[Beowulf] Barcelona numbers

[Vincent Diepeveen]

hi Bill,

thanks for your input.

that simple C program that measures latency,
can you try it with a more realistic working set size also
to measure RAM latency, so with like 2GB in total or so?

That's real interesting to compare to intel and 32MB though interesting is 
far less interesting than the improvement that DDR2 ram gives the barcelona 
over higher clocked opterons with the old DDR ram.

note i wrote a test proggie that is doing the same thing in a bit more 
complex manner (using shared memory).

For many people here the random read latency to BIG ram sizes is quite 
important.

That's the kind of test that you don't see soon online, they just are busy 
with the 150MB max that spec gives you, which is a joke and soon gives too 
much cheating for spec (as we can already see for some spec tests now which 
use something that fits nicely within combined level caches). Back in 1999 
even at pc's i ran with more than that for my crunching software (chess, 
namely i used 400MB).

Today several GB's of ram is interesting to measure.

Regards,
Vincent

----- Original Message ----- 
From: "Bill Broadley" <bill at cse.ucdavis.edu>
To: "Beowulf Mailing List" <beowulf at beowulf.org>
Sent: Monday, September 10, 2007 9:45 PM
Subject: [Beowulf] Barcelona numbers


>
> Finally the NDAs have expired and there's tons of technical info available 
> on
> many hardware websites.  I figured I'd post some more numbers.  In general
> I'm impressed with the bandwidth (considering it uses the same dimms), and
> with the parallelism in the memory system.
>
> Dual socket quad core opteron 2350's (2.0 GHz) running the current 
> McCalpin'S
> STREAM compiled with pathscale-3.0 -mp -O4:
> Total memory required = 228.9 MB.
> Function      Rate (MB/s)   Avg time     Min time     Max time
> Copy:       15355.3139       0.0104       0.0104       0.0105
> Scale:      15249.5885       0.0105       0.0105       0.0105
> Add:        14954.2883       0.0161       0.0160       0.0162
> Triad:      15061.2389       0.0160       0.0159       0.0160
> -------------------------------------------------------------
> Solution Validates
>
>
> (no source changes except for the N= line).
>
> I latency benchmark I wrote that for for each thread accesses
> 32 MB randomly via:
>
> while (p != 0)
> {
>      p = a[p];
> }
>
> thread=0, 7.293 seconds, latency=108.48 ns
> With 1 thread(s), max latency was 7.293 seconds, effective latency=108.48 
> ns.
>
> thread=0, 7.300 seconds, latency=108.59 ns
> thread=1, 7.290 seconds, latency=108.44 ns
> With 2 thread(s), max latency was 7.300 seconds, effective latency=53.84 
> ns.
>
> thread=0, 7.418 seconds, latency=110.35 ns
> thread=1, 7.363 seconds, latency=109.53 ns
> thread=2, 7.422 seconds, latency=110.41 ns
> thread=3, 7.359 seconds, latency=109.47 ns
> With 4 thread(s), max latency was 7.422 seconds, effective latency=26.91 
> ns.
>
> thread=0, 7.417 seconds, latency=110.33 ns
> thread=1, 7.394 seconds, latency=109.98 ns
> thread=2, 7.433 seconds, latency=110.57 ns
> thread=3, 7.382 seconds, latency=109.81 ns
> thread=4, 7.448 seconds, latency=110.79 ns
> thread=5, 7.379 seconds, latency=109.76 ns
> thread=6, 7.443 seconds, latency=110.71 ns
> thread=7, 7.411 seconds, latency=110.24 ns
> With 8 thread(s), max latency was 7.448 seconds, effective latency=13.03 
> ns.
>
> For comparison an opteron 270 (2.0 GHz as well):
>
> thread=0, 6.301 seconds, latency=93.72 ns
> With 1 thread(s), max latency was 6.301 seconds, effective latency=93.72 
> ns.
>
> thread=0, 6.302 seconds, latency=93.73 ns
> thread=1, 6.276 seconds, latency=93.35 ns
> With 2 thread(s), max latency was 6.302 seconds, effective latency=46.47 
> ns.
>
> thread=0, 12.071 seconds, latency=179.55 ns
> thread=1, 12.084 seconds, latency=179.75 ns
> thread=2, 11.952 seconds, latency=177.79 ns
> thread=3, 12.055 seconds, latency=179.31 ns
> With 4 thread(s), max latency was 12.488 seconds, effective latency=45.27 
> ns.
>
>
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit 
> http://www.beowulf.org/mailman/listinfo/beowulf
>