GPU version tweaking for current WUs

Message boards : Number crunching : GPU version tweaking for current WUs

To post messages, you must log in.

1 · 2 · Next

AuthorMessage
Sergei Chernykh
Project administrator
Project developer

Send message
Joined: 5 Jan 17
Posts: 388
Credit: 72,451,573
RAC: 0
   
Message 463 - Posted: 8 Jun 2017, 10:29:45 UTC

GPU version becomes more and more demanding for CPU with current type of work units. This was expected, and it's already affecting owners of GTX 1070 (and faster) GPUs. The fix is to run 2 or more tasks per GPU. My tests have shown that the following app_config.xml
<app_config>
   <app>
      <name>amicable_2_64</name>
      <max_concurrent>3</max_concurrent>
      <report_results_immediately/>
      <fraction_done_exact/>
      <gpu_versions>
          <gpu_usage>0.5</gpu_usage>
          <cpu_usage>1</cpu_usage>
      </gpu_versions>
    </app>
</app_config>

is enough to get 100% load on GTX 1080 Ti now, but I'll have to switch to 4 tasks per GPU soon.
ID: 463 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
KaptainBlaZzed

Send message
Joined: 28 Mar 17
Posts: 1
Credit: 100,600,495
RAC: 0
   
Message 464 - Posted: 8 Jun 2017, 16:57:05 UTC - in response to Message 463.  

thx for this guidance.

I was having issues with my my 1070 and 1080's getting horrible run times while running a single WU and reserving more CPU threads did not help.

I am now running 3 WU on my 1070 and 4 WU's on my 1080's.
My cpus are slow Xeons so i need to run more WU's to keep usage up.
I will probably have to bump it up to 4 and 5 WU respectively to keep usage up.
ID: 464 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Matt Kowal
Avatar

Send message
Joined: 12 Feb 17
Posts: 14
Credit: 175,874,429
RAC: 0
   
Message 465 - Posted: 9 Jun 2017, 14:29:59 UTC

Thank you, Sergei.

I have updated the config for my AMD 7970 to run 2 WUs concurrently.
ID: 465 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Su_Root@jisaku

Send message
Joined: 16 May 17
Posts: 3
Credit: 10,078,653
RAC: 0
  
Message 469 - Posted: 13 Jun 2017, 12:00:22 UTC

Thanks
My GTX 1050Ti is around 75% load, it's fixed.
ID: 469 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Sergei Chernykh
Project administrator
Project developer

Send message
Joined: 5 Jan 17
Posts: 388
Credit: 72,451,573
RAC: 0
   
Message 473 - Posted: 16 Jun 2017, 6:40:14 UTC
Last modified: 16 Jun 2017, 6:40:42 UTC

GTX 1080 Ti requires 8 simultaneous tasks now to get 100% load, my Ryzen 7 1700 @ 3.6 GHz is at 70% load already. Damn, it will soon be not powerful enough to feed 1080 Ti.
ID: 473 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
RobertN

Send message
Joined: 30 Mar 17
Posts: 7
Credit: 153,642,561
RAC: 5
   
Message 474 - Posted: 16 Jun 2017, 12:37:07 UTC

This is with a GTX 1080:



GPU load is around 80%. :(
ID: 474 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Sergei Chernykh
Project administrator
Project developer

Send message
Joined: 5 Jan 17
Posts: 388
Credit: 72,451,573
RAC: 0
   
Message 475 - Posted: 16 Jun 2017, 12:57:57 UTC - in response to Message 474.  

