folding@home client cpu biased?

[safemode]

I dont know if it's just more convenient or not but it looks to me like the folding@home client at least in linux is geared more towards the P4 than the athlon. I'm wondering if it's because the client just has more sse2 development behind it or if the advanced instructions on the athlon cpus from tbird and up just cant compete head to head with sse2. and just what compiler does the folding@home client get compiled by?


just wondering.

[safemode]

holy crap the -rt command is causing my athlon to go into crack fiend speed mode. :) nice. much better compared to the P4 now.

[Chinasaur]

Safemode,

Glad you are using the -rt switch..it is faster because the entire client and protein data are in RAM.

Also, Athlons are better at DF than P4's. My 1.5Ghz Athlon XP's equal the output of a P4 2.0 or 2.1.

If you refer to it as the DF client people won't be confused...you used 'folding@home' which is another protein folding project.

Fold on :)


:cheers:

[JTrinkle]

on my 7 boxes I went from approximately 1,050,000/day to just over 1,600,000/day by

1) adding ram to make 256 meg on 2 boxes so I could use the -rt switch.
2) also using the -qt switch on the win98 boxes (more of a difference than you would think)

YMMV

-JT

[safemode]

xp's are probably a significant amount better than tbird's simply by being palamino / thoroughbred. That and having sse ability cant be bad either. My athlon tbird 1.13 is only (very) slightly slower than my P4 1.7Ghz.

For all purposes they're about equal in speed now. I've got 3 times the amount of processes on the athlon computer than the P4.

I dont think it's as simple as being loaded into ram. I'm getting something around 88MB of memory usage from the client now. Perhaps it's not writing to disk now that's different. Not the loading the data for reading into ram. we just dont use a scratch file on the hdd, it's all done in ram.

[safemode]

also, i dont really care that it makes things faster. I was just wondering why i was seeing the results i was seeing and it seems to simply be an interactivity thing with io on the two systems. I run distributed programs like this during the winter to add some heat to the dorm. nice silent and legal heater. keeps my cpu at around 40C so there's gotta be a lot of heat getting into the air which is good. The p4 is air cooled so ..heh. it will get into the mid upper 50C's but it's used to that. All i have it do is cpu bound stuff.

[bwkaz]

Regarding SSE, SSE2 and all that, I remember Howard saying once or maybe twice that the client depends a lot less on whatever fancy floating-point math any given processor can do, and a lot more on how fast it can just chase pointers. The vast majority of the DF client's time is spent chasing pointers, and a fast floating-point processor, while it will help a little, cannot make up entirely for a slow pointer-chasing unit (a.k.a. the integer math part of the processor).

If I remembered the thread, I'd link to it, but I don't. :o Maybe someone else does?

[Scoofy12]

A couple of answers:
The DF client doesn't use SSE or MMX extensions, most of the algorithm time is spent in pointer traversal, according to howard. see here (http://www.free-dc.org/forum/showthread.php3?s=&threadid=1044) or here (http://www.free-dc.org/forum/showthread.php3?s=&threadid=570)
The -rt switch causes the whole thing to be loaded into memory at once rather than (i assume) malloc()ing and free()ing all the time and hitting the disk.

You can download the linux version compiled with either icc (intel compiler) or gcc. The icc one is faster by a bit.

Fold on, keep em crankin' !