differences between Opterons

[lindheim]

I started doing some testing with OpenSpeedShop, using Papi for hardware counters, and am finding inconsistencies between
the older dual core (275) and the newer quad core (2380) Opterons for our compute nodes in one of our clusters.
Going back to the Papi tests, I see that f.ex. for the ctests/flops, I get very different counts for the two.
The source for this test, states that the number of flops I should be seeing on Intel compatible architectures should be 2*(INDEX^3).
With INDEX=1000, I get Total flpins: 2000000000 on the quad core systems, but I get Total flpins: 6000000000 on the dual core systems.
I have tried playing with the environment variable PAPI_OPTERON_FP, to see it this will make a difference, no avail.
Any suggestions for how to make the two types of Opteron systems consistent?

Thanks,
Jan Lindheim

[vweaver1]

I started doing some testing with OpenSpeedShop, using Papi for hardware counters, and am finding inconsistencies between
the older dual core (275) and the newer quad core (2380) Opterons for our compute nodes in one of our clusters.
Going back to the Papi tests, I see that f.ex. for the ctests/flops, I get very different counts for the two.
The source for this test, states that the number of flops I should be seeing on Intel compatible architectures should be 2*(INDEX^3).
With INDEX=1000, I get Total flpins: 2000000000 on the quad core systems, but I get Total flpins: 6000000000 on the dual core systems.
I have tried playing with the environment variable PAPI_OPTERON_FP, to see it this will make a difference, no avail.
Any suggestions for how to make the two types of Opteron systems consistent?

Are you running the exact same executable on both systems? What compiler are you using? gcc is known to sometimes poorly optimize code and cause factor of 3 increases in floating point values like this. Moving from gcc-4.1 to say gcc-4.3 would cause the results you see.

[lindheim]

Yes, the exact same excutable and OS image.
The compiler is the default gcc (v4.1.2) with RedHat5.

Jan

[vweaver1]

From some quick research it sounds like your new processor is a Shanghai, which has an improved FPU that can perform operations in 128-bit chunks. The older AMD FPUs broke these into two 64-bit chunks. So my guess is you are seeing fewer floating point instructions because of this.

Part of the problem is that in papi_events.csv PAPI_FP_OPS and PAPI_FP_INS are defined the same way, which is why you are seeing fewer FP_INS even though the binaries are identical.

Vince