GPU Upgrade Shows No Improvement in Work Unit Completion

Sotho
Sotho
Joined: 6 Jul 13
Posts: 1
Credit: 5389830
RAC: 1968

PCIe bandwich is important.

PCIe bandwich is important. There is significant difference between PCIe 4X ,8X and 16X (PCIe 2.0)

cliff
cliff
Joined: 15 Feb 12
Posts: 176
Credit: 283452444
RAC: 0

Hi Phil, RE: Thank

Hi Phil,

Quote:

Thank you Archae. The 970 elapsed times running 2 WUs is 225 minutes. Like yours my single WUs takes 95 minutes to complete. It is actually taking longer to complete WUs running 2 cores.

After dinner I'm going to research the thread you directed me to. I really appreciate the advice because I'm back crunching after a long Illness and a 3 year hiatus.


You 'may' also need to check the Nvidia driver version in use. Some of the latest drivers are slower than earlier ones:-(
I write 'may' as I am running AMD not Intel, so the platform might be significant.

My main rig CPU is an AMD FX9590 @5022GHz [4core + 4 Logical]and I run 2 x gtx980ti gpu's and I dedicate each gpu to a specific project [E@H & MW@H] and only run 1 WU at a time. I'm running NV 358.91 drivers as I've found later ones crunch slower.

Also I don't run CPU work right now, since my rig is a tad overclocked:-)

Its also seemingly OS dependant.. I run Win 7 Ultimate and allow 1 core free for
feeing the GPU's. I'm crunching the Parkes beta tasks and they are faster then the 32 series. You need to check the use beta programs in the E@H setup on the server to get them.

I was running Win10 but I had a boot problem and had to remove the drive it was on, when I put Win10 on another SSD it ran tasks slower than Win7 so I'm back with that OS now, timings as stated.

I also set my cache at 0.5 & 0.5 since I only run about 5/6 hours per day on cheap rate electrickery:-) I can the reset the cache to a higher figure as/when I need to.

So far my times are roughly 44m40s per WU for E@H tasks. They seem to vary dependant on the individual task.. dispite that the tasks are 'supposed' to be the same size:-)The contents will vary of course..

I also lost a few WU during a recent hospital stay [6 days] and being ill isn't something you can plan for, but a small cache limits the damage:-)

You may also want to set DCF debug in the Boinc prefs to help sort out DCF.

Regards,

Cliff,

Been there, Done that, Still no damm T Shirt.

Florida Rancher
Florida Rancher
Joined: 4 Oct 13
Posts: 31
Credit: 23998436
RAC: 0

I'm very much a novice with a

I'm very much a novice with a great willingness to learn. I love acquiring knowledge and learning the technical aspects behind what makes Boinc tick intrigues me. I want to achieve maximum performance where possible. My knowledge is disjointed because I'm trying to learn by reading message board posts. This is the first time I've felt inclined to post my own questions. Thank you for bearing with me.

Here are the answers to your following questions:

1. what multiplicity are you running the jobs at (two at once is common, but values from 1 to 6 are not rare). 2

2. how many BOINC cpu tasks are running? After I made Juergen's changes it went from 7 to 8

3. which applications are running? (for comparison purposes it is good that a single GPU application type and a single CPU application type, from just one project is all that is running). Since no two applications behave the same, it is even more important that the workload running is the one you wish to run efficienty. Only Einstein@Home

4. does the machine run 24x7 BOINC, with not enough other work to matter? Yes, 24/7

After making the changes suggested by Juergen my stats took a great leap yesterday. My credit per day rose by almost 50,000 to 141,986. My average work done is still taking a hit from my time in the hospital but is now at 71,122.

I am running 8 CPU tasks and 4 GPU tasks between two cards.

archae86
archae86
Joined: 6 Dec 05
Posts: 3145
Credit: 7023204931
RAC: 1829022

RE: After making the

Quote:

After making the changes suggested by Juergen my stats took a great leap yesterday. My credit per day rose by almost 50,000 to 141,986. My average work done is still taking a hit from my time in the hospital but is now at 71,122.

I am running 8 CPU tasks and 4 GPU tasks between two cards.


The credit per day is severely distorted by variations in how much of your returned work is waiting for quorum partners.

For productivity assessment while making changes, I'd advise focusing on just two things:

1. invalid and error results -- if any -- there should be none.
2. average elapsed time per WU multiplied by number currently running simultaneously.

In your current situation, I believe the biggest "bang for the buck" constructive change you could make would be to switch from running the stock CUDA32 application for GPU GRP6 (Parkes)work to running the CUDA55 version.

On cards of your 970 generation, the CUDA55 variant consistently gives appreciably shorter execution times, and also requires less CPU support, so is likely less afflicted with variation arising from the busy-ness of your machine with other work.

To enable the CUDA55 version, you need to set your appropropriate venue/location to "allow test work".

To do this:
1. ascertain the location/venue in which your PC runs by clicking the the "details" link for that PC from the computer list reached from your Einstein account web page. (it will be one of default, home, work, or scholl)
2. from your main account page click on the Einstein preferences link
3. switch view so that you see current settings for all four locations
4. click edit at the bottom of the column for the location your PC occupies.
5. click the box for "run test applications" so a tick shows
6. click "update preferences" down at the bottom

The next time your PC fetches work, it will set the work to run using CUDA55 codes, which you'll be able to observe by application name in your task list (either online or on the PC).

