Re: 2.6.38 sbrk regression

[raz ben yehuda <raziebe@gmail.com>]

oddly I am failing to compile 2.6.38-rc, I am getting:
linux-2.6]$ make bzImage
  CHK     include/linux/version.h
  CHK     include/generated/utsrelease.h
  CALL    scripts/checksyscalls.sh
  CHK     include/generated/compile.h
  CC      arch/x86/lib/memmove_64.o
gcc: arch/x86/lib/memmove_64.c: No such file or directory
gcc: no input files
make[1]: *** [arch/x86/lib/memmove_64.o] Error 1
make: *** [arch/x86/lib] Error 2

and there is no such file. any idea ?

thank you

On Wed, 2011-04-13 at 14:51 +0200, Andrea Arcangeli wrote:
> Hi,
> CC'ing Mel.
> 
> On Wed, Apr 13, 2011 at 02:03:58PM +0300, raz ben yehuda wrote:
> > Andrea Hello
> > 
> > I am running the AIM benchmark suite looking for regressions in 2.6.38.
> > The test page_test tests sbrk(). 2.6.38 is 13% less than 2.6.27 and 11%
> > less than 2.6.37.
> > The regression starts somewhere in the THP patch:  
> > SHA1 4e9f64c42d0ba5eb0c78569435ada4c224332ce4  to
> > SHA1 152c9ccb75548c027fa3103efa4fa4e19a345449 . 
> > 
> > I cannot git bisect it any more as the kernel won't compile.
> > Even if i disable THP in the kernel I still get a regression. 
> > I performed the benchmark on Xeon blade 3GHZ. But it also happens in
> > other types of machines.
> > 
> > 2.6.38
> > Apr 13 13:31:16 2011 AIM Independent Resource Benchmark - Suite IX 5120
> > page_test 30010 85.0716       144621.79 System Allocations &Pages/second
> > 
> > 2.6.37
> > Apr 13 13:44:39 2011 AIM Independent Resource Benchmark - Suite IX 5120
> > page_test 5020 100.797       171354.58 System Allocations & Pages/second
> > 
> > I am going to profile it hopefully I will have more information.
> 
> The compaction kswapd caused regressions in fs benchs like specsfs,
> it's fixed in 2.6.39-rc. Obviously it will go away if you set
> CONFIG_COMPACTION=n, but 2.6.39-rc should work best with COMPACTION=y
> too. Could you try latest 2.6.39 git?
> 
> Also please make sure CONFIG_SLUB=n and CONFIG_SLAB=y. We must fix
> SLUB so it stops allocating with GFP_KERNEL in the higher order alloc,
> I think it should only do a quick check of the buddy with
> GFP_ATOMIC. The pageblock types are the thing that prevents
> fragmentation, no need of special logic for that in slub. It's
> unlikely that for short lived allocations succeeding the high order
> allocation provides enough speedup to the code using the memory, in
> order to at least break even with the cost of compacting the memory to
> succeed the allocation.