Welcome to MilkyWay@home

Posts by Mike Simicsak

1) Message boards : Number crunching : New BOINC Manager = no MWAH WUs (Message 54044)
Posted 16 Apr 2012 by Mike Simicsak
Post:
Assumption: You are running more projects than MW@H? I am too. I run 4. SETI in particular will flood you with a ton of tasks. As has already been stated MW@H is limited to 40 tasks per GPU. I am assuming it's another project that flooded you with WU's.

I was running 4 projects. I had two projects set to "No New Tasks". I was wondering if there was a chance that the "No New Tasks" was bleeding over from Collatz/PrimeGrid onto MW@H which is now my main project. I removed one and with the improved scheduling in ver7 I am trying to accept tasks from all Projects again. For me it seemed like the "minimum work buffer" solved it although I was kind of doing both things at the same time. I'm still curious if the "No New Tasks" code is implicated?

Do you have "No New Tasks" set for any other project? Can you try (temporarily) removing all other projects and then adjust the "minimum work buffer" setting to see if when MW@H is by itself that resolves it?

If you haven't already another thing I did was remove MW@H, restart BOINC, and then add it back in just to make sure there were no ver 6.x to 7.x upgrade issues. It was after this that I found and adjusted the minimum buffer setting.
2) Message boards : Number crunching : Download more WUs (Message 54041)
Posted 15 Apr 2012 by Mike Simicsak
Post:
See this thread: http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=2855#54031
From BOINC client (7.0.25), click tools, preferences. Switch to Network preferences. Adjust "minimum work buffer" value. You want to increase it.
3) Message boards : Number crunching : 7970 crunching more WUs (Message 54040)
Posted 15 Apr 2012 by Mike Simicsak
Post:
For me, the bulk of my credits come from the GPU. I have recently had an issue with Apple Mobile Device manager where it was "running away" and consumming an entire core. Adding in contention with the 5 CPU tasks I generally run, the GPGPU Open_CL WU's were getting no CPU time and the GPU was idle. It is a good illustration that keeping the GPU busy requires some CPU cycles. It took me about a month to figure out what the problem was and to get it corrected. My averages are still depressed from this incident.

Previously the MW@H WU's showed about 5% CPU load. Even with the recent change (upg BOINC client to 7.0.25), my experience is that it does not take 96% of a core to keep the GPU busy; it's still much more like 5%. My observation is that this appears to be a frontend/backend load to get the WU in and out of the GPU.

I do have an IT background, so hopefully this will make sense. My idea was that the GPU tasks should be isolated from the CPU tasks to avoid any contention. As had already been posted this is partially accomplished by "reserving" cores by reducing the maximum CPU% available to the BOINC client by converting the fraction 1/cores to a percentage and subtracting from 100%. In my case with 6 cores, 1/6 = 16% and I set the maximum CPU utilization to 84%. If you observe task manager after making this change, it will appear that all cores are loaded but none are 100% busy. The next part involves utlizing the concept of proccess / processor affinity. In simple terms this sort of means "gluing" a task/program/process to a CPU core. This can be observed through task manager. Goto the "Processes" tab, right click on a process and select "set affinity" from the popup menu. There will be a check box for each core. You can change CPU/Core affinity by changing the check boxes. On my system GPU WU take about 1 minute but CPU WU run for about an hour. The "easiest" way to make a difference is to de-select core zero from all the CPU tasks since they run the longest. This should produce a nice 100% busy line for all the CPU WU cores. The ever changing GPU WU will not prefer any core but since all the cores except zero are busy, it will run on core zero.

I've been working with my son to write a C# program to change a task's processor affinity or to say it another way to automate the process I've just described. The program reads a list of active processes and changes the processor affinity for GPU / "openCL" tasks to "prefer" processor zero. It also changes the CPU tasks to "avoid" processor zero. It's cool to watch this in task manager as the "curve" changes from random up down on all 6 cores to core zero being random and the other 5 cores are straight lines at 100%. With the different workloads isolated to different processor cores and with "nothing" else running on the system, it's easy to see that a GPU WU still only uses about 5% of a processor core. With the CPU load I'm seeing, I'm suggesting that it might be possible, even with a 79x0, to use 1 core to keep that beast busy if you can make sure the GPU pre/post processor task always has access to CPU time by utilizing processor affinity.

I've had this working for about 2 weeks now and my average MW@H scores have jumped from about 180,000 credits a day to close to 215,000 credits a day. We're considering generalizing this, making it more user friendly, making it more cross-project usable, and posting it for general availability. Is anyone interested in using a tool like this?

Is there anyone out there who uses either NVidia and/or Multiple GPU's that might be willing to help us test?
4) Message boards : Number crunching : New BOINC Manager = no MWAH WUs (Message 54030)
Posted 15 Apr 2012 by Mike Simicsak
Post:
Thanks Floyd that solved it for me too! :-)


To expand:
This change didn't make sense to me which is why I didn't try it before you suggested it. It seems like there is a limit to how many tasks you can have locally in buffer/queue and I was already there. Even after changing it, the number of tasks hasn't changed. The value was set to 0.10 (days) which I would think is 2.4 hours. At max, I only see about 40 tasks which at an estimated 1:15 would only be 50 minutes of work so at first glance it already seemed set higher than what was being queued locally.

Anyway, I changed it to 0.50 and voila, it's working. Thanks again.
5) Message boards : Number crunching : New BOINC Manager = no MWAH WUs (Message 54025)
Posted 15 Apr 2012 by Mike Simicsak
Post:
Upgraded to 7.0.25. WU's download and crunch but no longer auto upload. Primary app is OpenCL on ATI 5870. Will download anywhere from 10-40 WU's but once they are done, they just sit there. If I click update, they upload and new ones are downloaded. On the 5870 they take about 1 min 5-8 sec so it doesn't stay busy for very long. I've removed the project and added it back but no difference. Also disabled firewall. Any suggestions?

It used to auto-upload each WU before in the 1 minute interval before the next one (WU) completed.

Hmmm, found this:
4/14/2012 5:47:32 PM | Milkyway@Home | Reporting 15 completed tasks, not requesting new tasks




©2024 Astroinformatics Group