Date:	Thu, 11 Apr 2013 15:55:46 +0900
From:	Minchan Kim <minchan@...nel.org>
To:	KOSAKI Motohiro <kosaki.motohiro@...il.com>
Cc:	linux-mm@...ck.org, linux-kernel@...r.kernel.org,
	Michael Kerrisk <mtk.manpages@...il.com>,
	Arun Sharma <asharma@...com>,
	John Stultz <john.stultz@...aro.org>,
	Mel Gorman <mel@....ul.ie>, Hugh Dickins <hughd@...gle.com>,
	Dave Hansen <dave@...ux.vnet.ibm.com>,
	Rik van Riel <riel@...hat.com>, Neil Brown <neilb@...e.de>,
	Mike Hommey <mh@...ndium.org>, Taras Glek <tglek@...illa.com>,
	KOSAKI Motohiro <kosaki.motohiro@...fujitsu.com>,
	KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>,
	Jason Evans <je@...com>, sanjay@...gle.com,
	Paul Turner <pjt@...gle.com>,
	Johannes Weiner <hannes@...xchg.org>,
	Michel Lespinasse <walken@...gle.com>,
	Andrew Morton <akpm@...ux-foundation.org>
Subject: Re: [RFC v7 00/11] Support vrange for anonymous page

On Wed, Apr 10, 2013 at 04:22:58PM -0400, KOSAKI Motohiro wrote:
> (3/12/13 3:38 AM), Minchan Kim wrote:
> > First of all, let's define the term.
> > From now on, I'd like to call it as vrange(a.k.a volatile range)
> > for anonymous page. If you have a better name in mind, please suggest.
> > 
> > This version is still *RFC* because it's just quick prototype so
> > it doesn't support THP/HugeTLB/KSM and even couldn't build on !x86.
> > Before further sorting out issues, I'd like to post current direction
> > and discuss it. Of course, I'd like to extend this discussion in
> > comming LSF/MM.
> > 
> > In this version, I changed lots of thing, expecially removed vma-based
> > approach because it needs write-side lock for mmap_sem, which will drop
> > performance in mutli-threaded big SMP system, KOSAKI pointed out.
> > And vma-based approach is hard to meet requirement of new system call by
> > John Stultz's suggested semantic for consistent purged handling.
> > (http://linux-kernel.2935.n7.nabble.com/RFC-v5-0-8-Support-volatile-for-anonymous-range-tt575773.html#none)
> > 
> > I tested this patchset with modified jemalloc allocator which was
> > leaded by Jason Evans(jemalloc author) who was interest in this feature
> > and was happy to port his allocator to use new system call.
> > Super Thanks Jason!
> > 
> > The benchmark for test is ebizzy. It have been used for testing the
> > allocator performance so it's good for me. Again, thanks for recommending
> > the benchmark, Jason.
> > (http://people.freebsd.org/~kris/scaling/ebizzy.html)
> > 
> > The result is good on my machine (12 CPU, 1.2GHz, DRAM 2G)
> > 
> > 	ebizzy -S 20
> > 
> > jemalloc-vanilla: 52389 records/sec
> > jemalloc-vrange: 203414 records/sec
> > 
> > 	ebizzy -S 20 with background memory pressure
> > 
> > jemalloc-vanilla: 40746 records/sec
> > jemalloc-vrange: 174910 records/sec
> > 
> > And it's much improved on KVM virtual machine.
> > 
> > This patchset is based on v3.9-rc2
> > 
> > - What's the sys_vrange(addr, length, mode, behavior)?
> > 
> >   It's a hint that user deliver to kernel so kernel can *discard*
> >   pages in a range anytime. mode is one of VRANGE_VOLATILE and
> >   VRANGE_NOVOLATILE. VRANGE_NOVOLATILE is memory pin operation so
> >   kernel coudn't discard any pages any more while VRANGE_VOLATILE
> >   is memory unpin opeartion so kernel can discard pages in vrange
> >   anytime. At a moment, behavior is one of VRANGE_FULL and VRANGE
> >   PARTIAL. VRANGE_FULL tell kernel that once kernel decide to
> >   discard page in a vrange, please, discard all of pages in a
> >   vrange selected by victim vrange. VRANGE_PARTIAL tell kernel
> >   that please discard of some pages in a vrange. But now I didn't
> >   implemented VRANGE_PARTIAL handling yet.
> > 
> > - What happens if user access page(ie, virtual address) discarded
> >   by kernel?
> > 
> >   The user can encounter SIGBUS.
> > 
> > - What should user do for avoding SIGBUS?
> >   He should call vrange(addr, length, VRANGE_NOVOLATILE, mode) before
> >   accessing the range which was called
> >   vrange(addr, length, VRANGE_VOLATILE, mode)
> > 
> > - What happens if user access page(ie, virtual address) doesn't
> >   discarded by kernel?
> > 
> >   The user can see vaild data which was there before calling
> > vrange(., VRANGE_VOLATILE) without page fault.
> > 
> > - What's different with madvise(DONTNEED)?
> > 
> >   System call semantic
> > 
> >   DONTNEED makes sure user always can see zero-fill pages after
> >   he calls madvise while vrange can see data or encounter SIGBUS.
> 
> For replacing DONTNEED, user want to zero-fill pages like DONTNEED
> instead of SIGBUS. So, new flag option would be nice.

