Benchmark Weirdness

[R.D. Silverman]

I managed to get access to two Intel Core Duo Windoze machines
over the weekend. One running at 1.8GHz, the other at 2.4GHz.

I ran two copies of my lattice siever on each.

The 1.8G machine took about 11.2 seconds per special q on each processor
and each processor ran at the same speed. This is about 15-17% faster for
each processor than each of two threads running on my P IV 3.4GHz
hyper-threaded machine.

However, something weird happened on the 2.4G machine. One processor
ran at 8.6 seconds per special q, while the other ran at 9.2 seconds per
special q. This was consistently the case all weekend.

Why might one core run 7% faster than the other on the 2.4G machine?
Why didn't I see the same thing on the 1.8G machine?

I have a *guess* (and it is only a guess) that for the faster machine,
each core was assigned a share of the L2 cache *statically*. i.e. the
process to start first grabbed a higher percentage of the available L2
cache and it stayed that way.... ???? Puzzled?

I was also surprised to see almost NO per-processor performance degradation
when two processors were running as opposed to one. I expected to see
a slowdown due to cache and memory contention. I saw none.

[xilman]

Originally Posted by R.D. Silverman

I managed to get access to two Intel Core Duo Windoze machines
over the weekend. One running at 1.8GHz, the other at 2.4GHz.

I ran two copies of my lattice siever on each.

The 1.8G machine took about 11.2 seconds per special q on each processor
and each processor ran at the same speed. This is about 15-17% faster for
each processor than each of two threads running on my P IV 3.4GHz
hyper-threaded machine.

However, something weird happened on the 2.4G machine. One processor
ran at 8.6 seconds per special q, while the other ran at 9.2 seconds per
special q. This was consistently the case all weekend.

Why might one core run 7% faster than the other on the 2.4G machine?
Why didn't I see the same thing on the 1.8G machine?

I have a *guess* (and it is only a guess) that for the faster machine,
each core was assigned a share of the L2 cache *statically*. i.e. the
process to start first grabbed a higher percentage of the available L2
cache and it stayed that way.... ???? Puzzled?

I was also surprised to see almost NO per-processor performance degradation
when two processors were running as opposed to one. I expected to see
a slowdown due to cache and memory contention. I saw none.

Did you run the benchmarks repeatedly, with other activity on the machine as well?

If you ran only once it's entirely possible that some other process, perhaps in the OS, was fighting one siever for the cache.


Paul

[Cruelty]

Do you know something specific about memory subsystem of each of this machines (+ FSB speed + chipset model). Perhaps in given configuration 1.8GHz nor 2.4GHz machine were able to fully utilize available badwidth. From what I have observed you will notice significant performance hit on C2D platform with CPU speed of 2.66GHz and memory @ DDR2-667 or 2.9-3GHz and memory @ DDR2-800.
As for the 7% difference between two processes I would first check if their priority was equal (e.g. idle) and then set affinity for both processes