Re: [PATCH 0/5] Speed up mremap on large regions

[Lokesh Gidra]

On Thu, Oct 1, 2020 at 9:00 AM Kalesh Singh <kaleshsingh@xxxxxxxxxx> wrote:
>
> On Thu, Oct 1, 2020 at 8:27 AM Kirill A. Shutemov
> <kirill.shutemov@xxxxxxxxxxxxxxx> wrote:
> >
> > On Wed, Sep 30, 2020 at 03:42:17PM -0700, Lokesh Gidra wrote:
> > > On Wed, Sep 30, 2020 at 3:32 PM Kirill A. Shutemov
> > > <kirill.shutemov@xxxxxxxxxxxxxxx> wrote:
> > > >
> > > > On Wed, Sep 30, 2020 at 10:21:17PM +0000, Kalesh Singh wrote:
> > > > > mremap time can be optimized by moving entries at the PMD/PUD level if
> > > > > the source and destination addresses are PMD/PUD-aligned and
> > > > > PMD/PUD-sized. Enable moving at the PMD and PUD levels on arm64 and
> > > > > x86. Other architectures where this type of move is supported and known to
> > > > > be safe can also opt-in to these optimizations by enabling HAVE_MOVE_PMD
> > > > > and HAVE_MOVE_PUD.
> > > > >
> > > > > Observed Performance Improvements for remapping a PUD-aligned 1GB-sized
> > > > > region on x86 and arm64:
> > > > >
> > > > >     - HAVE_MOVE_PMD is already enabled on x86 : N/A
> > > > >     - Enabling HAVE_MOVE_PUD on x86   : ~13x speed up
> > > > >
> > > > >     - Enabling HAVE_MOVE_PMD on arm64 : ~ 8x speed up
> > > > >     - Enabling HAVE_MOVE_PUD on arm64 : ~19x speed up
> > > > >
> > > > >           Altogether, HAVE_MOVE_PMD and HAVE_MOVE_PUD
> > > > >           give a total of ~150x speed up on arm64.
> > > >
> > > > Is there a *real* workload that benefit from HAVE_MOVE_PUD?
> > > >
> > > We have a Java garbage collector under development which requires
> > > moving physical pages of multi-gigabyte heap using mremap. During this
> > > move, the application threads have to be paused for correctness. It is
> > > critical to keep this pause as short as possible to avoid jitters
> > > during user interaction. This is where HAVE_MOVE_PUD will greatly
> > > help.
> >
> > Any chance to quantify the effect of mremap() with and without
> > HAVE_MOVE_PUD?
> >
> > I doubt it's a major contributor to the GC pause. I expect you need to
> > move tens of gigs to get sizable effect. And if your GC routinely moves
> > tens of gigs, maybe problem somewhere else?
> >
> > I'm asking for numbers, because increase in complexity comes with cost.
> > If it doesn't provide an substantial benefit to a real workload
> > maintaining the code forever doesn't make sense.
>
mremap is indeed the biggest contributor to the GC pause. It has to
take place in what is typically known as a 'stop-the-world' pause,
wherein all application threads are paused. During this pause the GC
thread flips the GC roots (threads' stacks, globals etc.), and then
resumes threads along with concurrent compaction of the heap.This
GC-root flip differs depending on which compaction algorithm is being
used.

In our case it involves updating object references in threads' stacks
and remapping java heap to a different location. The threads' stacks
can be handled in parallel with the mremap. Therefore, the dominant
factor is indeed the cost of mremap. From patches 2 and 4, it is clear
that remapping 1GB without this optimization will take ~9ms on arm64.

Although this mremap has to happen only once every GC cycle, and the
typical size is also not going to be more than a GB or 2, pausing
application threads for ~9ms is guaranteed to cause jitters. OTOH,
with this optimization, mremap is reduced to ~60us, which is a totally
acceptable pause time.

Unfortunately, implementation of the new GC algorithm hasn't yet
reached the point where I can quantify the effect of this
optimization. But I can confirm that without this optimization the new
GC will not be approved.


> Lokesh on this thread would be better able to answer this. I'll let
> him weigh in here.
> Thanks, Kalesh
> >
> > --
> >  Kirill A. Shutemov
> >
> > --
> > To unsubscribe from this group and stop receiving emails from it, send an email to kernel-team+unsubscribe@xxxxxxxxxxx.
> >