HyperThreading-enabled computer: Factoring & sieving

[Mystwalker]

Disclaimer: I couldn't decide for a subforum to put this in and didn't want to open 2 threads. If someone cares, change it as you like.

Whoever uses a P4 with hyperthreading (and is into the factoring and/or sieving subproject) might consider running a copy of each client simultaneously.

My experience (3 GHz, 800 MHz FSB):

2x sieving = 2x265 kp/s

2x factoring = 83 min / 2 tests = 41.5 min / test


1x sieving + 1x factoring = 265 kp/sec + 62 min / test

Looks like a decent improvement (not considering the effectiveness of factoring vs. sieving).

Any other experiences?

[jjjjL]

I run two factoring threads on my hyperthreaded P4 2.8GHz.

Stage 2 puts pressure on memory so it's best not to have the clients running in lock step with each other. I started one client and then started the other when the first started stage 2 of it's first test. This has kept the overlapping on stage 2 to a minimum and probably increased performance substantially. It probably makes a big difference. How much memory were you allocating to each factorer and how much total does your system have?

Cheers,
Louie

[Mystwalker]

This system has 1 GB of memory. Each factorer gets 256 MB.

Stage 2 puts pressure on memory so it's best not to have the clients running in lock step with each other.

Hm, that could be the explanation, as typically both factorer processes get nearly the same amount of CPU load (their percentage stays the same over several tests). That way, the time to compute a test stays constant as the overlapping stays the same (100% overlapping ).
Now I understand your idea of a modified multithreaded factorer, which could rather start a third factoring in stage 1 than run a 2nd factorer in stage 2.

[mklasson]

I'm afraid proth_sieve v0.40 is a bit wacky when it comes to rate calculation. On Linux and FreeBSD it only considers the time taken by the current process (i.e. the displayed rate is not affected by other running programs). It happened when I rewrote the timing routines to make them compile under freebsd. Didn't know about it at the time... I'll try to fix it in the next release.