That will help, I promise.
(this has already been suggested to you--I'm filling in details hoping to enable the change.)

(banished: ID 567202)
(banished: ID 5...
Joined: 20 Sep 11
Posts: 8
Credit: 28188130
RAC: 0

Anything below a 980 does not

Anything below a 980 does not add much or might be a step back these 960 970 cards suck
I found that out with the 960 I got

tbret
tbret
Joined: 12 Mar 05
Posts: 2115
Credit: 4812690330
RAC: 85748

RE: Anything below a 980

Quote:
Anything below a 980 does not add much or might be a step back these 960 970 cards suck
I found that out with the 960 I got

I think, if you'll do the comparison, that "suck" is a little harsh for a 970 compared to a 770 or a 670. It costs less, is faster, and generates less heat.

I don't have any experience with a 960, so if you say the 960s "suck" I have to take your word for it.

Florida Rancher
Florida Rancher
Joined: 4 Oct 13
Posts: 31
Credit: 23998436
RAC: 0

I just made the change you

I just made the change you suggested so I'll let you know what I observe in a couple of days.

The advice you are giving me I take to heart with enthusiasm. Being a cattle rancher doesn't make me the greatest of computer experts but I enjoy the hobby aspects of Boinc with a willingness to help a good cause while adding to my knowledge of understanding how it works.

Florida Rancher
Florida Rancher
Joined: 4 Oct 13
Posts: 31
Credit: 23998436
RAC: 0

Thank you Gary. Great stuff!

Thank you Gary. Great stuff! I'm rereading your post for the 3rd time and trying to digest your well-explained technical aspects of GPU crunching.

I've made any changes you've suggested so we'll see how things improve down the road. I'm not disappointed so much with the performance of my cards but that I fell for the trap of thinking that Nvidia card performance improves with an increase in model numbers.

I wrongly assumed a 970 would offer a marked improvement over a 3 year old 680. Also the 730 is a poor card but my last PCIe 2.0 x16 slot will only accept a single slot card.

I have five questions for you:

1. Under Einstein preferences what does "Run CPU versions of applications for which GPU versions are available" refer to?

2. How do I get into the CUDA beta testing program?

3. How is it that I have 8 CPU tasks and 4 GPU tasks running .2 CPUs when I only have 8 cores?

4. How to I insert a screenshot in a message?

5. Would you mind taking a look at my settings?

Einstein Preferences:
Use CPU checked
Use NVIDIA GPU checked
Run test applications checked
Run BRP checked
Run GWS checked
GPU utilization factor of BRP apps .5
GPU utilization factor of FGRP apps .5


hcc1
4

.500
1.250


1
1
1
0
0
0
0
1
0
0
1
1
1
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0


1
0
0
0
http://boinc.berkeley.edu/download.php?xml=1

http://boinc.berkeley.edu/download.php
0
0
0
0
0
0
0
0
0
0
0
default
0
300
10
2000
8
2
0
0
0
-1
http://www.google.com/
0
0
0
0
0
-1
-1


80

80




10.000000
0
0
30
0
0
0.000000
0
0
0
1
0
0
0

Regards,
Phil

Florida Rancher
Florida Rancher
Joined: 4 Oct 13
Posts: 31
Credit: 23998436
RAC: 0

I forgot to send include my

I forgot to send include my Boinc Preferences:

Use at most 100% of the CPUs
Use at most 75% of CPU time
Store at least 5 days of work
Switch between tasks every 60 minutes
Request tasks to checkpoint at most every 60 seconds
Use no more than 200 GB
Leave at least 0.001 GB free
Use no more than 50% of total
When computer is in use, use at most 90%
When computer is not in use, use at most 100%
Leave non-GPU tasks in memory while suspended
Page/swap file: use at most 75%

Gamboleer
Gamboleer
Joined: 5 Dec 10
Posts: 173
Credit: 168389195
RAC: 0

As someone mentioned, the 970

As someone mentioned, the 970 (all the Maxwell cards) are somewhat disappointing in performance relative to previous generation nVidia cards, but they are much more power efficient.

I have a 950 and 960. They both work best with 2 simultaneous BRP6 units. The 970 is probably the same, and could possibly handle 3. If I understood another post correctly, it's paired on a system with a 730. If it were my system, I would pull the 730 and keep the cores free for more CPU tasks. This will also make it easier to configure how many simultaneous GPU tasks are done, as I doubt the 970 and 730 would run the same number optimally and you would need a special config file for one card to run X number of tasks, and the other to run Y. When you only have one card (or two similar cards), you can simply adjust the GPU utilization factor on the web.

On an Intel machine that does GPU and CPU work and which I use while it is crunching, I leave one core free per GPU task plus an extra core free, allowing 100% CPU time. If I set the value to stop BOINC when non-BOINC usage gets above 60%, I find this allows me to browse the web and run Netflix, and BOINC only stops when I close many web tabs at once. When I decide to play a game that requires my GPU, I suspend BOINC or set it to stop GPU computing when the computer is in use.

To do this, just keep reducing the % of cores used in your Options > Computing Preferences menu on BOINC until you have your GPUs running as many units as you specified through your web preferences (2 simultaneous per GPU with your current 0.5 setting), plus CPU tasks, the total number of both task types together equaling one less than the number of cores in your system.

Comment viewing options

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