| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Active Research
Forum: OpenPandemics - COVID-19 Project
Thread: GPU works overloads CPU works |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 9
|
|
| Author |
|
|
HyperComputing
Advanced Cruncher Joined: Aug 10, 2019 Post Count: 74 Status: Offline Project Badges: 
|
My unit have CPU and GPU availables for works.
----------------------------------------This unit is working on OpenPandemics - COVID-19 and Microbiome Immunity Project. Usually there is 72 hours of CPU works ready to start (OPN and MIP). This is matching the unit's settings.  When there is GPU works : When a GPU works ends it download a new GPU work. (This is how it may work) But when there is no more GPU work to download, it get a CPU work. This is a problem because after the last GPU work is done there is 250 hours of CPU works ready to start. (OPN and MIP) This is not matching the unit's settings.  Of course when a CPU work will ends it will not download a new one until the amount fall back to 72 hours of CPU works ready to start. When there is no more GPU work for a project, i want GPU work for another project and not CPU works for the same project.(I know that other projects are not working with GPU)
My GPU compute WUs as fast as I'm crushing bubble wrap.
|
||
|
|
PMH_UK
Veteran Cruncher UK Joined: Apr 26, 2007 Post Count: 769 Status: Offline Project Badges: 
|
Suggest limiting the number of units downloaded and consider reducing days downloaded to nearer 2.
----------------------------------------BOINC may be misled by short GPU units. Paul.
Paul.
|
||
|
|
HyperComputing
Advanced Cruncher Joined: Aug 10, 2019 Post Count: 74 Status: Offline Project Badges:
|
This will not prevent it to download CPU works when there is no more GPU works to download.
----------------------------------------I just want that it not download CPU works when there is no more GPU works to download. CPU works process and download already work correctly.
My GPU compute WUs as fast as I'm crushing bubble wrap.
|
||
|
|
HyperComputing
Advanced Cruncher Joined: Aug 10, 2019 Post Count: 74 Status: Offline Project Badges:
|
OK, wrong diagnosis.
----------------------------------------It not get a CPU work when there is no more GPU work to download. The device is doing OPN1's work in 2h and MIP1's work in 1h. There's in cue around 40 works waiting with 2h and 1h of estimated time for each. (total of waiting works in cue around 72h) But sometime all estimated time for OPN1, and only OPN1, change to 9h each. (new total of waiting works in cue around 250h) Is it because of GPU's works ? I think yes.
My GPU compute WUs as fast as I'm crushing bubble wrap.
|
||
|
|
sam6861
Advanced Cruncher Joined: Mar 31, 2020 Post Count: 107 Status: Offline Project Badges:
|
CPU slow down from overheat (check with Windows HWinfo64 tools), or CPU slow down from integrated graphics using a slow single channel RAM, can throw time estimation off.
But this can just be highly variable runtime on some MIP1 and OPN1. I have some MIP1 complete as little as 0.4 hours, and some MIP1 run a long time 1.6 hours. This throws runtime estimation off. 1. Boinc grabs 1 day of work with time estimation 0.5 hours each. Lots of tasks gotten. 2. Boinc completes some tasks as 1.5 hours. 3. Boinc submits some results and get a few more tasks, changes time estimation of all not yet started MIP1 or OPN1 tasks closer to 1.5 hours. This is when it suddenly appear to be 2 or 3 days of work. |
||
|
|
bluestang
Senior Cruncher USA Joined: Oct 1, 2010 Post Count: 272 Status: Offline Project Badges: 
|
What you are looking to do is not really a possibility since the Boinc Manager will always download what is available according to what you have set in your WCG settings.
----------------------------------------If you only want to run GPU work then the only option is to set OpenPandemics as your only project and then under Device Profiles set one of them with the "Allow research to run on my CPU?" To No. If you want to run CPU and GPU on then set up different Device Profiles for what you what to accomplish and then run Multiple Boinc instances and set them to different Device Profiles...one for CPU only and one for GPU only. Just make sure on the CPU only you leave enough CPU threads to run you GPU tasks. Hope that makes sense. ---------------------------------------- [Edit 2 times, last edit by bluestang at Jun 6, 2021 7:54:46 PM] |
||
|
|
HyperComputing
Advanced Cruncher Joined: Aug 10, 2019 Post Count: 74 Status: Offline Project Badges:
|
I think the estimated time for CPU works should not take into account the actual time of GPU works.
----------------------------------------They are two different processes and one should not conflict with the other. GPU works don't take up CPU resources to the point of making such a big difference.
My GPU compute WUs as fast as I'm crushing bubble wrap.
|
||
|
|
bluestang
Senior Cruncher USA Joined: Oct 1, 2010 Post Count: 272 Status: Offline Project Badges:
|
Yeah, I agree but the Boinc scheduler isn't that smart to know the difference. Also, each GPU task needs a full CPU thread so that complicates it even further.
----------------------------------------Setting up 2 instances (1 running CPU only and 1 running GPU only) each with their own "Device" Profile" is the best solution you're going to get. |
||
|
|
nanoprobe
Master Cruncher Classified Joined: Aug 29, 2008 Post Count: 2998 Status: Offline Project Badges: 
|
GPU works don't take up CPU resources to the point of making such a big difference. Actually they do. Especially with Nvidia GPUs.
In 1969 I took an oath to defend and protect the U S Constitution against all enemies, both foreign and Domestic. There was no expiration date.
  |
||
|
|
|