni >
| |