Re: OT: SETI@HOME on Alphas
Delete now if you don't care about SETI@HOME.
The "fast" client that was giving sub 1 hour work unit times on EV6
Turbolasers was an optimized version of the 1.3 client. In that version
of the SETI@HOME client, they were using a particularly inefficient FFT
Some of us Compaq folks noticed that SGI was posting some "hot" numbers
and we took that as a challenge. There was no way a MIPS cpu was going
to smoke an Alpha! Persumably, they did the same thing we did -
substituted a different FFT algorithm. We just used the FFT functions in
the CXML library. We did the port and tested it and it checked out fine.
We submitted the binaries (for Tru64, AlphaNT and OpenVMS) back to SETI,
but they never released it to the public. Most folks here inside Compaq
ran that one until they stopped accepting results from the 1.3 client.
When it came time to update the client, the SETI folks made Compaq and
SGI promise not to substitute different FFT's in the interest of getting
consistent scientific results. We agreed and presumably SGI did too. The
SETI project had gotten a new/better FFT algorithm anyway which wasn't
as bad as the original and in other benchmarks, gave comparable
performance to the FFT routines in the CXML library. This sped up the
processing time, but they also gave the client more work to do at the
same time. So, the actual time to process a work unit didn't drop all
that much (comparing unoptimized clients here). They added even more
processing in the 3.0+ clients as machines got faster.
So, that is the history. The current clients are built from the exact
same source as all the other platforms. They are all built with the
Compaq "cxx" compiler. On Linux, the client is linked statically since
not everyone has the Compaq "libcpml" and "libots" installed (like
Debian users mainly; RedHat and SuSe users usually install them).
If you want to know more, contact me "off list" so we don't clog up the
Debian list with OT SETI stuff.
> -----Original Message-----
> From: Stefan Schroepfer [mailto:firstname.lastname@example.org]
> Sent: Monday, February 25, 2002 5:09 PM
> To: Donsbach, Jeff
> Cc: email@example.com; Debian-Alpha
> Subject: OT: SETI@HOME on Alphas (was: Re: flakey miata 433a)
> Donsbach, Jeff wrote:
> > depends what version of the SETI@HOME client you are talking about..
> Jeff, this is going more and more off-topic, but
> could you please drop a few words about ancient
> seti@home clients here ...
> There are some "about one hour" results out there
> calculated on ev67 Alphas using a optimised 2.4(?)
> Seti@Home client that utilised DEC/Compaq FFT
> libraries ... is that right so far? (Obviously
> Berkeley does not allow such vendor-optimisations
> at this time.)
> If that is true, how large was the benefit using
> these libraries compared to a unmodified Berkeley
> distributed S@H client at that time?
> Stefan Schroepfer
> (who is running the S@H client on two Alphas using
> Debian and NT4 ...)