[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: UP1500 Openssl speed

On Wed, 2003-05-07 at 04:18, Matt Simis wrote:
> "Thomas Evans" <tom@23palmer.net> wrote in message
> news:20030506231011$6e96@gated-at.bofh.it...
> > >Interesting that the Alpha consistantly outperforms the Athlon on
> > >"verify/s" but looses ground on signs/s.
> >
> > It's been a while, but one of my tasks before
> > leaving Compaq was analyzing openssl for
> > potential optimizations to boost it's
> > performance on Alpha.
> >
> > The lack of rotate instructions seemed to
> > be a big factor in performance differences, IIRC.
> >
> >
> > ...tom
> Out of interest, did these "potential optimizations" translate into
> realworld improvements or just for OpenSSL benchmarks, kind of like the way
> everyone optimises (well, maybe not AMD) Spec score?

We were using the OpenSSL benchmarks to start.  All the
algorithms are integer based and many already had
support for 64bit integers when appropriate, so
most of the Alpha-isms were taken care of by a good
compiler and the right build switches.

The pattern that emerged was that the algorithms
that made heavy use of rotates were the dogs on Alpha.
We concentrated on those, but only on finding a way
to better schedule the instructions for performing
the rotates.

We concluded that the better thing to
do would be to find someone that understands these
algorithms and rewrite them without rotates.
Similar to the way the DES cracking contest
eventually came up with an algorithm for machines
with many registers, we felt that the best
benefit would come from rethinking the algorithms.
(Or convincing the Alpha architecture folks that
 rotates were a good idea....).

This wasn't an effort of the compiler or benchmarking
groups, we ere looking for ways to show that
Alpha was significantly better for certain tasks
and this was an initial investigation into
crypto uses.


Reply to: