Times (Elapsed / CPU) for BRP5/6/6-Beta on various CPU/GPU combos - DISCUSSION Thread

Maximilian Mieth
Maximilian Mieth
Joined: 4 Oct 12
Posts: 128
Credit: 8,642,951
RAC: 2,380

RE: RE: ... If you're not

Quote:
Quote:
... If you're not running any CPu tasks along with them, the CPU will be at idle / base frequency (1600 MHz I think) when the Einstein tasks start. Ramping it up to full speed takes some time. If Einstein is already finished, or at least most of it, the average CPU clock speed will be well below the maximum clock speed.

Thanks very much for pointing this out! I've sometimes seen people say that they run multiple GPU tasks and leave ALL the cores free. I'm sure that helps with both power consumption and temperature but may hinder GPU performance if all the CPU cores are likely to be running at idle frequency most of the time. I remember looking at such a host some time ago and expecting to find the fastest crunch times but actually seeing what seemed to be slightly worse performance. That all makes sense now. Thanks for the explanation.

I just tried it out and it works for me too! I got a performance boost of about 10% for my NVIDIA 610M. But the INTEL HD4000 does not seem to be faster. So I guess it only applies to "real" GPUs. I will have to monitor the temperature however. I hope it is not getting too hot.

Jeroen
Jeroen
Joined: 25 Nov 05
Posts: 379
Credit: 738,881,678
RAC: 1,204

I fired up an older GTX 580

I fired up an older GTX 580 to see how it would do with the new version of the BRP6 application. The card is performing quite well with this application.

[pre]
CPU: Intel Core i7 5960x @ 4.1 GHz
Threads: 8 - HT bypassed via process lasso
PCIe slot x16 Version 2.0
1st GPU: EVGA GTX 580 SC
RAM: 4 x 4GB DDR4 2600 MHz CL13
Concurrency: 0.5 @ 0.2 CPUs + 0.5 GPUs
CPU Tasks: 7x Asteroids@Home AVX CPU tasks
Free CPU cores: 1
OS: Windows 10 x64 Build 10049

Application Elapsed CPU time Sample Size
(Parkes PMPS XT) v1.52 6385 398 6

Daily production per elapsed time and concurrency: ~119K
[/pre]

I only have a few samples so I thought I would post here instead.

Jeroen

JBird
Joined: 22 Dec 14
Posts: 1,963
Credit: 4,046,216,051
RAC: 0

Do we know if a lower CPU

Do we know if a lower CPU fraction will increase the GPU load?
GTX 960 Maxwell stays at about 82% at the .2CPU fraction. Mem use is 301MB

Yet on SETI CUDA apps the CPU fraction I use is 0.04(with GPU at 0.5) and get GPU usage in the mid to high 90% range (637MB memory)

I don't see a cuda.cfg file to edit process priority

Both these *things improve runtime

Sid
Sid
Joined: 17 Oct 10
Posts: 149
Credit: 594,036,928
RAC: 611,751

RE: I fired up an older GTX

Quote:

I fired up an older GTX 580 to see how it would do with the new version of the BRP6 application. The card is performing quite well with this application.
Daily production per elapsed time and concurrency: ~119K

Jeroen

Such kind of cards are reasonably cheap on a second market.
However, I wonder how many heat it produces comparing with relatively new cards?

Jeroen
Jeroen
Joined: 25 Nov 05
Posts: 379
Credit: 738,881,678
RAC: 1,204

RE: Such kind of cards are

Quote:

Such kind of cards are reasonably cheap on a second market.
However, I wonder how many heat it produces comparing with relatively new cards?

The GTX 580 has a TDP of 244w and factory temperature target of 86C. However, with a custom fan profile, this card can be kept in the 70s under load while running the BRP6 application. Also, I found that I am able to run a lower than factory GPU voltage with my cards without sacrificing stability. This can help save a bit on power draw.

ExtraTerrestrial Apes
ExtraTerrestria...
Joined: 10 Nov 04
Posts: 770
Credit: 330,141,191
RAC: 645,623

The heat is usually not a

The heat is usually not a problem. On the other hand the power consumption - which causes the heat - can be a problem because you have to pay for it. If you're paying anything but US electricity prices that's going to hurt with a Fermi GPU, especially a high end 400 series. Assuming the GTX470 draws 200 W under BOINC load this would cost about 400€/year in Germany. At this point it's easy to see how more efficient cards in the range of 100 - 300€ would be a better deal for 24/7 number crunching, even though they cost more initially.

@JBird: there is no cuda.cfg at Einstein@Home. However, you can increase the process priority via Process Lasso. The only downside is you have to adjust this when ever the executable is renamed. Or.. I vaguely remember one can also define rules based on wildcards like *einstein*. I haven't tried to use anything like that, though.

Quote:
Do we know if a lower CPU fraction will increase the GPU load?


Are you referring to the "" tag in the app_config.xml? If so: lower values make BOINC reserve less CPU ressources to run the app. If this has any effect on performance (depending onnthe magnitude of the change) it's going to be detrimental. However, you can use this to make sure there are always as many CPU cores free as you want. A solid choice is to set the same value as for "" if you're using 1 GPU. This way one logical core is reserved.

MrS

Scanning for our furry friends since Jan 2002

archae86
archae86
Joined: 6 Dec 05
Posts: 3,021
Credit: 4,990,439,385
RAC: 3,038,717

RE: However, you can

Quote:
However, you can increase the process priority via Process Lasso. The only downside is you have to adjust this when ever the executable is renamed. Or.. I vaguely remember one can also define rules based on wildcards like *einstein*. I haven't tried to use anything like that, though.


You can and I do, though I'd advise a bit more specificity than the example.

I just looked, and my daily use machine currently has a Process Lasso CPU priority option configured to set the priority of tasks matching einstein*cuda32*.exe to "above normal".

Come to think of it, the "cuda32" part of that match may become obsolete in the very near future, if, as hoped, an executable compiled against a higher version of CUDA emerges.

Thanks for making me look at this. At first glance I think I'll revise my match string to "einstein*cuda*.exe".

ExtraTerrestrial Apes
ExtraTerrestria...
Joined: 10 Nov 04
Posts: 770
Credit: 330,141,191
RAC: 645,623

That's good advice. And while

That's good advice. And while we're at it: I'm also using Process Lasso to tie 4 CPU tasks to the first 2 physical cores with HT of my i7, to leave 2 physical cores free to optimally support the GPUs. This may be extreme, but the GPUs perform better than using 3 unregulated CPU tasks.

MrS

Scanning for our furry friends since Jan 2002

Stef
Stef
Joined: 8 Mar 05
Posts: 206
Credit: 110,568,193
RAC: 0

Heat isn't necessarily a

Heat isn't necessarily a problem. I'm planning to heat the living room with e@h next winter. 300W would be decent, I hope the R9 390/X will deliver a good performance. But I will have to see and compare what's on the market in October. And then sell it once it gets warmer outside.
That's the plan :)
The biggest problem is noise. I like it quiet. And dissipating 200-300W quietly isn't that easy without water cooling, which is too expensive.

The 750Ti provided a little too less heat past winter. The whole rig drew only about 110W including CPU tasks.

ExtraTerrestrial Apes
ExtraTerrestria...
Joined: 10 Nov 04
Posts: 770
Credit: 330,141,191
RAC: 645,623

Yes, cooling 300 W quietly is

Yes, cooling 300 W quietly is difficult. 200 W is relatively easy for a modern mid or full tower, though. You're going to hear it, but set up well it's only going to be a faint wind sound.

I wouldn't bet too much on that heating plan, tough. 300 W is almost nothing for a bigger room, depending on the insulation to the outside. And after half a year the big GPU will still be very competitive, yet you have to sell it at a significant loss since it's used.

I'd probably choose a mid range GPU, which is easy enough to cool, and run it until something appreciably more efficient comes out.

MrS

Scanning for our furry friends since Jan 2002

Comment viewing options

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