New GPU client and enhancement of project focus

Message boards : News : New GPU client and enhancement of project focus

To post messages, you must log in.

1 · 2 · 3 · 4 . . . 12 · Next

AuthorMessage
Thomas Koch
Project administrator
Project developer
Project scientist

Send message
Joined: 17 Feb 12
Posts: 436
Credit: 37,847
RAC: 0
Message 9874 - Posted: 19 Nov 2014, 13:41:39 UTC

Dear POEM@HOME members,

We are glad to announce that we now have a GPU version of our latest force field ready for deployment on our BOINC test server. This new client has been made possible by Vladimir Tankovich, a community member who approached us to help improve the efficiency of our OpenCL client. We gratefully accepted that help and set him the task to port our latest force field to OpenCL code. And he has done a great job. The performance of the new OpenCL client on our small GPU cluster is impressive, to say the least. For numbers, the AMD Radeon R290X in that cluster is 13 times faster than a full AMD Opteron 6380 CPU.

This latest force field provides a much more accurate description of the atomic interactions within proteins and also between the protein and the solvent surrounding it. In addition, new algorithms have made it also much more computationally efficient. Last but not least, we have some new Monte Carlo algorithms that efficiently run on hundreds of threads ready for deployment, like Multiple-Try-Metropolis Monte Carlo.

With these new tools at hand, we are going to shift the focus of this project a little from the question “Into which structures do proteins fold?” to the question “How do proteins fold into these structures?” The scientific reason is that proteins must not only fold into their native structure to fulfill their function, but they must do so in a reasonable amount of time. In addition, the native structure must be stable enough, but also allow for enough flexibility to allow the protein to fulfill its function. These are some of the aspects we are going to investigate with our new tools.

Finally, here are some technical details. The new force field is optimized for AMD’s GCN architecture, but runs also fairly well on NVIDIA Maxwell hardware. Older OpenCL 1.1 hardware from NVIDIA and AMD is also supported, but less efficient. Although owners of those older cards are also welcome to join the test, they may consider using them for other BOINC projects, which can use them more efficiently. Finally, other GPU vendors, e.g. Intel, are not supported.

Since we have only a very limited amount of different hardware available for testing, we are deploying this new GPU client on our test servers first. We hope to get as many results from different cards as possible to check if there are any problems left that slipped through internal testing. So please help us by running one of the jobs on the test server.

To do so, use the “Add project” wizard of your BOINC client and paste the Project URL “http://int-boinctest.int.kit.edu/poem/” into the corresponding text field, as our test server is not listed as an official BOINC project. We appreciate every help for testing, but please keep the following in mind:
- Test server results will not be used for scientific publications.
- Credits will not be listed in the BOINC cross-project statistics.
- Test server applications may crash your system, so be sure to save any valuable data beforehand.
Best practice is to return a few tasks, and immediately detach from the POEM@TEST project.

Kind regards,
your POEM@HOME Team
ID: 9874 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Yury Naydenov

Send message
Joined: 9 May 13
Posts: 7
Credit: 63,969,653
RAC: 0
Message 9875 - Posted: 19 Nov 2014, 15:22:51 UTC
Last modified: 19 Nov 2014, 16:14:35 UTC

Greetings!
Safely handled 5 of 5 jobs.
http://int-boinctest.int.kit.edu/poem/results.php?userid=245
Win 7 64 sp1, driver Ctalyst 14.9.1beta, video card AMD Radeon HD 6970, gpu load 98-99% 1cpu+1gpu.
6970+6870 (0.5CPUs+1gpu) load 98% each card.
ID: 9875 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Sabroe_SMC

Send message
Joined: 9 May 08
Posts: 15
Credit: 634,986,666
RAC: 0
Message 9876 - Posted: 19 Nov 2014, 15:50:22 UTC

http://int-boinctest.int.kit.edu/poem/results.php?hostid=318
6 WU's run very well
2x R9 280X@1080 Mhz , Win 7 Home Prem. 64bit
Catalyst 14.9 APP-SDK 2.9.1
workload both gpu's approx 93% each
ID: 9876 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Phil1966

Send message
Joined: 16 Jul 13
Posts: 27
Credit: 58,175,697
RAC: 0
Message 9877 - Posted: 19 Nov 2014, 16:43:40 UTC
Last modified: 19 Nov 2014, 17:12:13 UTC

Hello !

