Single AMD = Great for 2 Instances

[CaptainMooseInc]

I've been testing out FaD on my Sempron 2800 (the renamed Athlon XP 2800+ because AMD wants to save the Athlon name for its high-end 64bit processors) and have discovered that FaD runs better with 2 instances than just 1. For the time it takes my processor to complete 1 big WU, I can process 2 big WUs using TWO instances in roughly the same time (it takes about an hour and a half longer to finish up the second WU).

To make the program even quicker, do a CTRL+ALT+DEL and set the priority for the instance that is running at the lowest CPU percentage to ABOVE NORMAL. This will make the 2 instances of THINK.EXE run right around 50% each.

I've managed to up my production of WUs from my uno machine by about 50%.

I don't know about Intels in this matter...

I don't own one.... :moon:

[redukt]

About to try this out on my AMD64 3200+.

:spank:

God I love that icon.

Update: When you changed to 2 instances, did you see the CPU rating drop (in FaDSpy or wherever). From what FaDspy is telling me I have two CPUs with roughly the same rating on them.

[CaptainMooseInc]

Just looked at that for ya.

Yeah, I did see a drop. Normally I have a CPU rating of 180 or so (168 for big monster WUs).

In two instances the CPU ratings are 168.

In my mind, that makes me think my CPU rating is 336 with two instances, just a little less than double the 180.

Do you agree with this train of thought?

I don't care what the explanations behind everything are, all I know is that I'm putting out a lot more WUs than I have been.

-Jeff

[redukt]

I would agree, though I do not pretend to understand how its possible. WUs here I come.

(BTW, I added another boxxen while you guys weren't looking). Just some 1.4Ghz (or so) machine at work that runs all the time).

[CaptainMooseInc]

have you seen an increase yet?

[redukt]

Not yet. I ran into a problem - my setup has my main AMD machine as the server, and when I tell it to run two instances of think, so does every client machine. Ugh.

[CaptainMooseInc]

Interesting....

Umm...I would tell you how to belly around that, but I can't think of how....

-Jeff

[jasong]

I'm not saying it isn't true(when I tried it, I didn't change the priority from low), but from my experimentation, the overall rating drops.

Suggestion: Try timing how many seconds actually pass in one of the Think instances compared to clock time. Your advantage may be imaginary. On the other hand, the Priority change you mentioned may be the key.

[CaptainMooseInc]

I've noticed that running 2 instances w/o changing priority usually means 1 of the 2 instances hogs about 67-82% of the total available cycles during idle periods. By upping the priority on the less-fortunate instance, they trade off at around 48-52% each. My CPU rating for a single instance is usually 180. Since I've started using 2 instances no rating has been below 120 per instance and I'm still outproducing my former 1 instance self.

I figure that if I have 2 instances at 150 (usually at 168), then my rating is 300 instead of the normal 180 for me.

True, my findings could be pure fictional and rationalization...but I had to explain the extra WUs I'm putting out somehow....

-Jeff

[jasong]

I'm going to try it using the Priority advice posted. Many things, including the 256K cache size, could also play a part.

See you in about 4 days, guys and gals. :)

[magnav0x]

Yeah, I did see a drop. Normally I have a CPU rating of 180 or so (168 for big monster WUs).

In two instances the CPU ratings are 168.

In my mind, that makes me think my CPU rating is 336 with two instances, just a little less than double the 180.

Seeing how you are processing at about 50% you would actually end up with about the same CPU rating if not lower.

168 / 2 = 84 CPU rating each at 50%

84 x 2 = 168 CPU rating at 100%.

Due to this, I would think the point totals would actually probably end up about the same. You may finish 2 WU's in the same time, but I would imagine it would equal out, because each instance is taking twice as long to complete it's unit.

[Paratima]

Remember that FAD wu's vary wildly in their total processor workload. Even wu's on the same disease molecule, but running different drugs may turn in quite different runtimes, by hours or even days.

If you think you can get 120% out of a CPU's duty cycles, I've got a really sweet real estate deal I'd like to talk over with you. :cool:

The sad truth is that running two instances is inevitably going to run slower on a single CPU than one instance. The main reason is a bugger called "context switching". One CPU has one instruction pipe and one data pipe, and can think about one thing at a time, period. When you interrupt it to think about something else (like your 2nd instance), there's overhead to pay. Every time, always, zero exceptions.

Nice try, though. :p

[CaptainMooseInc]

I see your reasoning Paratima, but what I still don't get is the fact my WUs aren't taking any longer to get done. I'm putting out 2WUs in the same amount of time it took me to put out 1 WU. Maybe an extra 10mins of crunching at max. I don't run any of the heavy-duty long-running WUs from HIV and Cancer, just Proteome and the quicker Cancer targets. These aren't showing any signs of more time needed to complete than it would just 1 WU.

In my little happy world it's working for me and I don't know why. I'm probably just too dumb to understand what you're trying to stick through my head. :bonk:

Maybe I'll run 2 HIV WUs and see the time it'll take (I already know it'll double). For the short WUs it isn't taking any longer really.

Thanks,
-Jeff

[Paratima]

It's only going to save time <<<IF>>> the client loses time sleeping somewhere along the line. Possibly it naps before / during / after starting / completing / sending / downloading a new WU. IF that's the case and the 2nd instance then uses what would otherwise be "slack time", then I'd believe it made a difference. Normally, I see my CPU at 99% usage & it doesn't get any better than that. Two processes are just going to fight over that 99%, unless the OS is badly misreporting CPU usage. ;) (OK, so I suppose it's possible.)

[CaptainMooseInc]

I believe there may be some slacktime inbetween jobs that the other WU could use as a "catchup" period.

Make note though that my processes aren't really fighting over CPU time. I've allocated a higher prority to the less-fortune of the WUs to make them both run at about 47-52% of the total 99% available CPU cycles. They trade off right in that area to make up the 99%.

-Jeff