HT on P4s

[Khemikal]

I keep seeing these people claiming to get 1100+ppw off of oced HT P4s. Does anyone have info on exactly F@H handles running multiples of itself on the same proc through HT. Im a previous Primester and the optimizations were so tight for SSE2 that running two instances was slower because of the communications doing the HT took up cycles that would have been used for prime. I understand it works, but Im wanting the 'Why?' answered if anyone is willing to give it a shot :D

[hkgonra]

I know there are some links floating around this forum somewhere ........ I do know it is faster.

[nikhsub1]

The basic version is that obviously HT does pretty well with being able to 'multithread' with 2 instances of HT. I see about a 20% improvement in PPW with running 2 instances over just one.

[FattyMcNastyXK]

well, i got a 2.6c HT and when i run one instance of folding, it only takes up 50% of my cpu, so i just run another and set the affinity of one folding core to cpu0 and the other to cpu1, I think it works great. I think i'd be getting that much if i ran it constantly, but I about 500 points for the weeks total and thats just running it at night, and i think i missed a couple of nights.
beat me nik

[Khemikal]

But does any1 know why it is better. How does F@H handle the HT thing and why 2 threads with HT is handled better than 1 thread w/o HT.

[FattyMcNastyXK]

well, it handles it well because its like 2 cpus except i don't know how fast each cpu is going, maybe each cpu is like half of your whole cpu.

[Bobby]

Originally posted by Khemikal
But does any1 know why it is better. How does F@H handle the HT thing and why 2 threads with HT is handled better than 1 thread w/o HT.

There is a diagram showing why HT is better but I can't find it but think of it this way...

There are 10 cycles per millisecond (for arguement's sake), each cycle representing a process unit. Out of the 10 cycles, 8 or 9 of those cycles are used even though task manager shows 50% (100% if HT was disabled) usage with 1 instance of F@H. With HT, the 2nd process thread will fill up the empty cycles going through the processor making the cpu even more efficient.

Then there is the down time between WU's that the cpu is idle unless you have a 2nd instance using those cpu cycles and possibly a few cycles in between each frame within the WU.

I questioned this when I first started folding and I get an extra 1 or 2 WU's turned in every week over just running 1 instance.

-Bobby

[Khemikal]

If only 8-9 of the cycles are being used then the claim that it uses all remaining cycles is not a true good claim. If this is so the code needs to be tighter. I realize this might be asking some work, but that is a claim it should hold. Having to run 2 instances shouldnt be the only way to fill all cycles.

I might be misunderstanding you, but this is getting closer to an answer im looking for. Lemme see if i can find a link to some prime talk that is close to what im looking for a good answer about F@H.

[nikhsub1]

From this thread: http://forum.oc-forums.com/vb/showthread.php?s=&threadid=227195

There is much confusion about HT and folding and what it does and does not do... I have in the past recommended NOT to run 2 instances with an HT machine. Well, I am now changing that recommendation! Thanks to Plext in a thread HERE at OCAU (must be registered to view, I recommend you do) he tested one and two instances on the same CPU with AND without HT. The findings are as follows:

1. HT on and HT off with only ONE instance fold AT THE SAME SPEED! This means (like I have always said) that with HT enabled, even though taskmanager says 50% Load, your full CPU is getting used.

2. HT on with 2 instances results in a 15 - 20% gain in production than running just one instance!

Botton Line, Run 2 instances with an HT enabled machine!!!!

[Bobby]

Khemikal - The code is pretty damn good. Most applications only use 5-7 cycles which is why Intel started playing with HT. Even though your cpu is at 100%, it still waits for data to process.

15-20% gain sounds about right from my experiences.

-Bobby

[belorsch]

So 5 HT P4's with 2 clients running would net red status? Is that a P4 running at 3.0Ghz. If so is there an XP chip out right now that will net the same or close stats? I might have to look into going to P4's if not.

[nikhsub1]

Originally posted by belorsch
So 5 HT P4's with 2 clients running would net red status? Is that a P4 running at 3.0Ghz. If so is there an XP chip out right now that will net the same or close stats? I might have to look into going to P4's if not.
Well my P4 @3.6 yeilds me over 1200 PPW with 2 instances. I think 5 3.0's would get you VERY close to 5k/week

[hkgonra]

So you should expect between 800-1000 ppw on a HT p4 somewhere between 2.8 and 3.2 ? I might be adding at least one and maybe as many as 4 in the near future. :)

[nikhsub1]

Originally posted by hkgonra
So you should expect between 800-1000 ppw on a HT p4 somewhere between 2.8 and 3.2 ? I might be adding at least one and maybe as many as 4 in the near future. :)
Should be if those suckers get all gromacs! If you get some tinkers, which i have not seen on any boxen in a LONG while, I would think 800PPW would be a minimum.

[MLMIB]

I work at best buy, we're all getting 3 ghz p4's(intel deal) they're all gunna be folding(gotta stave off that frat folder some how :-/ )

[hkgonra]

I would KILL to get that deal !!!

You are one lucky man.

[dagamore]

when i upgraded from a 2.5 ghz p4 (2.0 eng smpl oc to 2.5) to a 3.0ghz p4c (800fsb) my time per wu on that machine with no other changes, droped about 2 hours on a 53.3 pointer, ie from 25.25 hours to 22.5 hours per wu, but i can run two consoles at ~ the same time per frame. I love my HT chip it rocks like a mo fo :D

[hkgonra]

Looks like I am only going to get 4 2.4ht's now :(

I guess that is better than nothing.

[Blunt]

how are you running more then on f@h when i try to start another it tells me there is already one running and closes the new one

[hkgonra]

You have to make the second one a different machine id.

[Blunt]

uhh... how?

[Bobby]

You need 2 separate directories for each instance.

-Bobby

[Blunt]

ahh so just install again

[Bobby]

Actually, just put a copy of FAH3Console.exe in it's own directory and execute it. Not much of an install.

-Bobby

[Blunt]

i had already reinstalled f@h to a diff directry and it still won't let me run 2

[Blunt]

ok i have one f@h windows version running. and one onclick running how do i make it so my oneclick is not minimized but hidden in the background