Nice to read some news ! :D

Currently running your test WU's on a GTX 970 Gigabyte Gaming G1

It is unfortunately still not possible to run POEM on 2 nVidia.

Appart of that remaining issue, the WU's look OK :)

NB : Have added an app_config to allocate 1 CPU Core / WU

Will do as proposed : Run 5 or 10 and then stop.

NB GPU Load > 90 % ! Great !

Don't know if the apps are similar to the one in "production" / Alpha.

If yes, the speed increase / optimization is fantastic.

Hope to see them on Poem@Home soon.

Best Regards

Phil1966
ID: 9877 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Vladimir Tankovich
Volunteer developer
Volunteer tester

Send message
Joined: 12 Nov 10
Posts: 182
Credit: 429,219,133
RAC: 0
Message 9878 - Posted: 19 Nov 2014, 16:45:05 UTC - in response to Message 9876.  

260x, 750 fine.
probably should enable 2 WUs per GPU to push use to 100%. test server only gives 1 per machine per GPU at this time.
ID: 9878 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Martin_POEM
Volunteer moderator
Project administrator
Project scientist

Send message
Joined: 18 Jan 11
Posts: 41
Credit: 26,196
RAC: 0
Message 9879 - Posted: 19 Nov 2014, 18:40:04 UTC

Hi all, POEM++ dev here.

Could you please all pay attention if you get any display driver resets when running jobs under windows. That would suggest there is still a problem with the code.

Likely the OpenCL app will hang after such a display driver reset. If so, you can try pausing and resuming the app to get the WU finished.

Apps are not directly comparable to what is running on the live server. But still we are very happy with current performance.

We will try enabling 2 gpu per WU later. But we have never tested that internally yet. And we dont want apps crashing because of memory allocation issues etc.
ID: 9879 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Phil1966

Send message
Joined: 16 Jul 13
Posts: 27
Credit: 58,175,697
RAC: 0
Message 9880 - Posted: 20 Nov 2014, 11:18:49 UTC - in response to Message 9879.  

Hi Martin,

thank you for your message and all your efforts !

Concerning :

We will try enabling 2 gpu per WU later. But we have never tested that internally yet.

I don't suggest to run a single WU on 2 different GPU, but 1 WU per GPU, on more than 1 GPU at the same time, even if you have 2, 3 or 4 GPU's. (as for GPUGRID, PrimeGrid, ...)

Currently, as it used to be the case in the passed, running POEM on more than 1 GPU is not possible.

Thank You

Best Regards

Phil1966


ID: 9880 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile [PUGLIA] Riccardo

Send message
Joined: 19 Jun 09
Posts: 4
Credit: 13,662,315
RAC: 0
Message 9881 - Posted: 20 Nov 2014, 13:03:57 UTC

100% fail on 38 WUs :-(

http://boinc.fzk.de/poem/forum_thread.php?id=1097

Win7 on I7-2600K and GTX-570
ID: 9881 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Vladimir Tankovich
Volunteer developer
Volunteer tester

Send message
Joined: 12 Nov 10
Posts: 182
Credit: 429,219,133
RAC: 0
Message 9882 - Posted: 20 Nov 2014, 22:32:51 UTC

yes, 2 NVidia GPUs on 1 machine fail. here is an example:

http://int-boinctest.int.kit.edu/poem/results.php?hostid=376

Also I was talking about letting test server issue more than 1 WU per GPU, so that client that has 1 video card can run them at the same time to avoid blocking GPU work on CPU to achieve 100% GPU use. This will work better for AMD since their opencl implementation does not use as much CPU (unlike NVIDIA's that constantly polls GPU, effectively blocking 1 thread, though achieving less latency between kernel calls).
ID: 9882 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Jim1348

Send message
Joined: 4 Jul 12
Posts: 104
Credit: 195,960,357
RAC: 0
Message 9883 - Posted: 21 Nov 2014, 0:44:33 UTC
Last modified: 21 Nov 2014, 0:47:23 UTC

I have finished 7 on two HD 7790s with no problem (AMD 14.4 drivers, Win7 64-bit), and will stop to give others a chance. But they seem to run fine, a little over 4 minutes per work unit. And GPU-Z shows the average GPU% at about 60%, but it jumps around a lot between zero and maximum. I guess that is the nature of OpenCL.
ID: 9883 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Vladimir Tankovich
Volunteer developer
Volunteer tester

Send message
Joined: 12 Nov 10
Posts: 182
Credit: 429,219,133
RAC: 0
Message 9884 - Posted: 21 Nov 2014, 4:19:48 UTC - in response to Message 9883.  