If userspace people want it, I can do it. 
But not sure they want it at the moment becaue vrange is rather
different concept of madvise(DONTNEED) POV usage.

As you know well, in case of DONTNEED, user calls madvise _once_ and
VM releases memory as soon as he called system call.
But vrange is same with delayed free when the system memory pressure
happens so user can't know OS frees the pages anytime.
It means user should call pair of system call both VRANGE_VOLATILE
and VRANGE_NOVOLATILE for right usage of volatile range
(for simple, I don't want to tell SIGBUS fault recovery method).
If he took a mistake(ie, NOT to call VRANGE_NOVOLATILE) on the range
which is used by current process, pages used by some process could be
disappeared suddenly.

In summary, I don't think vrange is a replacement of madvise(DONTNEED)
but could be useful with madvise(DONTNEED) friend. For example, we can
make return 1 in vrange(VRANGE_VOLATILE) if memory pressure was already
severe so user can catch up memory pressure by return value and calls
madvise(DONTNEED) if memory pressure was already severe. Of course, we
can handle it vrange system call itself(ex, change vrange system call to
madvise(DONTNEED) but don't want it because I want to keep vrange hinting
sytem call very light at all times so user can expect latency.

> 
> I played a bit this patch. The result looks really promissing.
> (i.e. 20x faster)

Thanks for the testing with Glibc!
Yes. Although I didn't post my KVM test result with jemalloc, it could be
very good, too.

> 
> My machine have 24cpus, 8GB ram, kvm guest. I guess current DONTNEED
> implementation doesn't fit kvm at all.

Yes. I expect virtual machine MMU/cache/TLB handling would be expensive
than bare box.

> 
> 
> # of     # of   # of
> thread   iter   iter (patched glibc)

What's the workload?

> ----------------------------------
>   1      438    10740
>   2      842    20916
>   4      987    32534
>   8      717    15155
>  12      714    14109
>  16      708    13457
>  20      720    13742
>  24      727    13642
>  28      715    13328
>  32      709    13096
>  36      705    13661
>  40      708    13634
>  44      707    13367
>  48      714    13377
> 
> 
> ---------libc patch (just dirty hack) ----------------------
> 
> diff --git a/malloc/arena.c b/malloc/arena.c
> index 12a48ad..da04f67 100644
> --- a/malloc/arena.c
> +++ b/malloc/arena.c
> @@ -365,6 +365,8 @@ extern struct dl_open_hook *_dl_open_hook;
>  libc_hidden_proto (_dl_open_hook);
>  #endif
> 
> +int vrange_enabled = 0;
> +
>  static void
>  ptmalloc_init (void)
>  {
> @@ -457,6 +459,18 @@ ptmalloc_init (void)
>      if (check_action != 0)
>        __malloc_check_init();
>    }
> +
> +  {
> +    char *vrange = getenv("MALLOC_VRANGE");
> +    if (vrange) {
> +      int val = atoi(vrange);
> +      if (val) {
> +       printf("glibc: vrange enabled\n");
> +       vrange_enabled = !!val;
> +      }
> +    }
> +  }
> +
>    void (*hook) (void) = force_reg (__malloc_initialize_hook);
>    if (hook != NULL)
>      (*hook)();
> @@ -628,9 +642,14 @@ shrink_heap(heap_info *h, long diff)
>         return -2;
>        h->mprotect_size = new_size;
>      }
> -  else
> -    __madvise ((char *)h + new_size, diff, MADV_DONTNEED);
> +  else {
> +    if (vrange_enabled) {
> +      syscall(314, (char *)h + new_size, diff, 0, 1);
> +    } else {
> +      __madvise ((char *)h + new_size, diff, MADV_DONTNEED);
> +    }
>    /*fprintf(stderr, "shrink %p %08lx\n", h, new_size);*/
> +  }
> 
>    h->size = new_size;
>    return 0;
> diff --git a/malloc/malloc.c b/malloc/malloc.c
> index 70b9329..3782244 100644
> --- a/malloc/malloc.c
> +++ b/malloc/malloc.c
> @@ -4403,6 +4403,7 @@ _int_pvalloc(mstate av, size_t bytes)
>  /*
>    ------------------------------ malloc_trim ------------------------------
>  */
> +extern int vrange_enabled;
> 
>  static int mtrim(mstate av, size_t pad)
>  {
> @@ -4443,7 +4444,12 @@ static int mtrim(mstate av, size_t pad)
>                        content.  */
>                     memset (paligned_mem, 0x89, size & ~psm1);
>  #endif
> -                   __madvise (paligned_mem, size & ~psm1, MADV_DONTNEED);
> +
> +                   if (vrange_enabled) {
> +                     syscall(314, paligned_mem, size & ~psm1, 0, 1);
> +                   } else {
> +                     __madvise (paligned_mem, size & ~psm1, MADV_DONTNEED);
> +                   }
> 
>                     result = 1;
>                   }
> (END)

I can't find VRANGE_NOVOLATILE call in your patch so I think you just
test with enough memory. I expect you wanted just fast prototype to see
the performance gain. Yes, looks good although above code isn't complete
since it doesn't handle purged pages.

Next step of us would optimize reclaim path so when memory pressure is
severe, vrange could have a improved result rather than vanilla for
avoing swap out. Any suggestion are welcome.

-- 
Kind regards,
Minchan Kim
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists