More CUDA tasks

Questions and Answers : Windows : More CUDA tasks
Message board moderation

To post messages, you must log in.

AuthorMessage
Profile Yanivicious
Avatar

Send message
Joined: 29 Mar 12
Posts: 157
Credit: 15,529,301
RAC: 0
United States
Message 1219500 - Posted: 17 Apr 2012, 4:39:57 UTC

I am getting proportionately much more CPU tasks than CUDA tasks. My max work buffer is set to 7 right now (don't know if that has anything to do with it), but BOINC keeps requesting more CPU tasks (probably 30 are ready to start in my queue, with only one CUDA task ready to start in my queue). The GPU crunches pretty quickly so its shame to have it sitting idle while the CPU tries to catch up. Any thoughts on grabbing a more decent ratio of CPU to GPU tasks or is there nothing I can do?
ID: 1219500 · Report as offensive
Profile Jord
Volunteer tester
Avatar

Send message
Joined: 9 Jun 99
Posts: 15184
Credit: 4,362,181
RAC: 3
Netherlands
Message 1219522 - Posted: 17 Apr 2012, 5:22:32 UTC - in response to Message 1219500.  

There are several problems with your post, you know? ;-)

According to your list with computers, you run 6.12.34 on the one with the Nvidia GPU.

6.12.34 still has the "Connect to + Additional days of work" cache to work with, not the 7.0's "minimum work buffer + maximum additional work buffer". So saying that your "max work buffer is set to 7" is not correct, as in 6.12 that means you run something like a CI of 1 + ADD of 6 = 7 days.

Furthermore, 7.0 has the separate CPU + GPU scheduler and work fetch modules, they will really request work for each piece of hardware separately and 'keep score' about that (the REC based scheduler). 6.12 did have separate CPU + GPU schedulers, but would piggy-back as much of the work requests into one request.

And why did you post this in the Windows forum, why not in the GPU forum? Aren't you trying to get GPU work? Aren't you interested in helping a next person with maybe the same problem with getting a solution, that he may then easily read in the right forum?

....
Anyway...
The CI + ADD that you use is?
What does it say for the GPU deferral in the project properties in BOINC Manager? (Projects tab, Seti, properties, scheduling section)
ID: 1219522 · Report as offensive
Profile Yanivicious
Avatar

Send message
Joined: 29 Mar 12
Posts: 157
Credit: 15,529,301
RAC: 0
United States
Message 1219529 - Posted: 17 Apr 2012, 5:57:43 UTC - in response to Message 1219522.  

Apologies, I overlooked the GPU forum, I do see it now that I looked (only 3 down from the windows forum) so I would post there in the future.

Yes I am using 6.12.34 on the machine with the GPU, the "Connect about every" is set to 0 and the "additional work buffer" is set to 7. I didn't realize that the additional work buffer is not the same as the "max work buffer" in BOINC 7, I figured them both to mean maximum length of tasks to keep download. I leave it at 7 so that in case the S@H servers go down for a few days I still have more than enough tasks to keep crunching until the servers come back.

GPU (and CPU) deferral in the project properties both read as ----

I am actually not sure why the "connect about every.." is set to 0, I am pretty sure my other computers have that set to .2 by default. I probably did that a few days ago by mistake.
ID: 1219529 · Report as offensive
Profile Jord
Volunteer tester
Avatar

Send message
Joined: 9 Jun 99
Posts: 15184
Credit: 4,362,181
RAC: 3
Netherlands
Message 1219534 - Posted: 17 Apr 2012, 6:06:02 UTC - in response to Message 1219529.  
Last modified: 17 Apr 2012, 6:08:10 UTC

Yes I am using 6.12.34 on the machine with the GPU, the "Connect about every" is set to 0 and the "additional work buffer" is set to 7. I didn't realize that the additional work buffer is not the same as the "max work buffer" in BOINC 7, I figured them both to mean maximum length of tasks to keep download. I leave it at 7 so that in case the S@H servers go down for a few days I still have more than enough tasks to keep crunching until the servers come back.


No, with BOINC 6.xx the total cache is CI + ADD, and BOINC will try to keep it topped off at that max for all eternity. With BOINC 7.0, the cache will only be added to when the total cache size drops under the minimum work buffer value. Only then will BOINC go and fetch more work and try to cache the 7 days worth.

However, the minimum work buffer value depicts the main buffer, so for BOINC 7.0, if you want 7 days of work, you set minimum work buffer to 7.0 and maximum additional work buffer to 0. (the other way around, see? ;-))
If you were to use 0 or 0.2 as minimum work buffer, you can see that your cache is running almost on empty with 7.0 before it gets filled up again. It's possible that one or more cores are without work.

GPU (and CPU) deferral in the project properties both read as ----

OK, that at least means that previous requests got through. So then it's either that your BOINC thinks you have enough work in cache, period, or it's an internal debt situation, where the GPU has taken slightly more time than the CPU and it's payback time (although I thought that was on 7.0 only), or it's a situation where your BOINC asked for GPU work at one point and there wasn't any in the feeder at that time, and now it's in that hidden away project back-off. Uhm, it'll show up somewhere, go check the different status columns. ;-)
ID: 1219534 · Report as offensive
Profile Yanivicious
Avatar

Send message
Joined: 29 Mar 12
Posts: 157
Credit: 15,529,301
RAC: 0
United States
Message 1219544 - Posted: 17 Apr 2012, 7:00:23 UTC - in response to Message 1219534.  

i see. in this regard BOINC 7 seems to have a less confusing work-fetching scheme. I would like to update this computer to BOINC 7 - do you know if downloading and installing BOINC 7 will that affect the installed optimized lunatics apps?
ID: 1219544 · Report as offensive
Profile Jord
Volunteer tester
Avatar

Send message
Joined: 9 Jun 99
Posts: 15184
Credit: 4,362,181
RAC: 3
Netherlands
Message 1219576 - Posted: 17 Apr 2012, 11:45:15 UTC - in response to Message 1219544.  

It shouldn't. Nothing in the data directory should be affected.
ID: 1219576 · Report as offensive
Profile Yanivicious
Avatar

Send message
Joined: 29 Mar 12
Posts: 157
Credit: 15,529,301
RAC: 0
United States
Message 1219609 - Posted: 17 Apr 2012, 14:13:52 UTC - in response to Message 1219576.  

Thank you
ID: 1219609 · Report as offensive
Profile Jord
Volunteer tester
Avatar

Send message
Joined: 9 Jun 99
Posts: 15184
Credit: 4,362,181
RAC: 3
Netherlands
Message 1219614 - Posted: 17 Apr 2012, 14:35:01 UTC - in response to Message 1219576.  

It shouldn't. Nothing in the data directory should be affected.

Well, technically, but for client_state.xml, that is. ;-)

ID: 1219614 · Report as offensive

Questions and Answers : Windows : More CUDA tasks


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.