Integration of ProPolice in FreeBSD
Steve Kargl
sgk at troutmask.apl.washington.edu
Sat Apr 19 15:56:50 UTC 2008
On Sat, Apr 19, 2008 at 09:46:38AM +0200, Jeremie Le Hen wrote:
> Hi Steve,
>
> On Fri, Apr 18, 2008 at 05:15:55PM -0700, Steve Kargl wrote:
> > On Fri, Apr 18, 2008 at 10:47:38PM +0200, Jeremie Le Hen wrote:
> > >
> > > Certainly. I would like to hear opinion from other committers if SSP
> > > should be enabled by default.
> >
> > I'm not a committer, but I'll ask a question anyway.
> >
> > Can you quantify the performance impact, in particular for
> > numerically intensive codes with heavy use of libm?
>
> I don't run such application, so I can't answer. Sorry. If you are
> willing to give a try, I would be pleased to help you to run your tests,
> or even run them on my side.
>
> BTW for the sake of my curiosity, is there a technical reason for
> ProPolice to be heavier for libm?
>
Most numerical applications, that I'm familiar with, tend
to contain nested loops that make calls to functions in
libm. Simple example in one of my codes is a 3 deep loop
that computes what is known as the thermal dose.
for (k = 0; k < kmax; k++)
for (j = 0; j < jmax; j++)
for (i = 0; i < imax; i++)
td += exp(a * b[k][j][i])
Now, put the above loops inside a time loop with n time steps.
exp() will be called kmax*jmax*imax*n times where this product
can be quite large (order of 5e11). Any overhead caused by PP
will increase the simulation time. A 1% increase in time is
probably tolerable, but a 10% increase would be detrimental to
simulations that takes days to complete (yes, I have a few that
run that long).
I'll see if I can get you some numbers this weekend.
--
Steve
More information about the freebsd-arch
mailing list