Jim, what is the clock speed on your cards? is the machine busy with something else? I have a very similar chip - 260X at 1175 and one WU averages it above 95%.
I see that you have runs of 200-250 seconds. mine are around 150-160.
ID: 9884 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Jim1348

Send message
Joined: 4 Jul 12
Posts: 104
Credit: 195,960,357
RAC: 0
Message 9885 - Posted: 21 Nov 2014, 6:07:06 UTC - in response to Message 9884.  
Last modified: 21 Nov 2014, 6:33:52 UTC

Trader,

The GPU clock is 1000 MHz, and the memory clock is 1500 MHz (both default for the card) according to GPU-Z. But I do see that the GPU clock is changing quite a bit, which I have not noticed before on other projects. I have six cores of my i7-4771 in use for CPU projects (typically 4 or 5 on Atlas, one or two on CPN or Cosmology). The leaves two cores that I have reserved for the HD 7790s.

I think I will free up a couple more CPU cores and see if it makes any difference. Thanks for the input.

EDIT:Running only 4 cores on CPU projects (all Atlas at the moment) resulted in a steady GPU clock and a constant 100% GPU utilization over most of the run. The average time is now 197 seconds (3 minutes 17 seconds), which is much nicer. I think that will do for now. Thanks again.
ID: 9885 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile [VENETO] boboviz

Send message
Joined: 30 Dec 09
Posts: 236
Credit: 48,564,069
RAC: 0
Message 9886 - Posted: 21 Nov 2014, 6:39:19 UTC

AMD FX 6300 + HD6850, all ok. Wus between 15/17 minutes
Video lag as the "old" gpu app....too much little the gpu
ID: 9886 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
keputnam

Send message
Joined: 25 Jun 14
Posts: 10
Credit: 39,555,105
RAC: 0
Message 9887 - Posted: 21 Nov 2014, 9:46:17 UTC
Last modified: 21 Nov 2014, 9:54:05 UTC

Win7 Pro/64Bit 8GB RAM

Intel Core i7 4771/3.4, Radeon 7870/2GB - Pitcairn

Seven WU successfully completed

CPU time between 7.16 & 5.18 seconds

Run Time between 158 & 185 seconds
ID: 9887 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
JugNut

Send message
Joined: 1 Sep 11
Posts: 15
Credit: 728,763,887
RAC: 0
Message 9888 - Posted: 21 Nov 2014, 19:53:44 UTC
Last modified: 21 Nov 2014, 20:50:31 UTC

Great news on the new GPU app.

Results of new app..

AMD HD 7970 on Win 7 x64 = works great. Time taken 1.5 - 5 mins

2x Nvidia GTX 970's = Computation error -1073741511 (0xffffffffc0000139) Unknown error number. Using Driver ver 344.48 on Win 7 x64. This is with or without excluding 1 x GPU in cc_config.

Cheers.
ID: 9888 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Bossmaniac
Avatar

Send message
Joined: 25 Jun 12
Posts: 7
Credit: 687,065,612
RAC: 0
Message 9889 - Posted: 21 Nov 2014, 20:09:24 UTC
Last modified: 21 Nov 2014, 20:18:59 UTC

Intel 4770k Nividia Titan driver: 337.88
4 wu, works ok, 99% load, no fail
Os Win7 pro

Intel Xeon(R) CPU E5-2670
ATI 7970 Catalyst 14.4
5 wu works ok,80-98% load
Os Win7 pro

Intel Xeon e5 2690v2
3x NVIDIA GeForce GTX 680 (2048MB) driver: 337.88
Instand fail of all Wu, with and wiothout exclusion with cc_config
Os Win7 pro

Also Instand Fail on NVIDIA GeForce GTX 550 Ti (1024MB) driver: 337.88
Os Win7 Pro
ID: 9889 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
JoeM

Send message
Joined: 15 Apr 13
Posts: 7
Credit: 78,429,992
RAC: 0
Message 9890 - Posted: 22 Nov 2014, 0:55:57 UTC

Two AMD 280X GPUs
Win 7 x64
Work units took around 5 minutes, one at 90 seconds.
4 POGS cpu apps running concurrently.
ID: 9890 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile BlueGhost

Send message
Joined: 19 Aug 11
Posts: 11
Credit: 56,429,329
RAC: 0
Message 9891 - Posted: 22 Nov 2014, 1:11:08 UTC

Hello Thomas.