This is with a GTX 1080:
...
GPU load is around 80%. :(

I've created a new application Amicable Numbers up to 2^64 (double check) for selective checking of the current search range. It'll have GPU-friendly work units, one WU will be enough for 100% load even on 1080 Ti. You can switch to it when it has work to do (hopefully this evening, I'm generating work now).
ID: 475 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
RobertN

Send message
Joined: 30 Mar 17
Posts: 7
Credit: 153,642,561
RAC: 5
   
Message 478 - Posted: 16 Jun 2017, 15:45:52 UTC

Works great! :)
ID: 478 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Profile Su_Root@jisaku

Send message
Joined: 16 May 17
Posts: 3
Credit: 10,078,653
RAC: 0
  
Message 479 - Posted: 16 Jun 2017, 16:50:42 UTC
Last modified: 16 Jun 2017, 16:51:17 UTC

Amicable Numbers up to 2^64
Run time : 1453.56
CPU time : 1453.56

https://sech.me/boinc/Amicable/result.php?resultid=4303593

↓↓↓

Amicable Numbers up to 2^64 (double check)
Run time : 559.51
CPU time : 8.97

https://sech.me/boinc/Amicable/result.php?resultid=4312215

Awesome update!!!
Thank you!
ID: 479 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
tictoc

Send message
Joined: 4 Feb 17
Posts: 1
Credit: 34,066,518
RAC: 75
  
Message 498 - Posted: 26 Jun 2017, 13:31:38 UTC

I'm now up to 12 tasks on my 1070. That puts my 1070 at about 85-90% GPU load. I tried running 14 tasks, but at that setting the tasks errored out within a few seconds.

Running this many concurrent GPU tasks, on a single GPU, is uncharted territory for a GPU project, so I am not too surprised that 14x tasks instantly failed.
ID: 498 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Tex1954

Send message
Joined: 4 Feb 17
Posts: 4
Credit: 22,825,667
RAC: 0
   
Message 516 - Posted: 29 Jun 2017, 19:04:33 UTC - in response to Message 498.  
Last modified: 29 Jun 2017, 19:10:42 UTC

I'm up to 4 tasks on a GTX-980 and still only at 65%-75%... BUT, I can't run anything else because it dramatically decreases the GPU utilization if I do. These new tasks are VERY CPU intensive...

I kinda thought the whole point of using a GPU was to offload the CPU and increase speed... maybe we need more than 21 Kernels?

8-)
ID: 516 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Sergei Chernykh
Project administrator
Project developer

Send message
Joined: 5 Jan 17
Posts: 388
Credit: 72,451,573
RAC: 0
   
Message 517 - Posted: 29 Jun 2017, 19:55:04 UTC - in response to Message 516.  

I'm up to 4 tasks on a GTX-980 and still only at 65%-75%... BUT, I can't run anything else because it dramatically decreases the GPU utilization if I do. These new tasks are VERY CPU intensive...

I kinda thought the whole point of using a GPU was to offload the CPU and increase speed... maybe we need more than 21 Kernels?

8-)

The part that runs on CPU (generating numbers to check on GPU) gets harder in the end of the search. Yes, it can also be done on GPU, but it's hard to implement without mistakes. Wait a few more days until the next search is started, it will be GPU-friendly as before.
ID: 517 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
cibban

Send message
Joined: 28 Jun 17
Posts: 4
Credit: 61,849
RAC: 0
Message 524 - Posted: 30 Jun 2017, 20:23:29 UTC

I can't get BOINC to run more than one WU whatever I do, with GPU at ~30%. I tried some different settings in app_config.xml but nothing changes. I have an old GeForce GTX 560 with 1024 MB RAM. Could that be the reason?
ID: 524 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Sergei Chernykh
Project administrator
Project developer

Send message
Joined: 5 Jan 17
Posts: 388
Credit: 72,451,573
RAC: 0
   
Message 525 - Posted: 30 Jun 2017, 20:55:08 UTC - in response to Message 524.  

Each WU requires RAM, so you can run only one WU with 1024 MB RAM.
ID: 525 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
cibban

Send message
Joined: 28 Jun 17
Posts: 4
Credit: 61,849
RAC: 0
Message 526 - Posted: 30 Jun 2017, 22:09:37 UTC
Last modified: 30 Jun 2017, 22:11:22 UTC

Thank you! I suspected that. Now I can go on with my day ans stop tweaking the untweakable. :)

Will the new 10^20 WU:s utilize my GPU better?
ID: 526 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
cibban

Send message
Joined: 28 Jun 17
Posts: 4
Credit: 61,849
RAC: 0
Message 533 - Posted: 1 Jul 2017, 10:38:07 UTC

Since this morning, I get no WU's at all. I'm guessing my GPU RAM is not sufficient for the 10^20 tasks. Could that be it?
ID: 533 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
Sergei Chernykh
Project administrator
Project developer

Send message
Joined: 5 Jan 17
Posts: 388
Credit: 72,451,573
RAC: 0
   
Message 534 - Posted: 1 Jul 2017, 10:43:45 UTC - in response to Message 533.  
Last modified: 1 Jul 2017, 10:45:15 UTC

Since this morning, I get no WU's at all. I'm guessing my GPU RAM is not sufficient for the 10^20 tasks. Could that be it?

