curious performance enhancement or punishment?

Astro
Astro
Joined: 18 Jan 05
Posts: 257
Credit: 1,000,560
RAC: 0

OK found a break(made one as

OK found a break(made one as I'm a supervisor...hehehehehe), Found the average cpu time for Primegrid PSP Sieve wus.

Intel Q6600, 817 seconds
AMD 9950one, 890
AMD 9600, 1014
AMDX2 6000, 827

The Sieve application IS a 64b one, and yet the X2 6000 (K8) almost goes core for core with the Q6600. Again, not a very big disparity compared with that I witness at Einstein. Will look at Docking next.

I'm not sure how to do docking. So far I only have 16 wus completed on teh Q6600 and so only collected the most recent 16 from the others. The wus runtime varies greatly I see variations as shown on the hosts:

Intel Q6600 1631 - 3378 seconds
AMD 9950one 2877 -4584
AMD 9600 1502 - 4905
AMD 6000 2986 - 4286

perhaps I should wait until I have a much larger sample? Perhaps the different types of wus take different amounts of time to process?

If I left it to the 16 samples the averages would be:

Intel Q6600, 2965 seconds
AMD 9950, 3985
AMD 9600, 3999
AMD 6000, 3621

I'll recalculate this tommorrow or so, when I have a larger base. For Example, the times for the 9600 include these 16:
4410, 3814, 4585, 4549, 4284, 4905, 4260, 4404, 3317, 1502, 3544, 4661, 4342, 3166, 4042, and 4199.

Astro
Astro
Joined: 18 Jan 05
Posts: 257
Credit: 1,000,560
RAC: 0

does anyone know how much

does anyone know how much memory an Einstein WU takes?

Good news, the installation of 32b libs has allowed Einstein to run on the 9950's. Currently at 03:10:00 and 33% comp.

Brian Silvers
Brian Silvers
Joined: 26 Aug 05
Posts: 772
Credit: 282,700
RAC: 0

RE: does anyone know how

Message 85892 in response to message 85891

Quote:
does anyone know how much memory an Einstein WU takes?

On my system the Windows SSE app (the "_1" app) is currently fluctuating between 64 and 70MB...

Brian Silvers
Brian Silvers
Joined: 26 Aug 05
Posts: 772
Credit: 282,700
RAC: 0

RE: does anyone know how

Message 85893 in response to message 85891

Quote:

does anyone know how much memory an Einstein WU takes?

Good news, the installation of 32b libs has allowed Einstein to run on the 9950's. Currently at 03:10:00 and 33% comp.

Well, my first task on my 3700+ took 40,833.34 seconds (11.34 hours). As was pointed out, I have no idea if that is a runtime maxima, minima, or somewhere in between... I'm under the assumption that Bikeman doesn't overclock. If that is true, then I have a 25% clock advantage over a single core from a 175 (nominal for a 175 and a San Diego 3700+ is 2200MHz, mine runs at 2750MHz).

Plan: Run a few more to try to see where the cyclic variance is currently at in the current frequency template. Later, try changing "AuthenticAMD" in the binary to be "AuthenticABC" (the "Naughty Intel" test).

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3,516
Credit: 468,471,303
RAC: 58,424

RE: I'm under the

Message 85894 in response to message 85893

Quote:

I'm under the assumption that Bikeman doesn't overclock. If that is true, then I have a 25% clock advantage over a single core from a 175 (nominal for a 175 and a San Diego 3700+ is 2200MHz, mine runs at 2750MHz).

Correct, I never overclock. But sometimes I try out things software-wise, so my hosts might not be perfect comparison targets. But anyway, I don't own a 175, my best AMD box is an X2 4580e @ 2.5 GHz. Currently none of my hosts is crunching under Windows.

Quote:

Plan: Run a few more to try to see where the cyclic variance is currently at in the current frequency template. Later, try changing "AuthenticAMD" in the binary to be "AuthenticABC" (the "Naughty Intel" test).

Ok, anybody who would like to do the same: note that you have to prepare an app_info.xml file first, to prevent BOINC from applying the MD5 checksum check on the executable (a modified version would fail the test, probably causing an automatic re-download of the app files).

CU
Bikeman

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3,516
Credit: 468,471,303
RAC: 58,424

RE: But anyway, I don't

Message 85895 in response to message 85894

Quote:

But anyway, I don't own a 175,

Upps...after re-checking....I do indeed have a 175 among my hosts :-), I'm sorry. Actually it is a rented machine at a hoster , and it is really a virtual server where I get about half a core. This one was completely off my radar, so please forgive my ignorance. Because E@H is run in a VM, I would not use that box as a benchmark either, tho.

CU
Bikeman

archae86
archae86
Joined: 6 Dec 05
Posts: 3,010
Credit: 4,853,582,067
RAC: 3,344,246

RE: Well, my first task on

Message 85896 in response to message 85893

Quote:

Well, my first task on my 3700+ took 40,833.34 seconds (11.34 hours). As was pointed out, I have no idea if that is a runtime maxima, minima, or somewhere in between...


In between. The frequency was 847.70, and sequence number 596.

According to the cycle length estimating function I posted a while back, the expected cycle period for 847.70 is 210.4, so sequence number 596 is about 83% of the way through the third cycle.

As your next three results are sequentially lower in the same frequency, I'd expect them to drop in time requirement at a nearly linear rate. (I think you are far enough away from the minimum not to have much of a wiggle problem)

However, your other results in queue on that host are from two other frequencies, and while I have pretty good handle on the cycle length, there is definitely a dependence on frequency for which I don't even have the beginnings of a model.

So, for your goal, I suggest running just one more result before switching to the patched executable, which will mean you have two remaining on which to test it.

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3,516
Credit: 468,471,303
RAC: 58,424

There's also the reference

There's also the reference unit posted by Bernd a while ago and the Windows batch file using this (see later in the same thread). This eliminates the need to care about cycle periods etc. The ref unit was designed for S5R3, but should be pretty representative for S5R4 as well.

CU
Bikeman

Donald A. Tevault
Donald A. Tevault
Joined: 17 Feb 06
Posts: 439
Credit: 73,516,529
RAC: 0

RE: RE: Maybe I'll try a

Message 85898 in response to message 85885

Quote:
Quote:

Maybe I'll try a Linux VM later ^^

Good luck with that. I tried that about a year ago with very poor results. I was using a VMware workstation (6.0.2, I think) with an Ubuntu 7.04 and 7.10 guest. Performance was in the tank due to overhead from the VM.

I'd be interested though if you have better luck...since this disparity has been around for almost 2 years now...

If you just want to give Linux a try, you'd be better off just booting from a live Linux CD, and installing BOINC to its virtual drive. That way, you won't have the overhead of a VM.

(Sorry Brian, that I didn't think to mention this before.)

Brian Silvers
Brian Silvers
Joined: 26 Aug 05
Posts: 772
Credit: 282,700
RAC: 0

RE: So, for your goal, I

Message 85899 in response to message 85896

Quote:

So, for your goal, I suggest running just one more result before switching to the patched executable, which will mean you have two remaining on which to test it.

I doubt if I'm going to have the time to tinker in the next 4 days, which is the amount of CPU time needed (approx), so I'll probably do that on another round. As I think I remember reading, the fact that we don't get as many of the same template at the same time makes it more difficult to do this kind of testing, particularly on my (now) slow system. Note: That was not a dig at the app, but just the fact that yes, I realize the technology has passed me by...

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.