[petsc-dev] Fwd: [Xlab] slow pagefault on KNC

[Karl Rupp]

Assuming that the penalty for page faults reduces ideally with the 
number of threads, one would obtain an average ~4us with 240 active 
threads. Nevertheless, 850 us is terribly large, so Amdahl's Law will 
hit hard...

Since I don't think the Xeon Phi will be the last type of accelerator 
from Intel, why not include such a test? I don't know how portable this 
is, though...

Best regards,
Karli


On 12/17/2012 11:23 AM, Jed Brown wrote:
> He should be using all the threads for this. Using one thread to fault a
> bunch of memory is a recipe for terrible performance all around.
>
>
> On Mon, Dec 17, 2012 at 9:21 AM, Barry Smith <bsmith at mcs.anl.gov
> <mailto:bsmith at mcs.anl.gov>> wrote:
>
>
>         Should we have a test like this in the PETSc benchmark directory?
>
>         Barry
>
>
>     Begin forwarded message:
>
>      > From: Kazutomo Yoshii <kazutomo at mcs.anl.gov
>     <mailto:kazutomo at mcs.anl.gov>>
>      > Subject: [Xlab] slow pagefault on KNC
>      > Date: December 17, 2012 10:48:06 AM CST
>      > To: "xlab at cels.anl.gov <mailto:xlab at cels.anl.gov>"
>     <xlab at cels.anl.gov <mailto:xlab at cels.anl.gov>>
>      >
>      > Hi,
>      >
>      > I noticed that pagefault is very slow on KNC.  It takes 850 usec
>      > while it takes ~1 usec on Xeon, so prefaulting 1GB of memory region
>      > takes 222 sec on KNC.
>      >
>      > This may not impact a big app runs for hours and hours, but I guess
>      > this definitely affects short-lived processes or threads, which might
>      > make MIC less fascinating.
>      >
>      > This could be a hardware problem(need to check Phi), a kernel bug,
>      > or maybe sage of SMP kernel on many-core.
>      > If this is the last case, it would be really interesting for me.
>      >
>      > Attached a simple page fault benchmark.
>      >
>      > - kaz
>     [see attached file:
>     pftest.c]_______________________________________________
>      > Xlab mailing list
>      > Xlab at lists.cels.anl.gov <mailto:Xlab at lists.cels.anl.gov>
>      > https://lists.cels.anl.gov/mailman/listinfo/xlab
>
>