Re: [RFC][PATCH 00/16] sched: Core scheduling

[Aubrey Li <aubrey.intel@gmail.com>]

On Sat, Feb 23, 2019 at 3:27 AM Tim Chen <tim.c.chen@linux.intel.com> wrote:
>
> On 2/22/19 6:20 AM, Peter Zijlstra wrote:
> > On Fri, Feb 22, 2019 at 01:17:01PM +0100, Paolo Bonzini wrote:
> >> On 18/02/19 21:40, Peter Zijlstra wrote:
> >>> On Mon, Feb 18, 2019 at 09:49:10AM -0800, Linus Torvalds wrote:
> >>>> On Mon, Feb 18, 2019 at 9:40 AM Peter Zijlstra <peterz@infradead.org> wrote:
> >>>>>
> >>>>> However; whichever way around you turn this cookie; it is expensive and nasty.
> >>>>
> >>>> Do you (or anybody else) have numbers for real loads?
> >>>>
> >>>> Because performance is all that matters. If performance is bad, then
> >>>> it's pointless, since just turning off SMT is the answer.
> >>>
> >>> Not for these patches; they stopped crashing only yesterday and I
> >>> cleaned them up and send them out.
> >>>
> >>> The previous version; which was more horrible; but L1TF complete, was
> >>> between OK-ish and horrible depending on the number of VMEXITs a
> >>> workload had.
> >>>
> >>> If there were close to no VMEXITs, it beat smt=off, if there were lots
> >>> of VMEXITs it was far far worse. Supposedly hosting people try their
> >>> very bestest to have no VMEXITs so it mostly works for them (with the
> >>> obvious exception of single VCPU guests).
> >>
> >> If you are giving access to dedicated cores to guests, you also let them
> >> do PAUSE/HLT/MWAIT without vmexits and the host just thinks it's a CPU
> >> bound workload.
> >>
> >> In any case, IIUC what you are looking for is:
> >>
> >> 1) take a benchmark that *is* helped by SMT, this will be something CPU
> >> bound.
> >>
> >> 2) compare two runs, one without SMT and without core scheduler, and one
> >> with SMT+core scheduler.
> >>
> >> 3) find out whether performance is helped by SMT despite the increased
> >> overhead of the core scheduler
> >>
> >> Do you want some other load in the host, so that the scheduler actually
> >> does do something?  Or is the point just that you show that the
> >> performance isn't affected when the scheduler does not have anything to
> >> do (which should be obvious, but having numbers is always better)?
> >
> > Well, what _I_ want is for all this to just go away :-)
> >
> > Tim did much of testing last time around; and I don't think he did
> > core-pinning of VMs much (although I'm sure he did some of that). I'm
>
> Yes. The last time around I tested basic scenarios like:
> 1. single VM pinned on a core
> 2. 2 VMs pinned on a core
> 3. system oversubscription (no pinning)
>
> In general, CPU bound benchmarks and even things without too much I/O
> causing lots of VMexits perform better with HT than without for Peter's
> last patchset.
>
> > still a complete virt noob; I can barely boot a VM to save my life.
> >
> > (you should be glad to not have heard my cursing at qemu cmdline when
> > trying to reproduce some of Tim's results -- lets just say that I can
> > deal with gpg)
> >
> > I'm sure he tried some oversubscribed scenarios without pinning.
>
> We did try some oversubscribed scenarios like SPECVirt, that tried to
> squeeze tons of VMs on a single system in over subscription mode.
>
> There're two main problems in the last go around:
>
> 1. Workload with high rate of Vmexits (SpecVirt is one)
> were a major source of pain when we tried Peter's previous patchset.
> The switch from vcpus to qemu and back in previous version of Peter's patch
> requires some coordination between the hyperthread siblings via IPI.  And for
> workload that does this a lot, the overhead quickly added up.
>
> For Peter's new patch, this overhead hopefully would be reduced and give
> better performance.
>
> 2. Load balancing is quite tricky.  Peter's last patchset did not have
> load balancing for consolidating compatible running threads.
> I did some non-sophisticated load balancing
> to pair vcpus up.  But the constant vcpu migrations overhead probably ate up
> any improvements from better load pairing.  So I didn't get much
> improvement in the over-subscription case when turning on load balancing
> to consolidate the VCPUs of the same VM. We'll probably have to try
> out this incarnation of Peter's patch and see how well the load balancing
> works.
>
> I'll try to line up some benchmarking folks to do some tests.

I can help to do some basic tests.

Cgroup bias looks weird to me. If I have hundreds of cgroups, should I turn
core scheduling(cpu.tag) on one by one? Or Is there a global knob I missed?

Thanks,
-Aubrey