Great news. Thanks to you, the POEM@HOME Team, and Vladimir Tankovich. Here are the data about my tests for the new GPU Version.

Computer 1 config:

AMD FX-6300 6x 3,5 GHz / 16GB DDR3 1866
Windows7 Ultimate 64Bit
BOINC Version 7.2.42 (64Bit)

AMD Radeon 7850 / 2GB GDDR5 RAM
(not overclocked, producer default settings, GPU clock rate: 1000MHz / RAM frequenz: 1225MHz)

Driver-Version: 14.301.1001-140915a-176154C
Catalyst-Version: 14.9
Direct3D-Version: 9.14.10.01061
OpenGL-Version: 6.14.10.13084
Mantle-Version: 9.1.10.0034
Mantle API Version: 94209

No app_info.xml

0.733 CPU`s + 1 ATI GPU

CPU utilization: constant 1% - difference to OpenCL version 2.01 -> 9%
GPU utilization: constant 97% - "wide" difference to OpenCL version 2.01
-> GPU utilization 46%
RAM utilization nearly equal by 64MB on both OpenCl versions

Completed 12 WU`s without any problems. Data from the WU`s:

Task ID / Work package / Runtime / CPU Time

157230 / 154558 / 529.04 / 27.03
157206 / 154535 / 537.28 / 26.77
157223 / 154551 / 172.60 / 13.60
157213 / 154542 / 538.67 / 27.46
157149 / 154485 / 573.02 / 25.65
156933 / 154304 / 534.49 / 27.42
157109 / 154284 / 532.80 / 28.02
157146 / 154386 / 539.30 / 27.38
157210 / 154539 / 542.75 / 27.03
157090 / 154449 / 552.96 / 27.66
157155 / 154491 / 539.91 / 27.81
157141 / 154483 / 168.75 / 13.90

-----------------------------------------------------------------------------

Computer 2 config:

AMD Athlon II X4 610e / 8GB DDR3 1333 RAM
Windows7 64Bit
BOINC Version 7.2.42 (64Bit)

AMD Radeon 7750 / 1GB GDDR5 RAM
(not overclocked, producer default settings, GPU clock rate: 800MHz / RAM frequenz: 1125MHz)

Driver-Version: 9.012-121219a-152192C-ATI
Catalyst-Version: 13.1
Direct3D-Version: 9.14.10.0945
OpenGL-Version: 6.14.10.12002

No app_info.xml

0.593 CPU`s + 1 ATI GPU

CPU utilization: constant 1% - difference to OpenCL version 2.01 -> 10%
GPU utilzation: constant 98% - "wide" difference to OpenCL version 2.01
-> GPU utilzation 49%
RAM utilization nearly equal by 60MB on both OpenCl versions

Completed 4 WU`s without any problems. Data from the WU`s:

Task ID / Work package / Runtime / CPU Time
157766 / 155019 / 1,468.02 / 31.37
157567 / 154685 / 1,466.69 / 31.20
157745 / 154998 / 465.06 / 19.25
157669 / 154929 / 1,607.48 / 32.32


Kind regards

BlueGhost
ID: 9891 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Martin_POEM
Volunteer moderator
Project administrator
Project scientist

Send message
Joined: 18 Jan 11
Posts: 41
Credit: 26,196
RAC: 0
Message 9892 - Posted: 22 Nov 2014, 21:11:22 UTC

Hi all,

could those who expierence problems with the gpu jobs (especially NVIDIA Fermi cards or multiple NVIDIA cards) please post their computer id for the test server here. That will make it easier for us.

We plan to release a new test client on tuesday, which will hopefully give us some more info on what is going wrong.
ID: 9892 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Jim1348

Send message
Joined: 4 Jul 12
Posts: 104
Credit: 195,960,357
RAC: 0
Message 9893 - Posted: 22 Nov 2014, 22:41:06 UTC - in response to Message 9892.  

I had problems on two Maxwell cards (GTX 750 Tis) on an Ivy Bridge Z77 motherboard. Only 3 out of 18 completed successfully (Win7 64-bit).

Computer ID: 395
ID: 9893 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
1 · 2 · 3 · 4 . . . 12 · Next

Message boards : News : New GPU client and enhancement of project focus


Copyright © 2017 KIT-INT