P4s Superior with Gromacs...or So We Thought

Leonardo

Below are two calculations for processing of identical Gromacs, Protein 385, p569_BBA5-ext. The data frame on the left is my hyperthreading P4; the right is my AMD XP. Usually, any comparison of Gromacs processing between these computers reveals the P4 wiping the floor with the Athlon. What is this reversal? Even if we take iinto account that the P4's p569 protein is only being Folded by 50% of the CPU's power (one service out of two Folding services on that computer), and we double the calculated points for a fair comparison, the AMD is clobbering the P4. Is this a new type of Gromac, not dependent on SSE2 for efficient processing?

hypermood

Just because they are running the same project does not mean that they are crunching the same data set. Obviously they are related sets (same project) but if you peek inside the log file you will see a line like:

[00:17:03] Project: 926 (Run 59, Clone 72, Gen 11)

If the two machines are not folding under the same constraints, you can't compare the frame times.

mmonnin

Wait a bit longer and see if its still like that. It shouldnt be. The test machine is a 2.8 P4 so any WU should produce at least 700 or so points on that machine.

Leonardo

I've been monitoring for over an hour. It's still 25 minutes per frame. I've never had Gromacs fold so slow on this machine.

Medlock

Leonardo wrote:

Even if we take iinto account that the P4's p569 protein is only being Folded by 50% of the CPU's power (one service out of two Folding services on that computer), and we double the calculated points for a fair comparison, the AMD is clobbering the P4.

Doubling the caluculated points wouldn't be accurate. Folding two instances doesn't get twice the production. More like 1.5-1.6 times normal production. I'll quote something thrax said in my thread a while ago.

Thrax wrote:

Well, yeah. It's not true SMP.. It's more like a bastardized, cobbled-together-because-the-p4-sucks, quasi-symmetric multi-thread(ish) execution engine.

Yeah, definitely Thrax!

Leonardo wrote:

Is this a new type of Gromac, not dependent on SSE2 for efficient processing?

I'm not sure. I can't remember if I've gotten any of these, but I don't remember this P4 ever folding any gromacs wu that slowly, besides the p212, 135 points. And that was hyperthreading, with two instances. You can check yourself to see if it is using SSE. Look in the FAHlog.txt file.

Leonardo

Well, yeah. It's not true SMP.. It's more like a bastardized, cobbled-together-because-the-p4-sucks, quasi-symmetric multi-thread(ish) execution engine.

Yeah, that's actually close to the truth. Hyperthreading was designed to keep the P4's long pipelines filled, in an attempt to compensate for the high latency times of pumping data through so complex a pipe.

1.5 or 1.6 the normal production? Sounds about right. But my point was that even if you compensate for the two instances running simultaneously, by extrapolating more points for slow Folding protein, it's still processing very slowly.

mmonnin

Dude WTF. Where is Thrax's post. I posted once in here earlier before and I came back and it was GONE. So I posted again. Are posts just disappearing?

Leonardo

I let the P569 run overnight; my initial inclination was to dump it. This morning EMIII reported that the unit was being processed at about 4 minutes/frame. Just a half hour later though, the frame times were back in the 20-minute zone. I checked the log file, and the protein had caused lots of errors during the night.

Well, I dumped it. Guess it was a corrupted work unit. Presently this Folding instance on the P4 rig is processing a P568_BBA5_N at 3:50/frame. Back in normal production efficiency now.

muddocktor

Yeah, it was probably a bad wu. It sounds like it was getting errors so it was disabling SSE, hense the very much greater folding times.