[Beowulf] Woodcrest Memory bandwidth
John Cutonilli
jcutonilli at haifire.com
Mon Aug 14 16:52:35 PDT 2006
I have an older system, but I thought the results were interesting.
ifort -O3 -static stream.f mysecond.o
Function Rate (MB/s) Avg time Min time Max time
Copy: 2874.5953 0.0113 0.0111 0.0115
Scale: 2864.0447 0.0114 0.0112 0.0125
Add: 3467.2027 0.0143 0.0138 0.0172
Triad: 3471.2678 0.0142 0.0138 0.0150
ifort -O3 -static -xN stream.f mysecond.o
Function Rate (MB/s) Avg time Min time Max time
Copy: 4102.0088 0.0079 0.0078 0.0084
Scale: 4100.8808 0.0079 0.0078 0.0082
Add: 4437.9277 0.0110 0.0108 0.0112
Triad: 4431.2855 0.0110 0.0108 0.0112
It's pretty interesting the difference a small change to the compiler
options will give.
The 10 GB/s number may be a reference to the 10.5 GB/s maximum bandwidth per
processor the Woodcrest 1333 MHz FSB cpus are theoretically capable of
achieving. This is in contrast to the 6.4 GB/s numbers for the Opterons and
older Intels (PC-3200)
John Cutonilli
> Bill and I passed a few emails off the list and it seems that he's heard
> reports of up to 10 GB/s. I tried on 2 different Woodcrest systems (one
> from Dell, one directly from Intel) here and they both report back
> similar numbers. It seems that the fortran version of the program with
> pathscale 2.3 (the numbers I reported above) reports different numbers
> than the C version of the program with pathscale 2.3:
>
> Function Rate (MB/s) Avg time Min time Max time
> Copy: 6129.2441 0.2090 0.2088 0.2097
> Scale: 6138.1509 0.2087 0.2085 0.2091
> Add: 5935.6727 0.3236 0.3235 0.3237
> Triad: 5972.4347 0.3217 0.3215 0.3221
>
> and the C version with Intel 9.1:
>
> Function Rate (MB/s) Avg time Min time Max time
> Copy: 6145.6340 0.2084 0.2083 0.2090
> Scale: 4830.1172 0.2652 0.2650 0.2658
> Add: 5289.7542 0.3632 0.3630 0.3638
> Triad: 5296.5246 0.3629 0.3625 0.3640
>
> Thanks,
>
> --
> Jason Holmes
More information about the Beowulf
mailing list