Sorry, but they require at least 1600 MB RAM each (depending on your kernel size setting). I did my best to fit them into 1024 MB, but it was impossible: prime numbers table alone is ~1208 MB in size (all primes up to 14142135624 => 633521375 primes, 2 bytes per each prime number).
ID: 534 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
cibban

Send message
Joined: 28 Jun 17
Posts: 4
Credit: 61,849
RAC: 0
Message 551 - Posted: 3 Jul 2017, 12:27:48 UTC

Thanks. I saw that in the eventlog too. I'm just going to have to get a graphics card from this decade. ;)
ID: 551 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
klepel

Send message
Joined: 31 Oct 18
Posts: 3
Credit: 38,214,322
RAC: 43,869
   
Message 965 - Posted: 1 Nov 2018, 18:10:39 UTC - in response to Message 463.  

The fix is to run 2 or more tasks per GPU. My tests have shown that the following app_config.xml
<app_config>
   <app>
      <name>amicable_2_64</name>
      <max_concurrent>3</max_concurrent>
      <report_results_immediately/>
      <fraction_done_exact/>
      <gpu_versions>
          <gpu_usage>0.5</gpu_usage>
          <cpu_usage>1</cpu_usage>
      </gpu_versions>
    </app>
</app_config>

is enough to get 100% load on GTX 1080 Ti now, but I'll have to switch to 4 tasks per GPU soon.

I do have two questions:
1) My GTX 1070 does not recieve WUs on my Ryzen 1700x with Lubuntu 18.10. What is wrong?
2) On my other system with 2 GTX970 under Windows 7, the above App_Config does not work. Does not run two Wus at the same time: I changed <name>amicable_2_64</name> to <name>amicable_10_20</name>.
I should say that I run App_Config files on other projects with no problems.

Has anybody any suggestions?

Thanks
ID: 965 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
klepel

Send message
Joined: 31 Oct 18
Posts: 3
Credit: 38,214,322
RAC: 43,869
   
Message 966 - Posted: 2 Nov 2018, 14:54:15 UTC - in response to Message 965.  

The fix is to run 2 or more tasks per GPU. My tests have shown that the following app_config.xml
<app_config>
   <app>
      <name>amicable_2_64</name>
      <max_concurrent>3</max_concurrent>
      <report_results_immediately/>
      <fraction_done_exact/>
      <gpu_versions>
          <gpu_usage>0.5</gpu_usage>
          <cpu_usage>1</cpu_usage>
      </gpu_versions>
    </app>
</app_config>

is enough to get 100% load on GTX 1080 Ti now, but I'll have to switch to 4 tasks per GPU soon.

I do have two questions:
1) My GTX 1070 does not recieve WUs on my Ryzen 1700x with Lubuntu 18.10. What is wrong?
2) On my other system with 2 GTX970 under Windows 7, the above App_Config does not work. Does not run two Wus at the same time: I changed <name>amicable_2_64</name> to <name>amicable_10_20</name>.
I should say that I run App_Config files on other projects with no problems.

Has anybody any suggestions?

Thanks

1) My GTX 1070 does not recieve WUs on my Ryzen 1700x with Lubuntu 18.04: STILL DOES NOT DOWNLOAD WUs.
2) Second try. The app_config works on the GTX970.
ID: 966 · Rating: 0 · rate: Rate + / Rate - Report as offensive    Reply Quote
1 · 2 · Next

Message boards : Number crunching : GPU version tweaking for current WUs


©2019 Sergei Chernykh