Intel HyperThread Question

[Celerator]

I was checking one of my borgs today, an Intel 2.8 with hyperthreading.
When I went into the Task Manager to see if it was running 100%, I noticed
that it was stable at 50%.

Is this correct?

 

[wraieghaeien]

Yes, because Hyperthreading tricks the OS into thinking there are 2 CPU's it will only show as 50%

Exactly how it should be.

 

[Celerator]

Thanks. That's good news. I thought for a minute that I'd been folding half-assed.

 

[Maximos]

You can run two instances of the program, utilizing the 2nd logical CPU.

 

[MN Scout]

You can run two instances of the program, utilizing the 2nd logical CPU.

Yes, that is an option, but you wont see any gain really in output. Maybe 15-20%. A good idea that has been proposed before it to install a second instance of F@H that is running timeless, and a setting of 15% cpu utilization. Then you'll get a few more points.

 

[Celerator]

Thanks. This forum is great for these timely answers to questions!

 

[IanG]

Yes, that is an option, but you wont see any gain really in output. Maybe 15-20%. A good idea that has been proposed before it to install a second instance of F@H that is running timeless, and a setting of 15% cpu utilization. Then you'll get a few more points.

Nice idea. I had been running two instances on my work box and I calculated it to be about 10% faster running two gromacs last week. Might switch to this setup now.

 

[wraieghaeien]

As an FYI, Stanford prefers only 1 unit to be run at a time on HT machines.
Yah its not best for points, but it is best for science, in that they get results
faster.

 

[IanG]

As an FYI, Stanford prefers only 1 unit to be run at a time on HT machines.
Yah its not best for points, but it is best for science, in that they get results
faster.

I've never really understood that - If I'm processing 2 units on my 2.8 P4, and get say a 10% boost from HT, that's still equivalent to two 1.5Ghz processor. Are Stanford saying they don't want people with older CPUs folding because it takes too long?

 

[SmokeRngs]

Yes, that is an option, but you wont see any gain really in output. Maybe 15-20%. A good idea that has been proposed before it to install a second instance of F@H that is running timeless, and a setting of 15% cpu utilization. Then you'll get a few more points.

Yeah, I did a bit of experimenting with this on my work machine which is a P4 2.8 with Hyperthreading. This is what I have run as long as I wasn't running QMD's. At around 15% CPU usage, you get some work done on a timeless while not affecting the frame times of the normal WU much.

It doesn't matter right now but do not run anything else when running QMD's. The QMD frame times are affected rather badly. This is only with 512 meg of RAM so that may have something to do with it.

Also, make sure e second instance is a timeless tinker. It does not use SSE or any other optimizations so the two proteins will not fight for the use of the optimizations. There can be a big difference in frame times if you are crunching on two proteins that use SSE since they have to fight over those resources.

I only have access to one P4 and the only access has been in the last year when I got a new system at work. Everything I own is either PII/PIII or AMD and yet I've done a lot more tweaking and testing with the P4 and it's my work machine. Yet, I think I have the best setup for a P4 to get some extra points and keep with Stanford's desire to get WU's back as fast as possible.

Right now it seems DGromacs with a second instance of a timeless tinker give the best points for a Hyperthreaded P4. I believe this 2.8 can pull around 260 PPD with that combination. Too bad it's chewing on a DbGromac.

 

[SmokeRngs]

As an FYI, Stanford prefers only 1 unit to be run at a time on HT machines.
Yah its not best for points, but it is best for science, in that they get results
faster.

That's the reason for running a second instance with timeless tinkers. Stanford has those units timeless for a reason, they are low priority and you can take your time doing them.

The tinkers don't use SSE like the rest of the units and so it doesn't slow the processing down much for the main instance when only run at 15%. I've done some testing on this using a production P4 2.8 with 512MB RAM. As long as the timeless tinker is running at around 15% and the main unit is an SSE optimized WU, there shouldn't be much difference in the frame times on the main unit.

 

[rayman2k2]

I've never really understood that - If I'm processing 2 units on my 2.8 P4, and get say a 10% boost from HT, that's still equivalent to two 1.5Ghz processor. Are Stanford saying they don't want people with older CPUs folding because it takes too long?

It's because running two instances on one core increases the potential for a miscalculation. Or so I've heard.

Miscalculations + Science = Bad

 

[Mohonri]

I've never really understood that - If I'm processing 2 units on my 2.8 P4, and get say a 10% boost from HT, that's still equivalent to two 1.5Ghz processor. Are Stanford saying they don't want people with older CPUs folding because it takes too long?

There are two cases here:
Case 1:
0 Hours: Stanford sends you the first WU
13 Hours: You send back the first results and get the second WU
26 Hours: You send back the second WU

Case 2:
0 Hours: Stanford sends you two WUs
24 Hours: You send back the results from both WUs

It seems to me that Stanford gains more by getting the first WU in 11 hours earlier than they lose by getting the second WU back two hours late. It would appear that some WUs have dependencies--the results from one WU make a second WU possible to do.

 

[wraieghaeien]

Mohori said it perfectly. Stanford gains more SCIENCE in getting 1 unit back every 10 hours, than it does getting 2 units back (at the same time from 2 instances on an HT) after 19 hours.