nv270:
1 app: 4500 seconds 30-40% GPU load
2 apps: 6500-7000 seconds average 40-45% GPU load
3 apps: one of three restarts every 3-5 seconds, haven't saved logs, something about "app exited with zero code but haven't returned result, restarting app"
I know, that I get 2 additional CPU WUs running with nv270 app, but two 27000 seconds apps won't cover performance loss in GPU WUs.
what version of app do you use? 1.08 has same performance for fullCPU and for nv270
ÐºÐ°ÐºÐ°Ñ Ñƒ Ñ‚ÐµÐ±Ñ Ð²ÐµÑ€ÑÐ¸Ñ Ð¿Ñ€Ð¸Ð»Ð¾Ð¶ÐµÐ½Ð¸Ñ? на 1.08 ÑкороÑÑ‚ÑŒ Ð¾Ð´Ð¸Ð½Ð°ÐºÐ¾Ð²Ð°Ñ Ð½Ð° обоих вариантах.
3 Ð¿Ñ€Ð¸Ð»Ð¾Ð¶ÐµÐ½Ð¸Ñ Ñ€ÐµÑтартуют, Ñкорее вÑего, из-за нехватки памÑти. у Ð¼ÐµÐ½Ñ Ñ‚Ð°ÐºÐ¾Ðµ наблюдаетÑÑ Ð½Ð° гигабайтной 560
what version of app do you use? 1.08 has same performance for fullCPU and for nv270
ÐºÐ°ÐºÐ°Ñ Ñƒ Ñ‚ÐµÐ±Ñ Ð²ÐµÑ€ÑÐ¸Ñ Ð¿Ñ€Ð¸Ð»Ð¾Ð¶ÐµÐ½Ð¸Ñ? на 1.08 ÑкороÑÑ‚ÑŒ Ð¾Ð´Ð¸Ð½Ð°ÐºÐ¾Ð²Ð°Ñ Ð½Ð° обоих вариантах.
3 Ð¿Ñ€Ð¸Ð»Ð¾Ð¶ÐµÐ½Ð¸Ñ Ñ€ÐµÑтартуют, Ñкорее вÑего, из-за нехватки памÑти. у Ð¼ÐµÐ½Ñ Ñ‚Ð°ÐºÐ¾Ðµ наблюдаетÑÑ Ð½Ð° гигабайтной 560
I use 1.08. About 3 apps: i don't remember exact values, but nidia-smi reported that there is more than enough memory to run 3 apps.
1 wu does not use fixed amount of gpu ram, but this amoun floats during calculation
3 wu takes from 800 up to 1040 mb that is why u got restarting.
reduce it's number to 2
This may add even more confusion, but anyway: Apparently something went wrong with the optimization during compilation of the 1.07 Linux CUDA App, i.e. the CPU-Part is probably slower than it could be.
We already have a set of new Linux CUDA Apps built, I intend to publish these tomorrow as version 1.08.
Sorry for the confusion. Oliver is on his well-deserved vacation now, and we wanted to fix the existing problems before he left, but that went a bit hastily.
BM
The 1.08 nv270 application is performing quite a bit better. The performance is almost the same as the full CPU version with a difference of about 1-minute per task at the most. Thanks for the update and work on this application.
My observations: on dual AMD Opteron 2218 + GF 9800GX2 (host 15037; ancient combination, I know), it takes around 40% of CPU to keep GPU happy. I can't say anything about GPU utilization (due to known bug in nvidia-smi 270), however nicely high GPU temeperatures (around 80°C) indicate utilization in vicinity of 100%.
I've observed G-RAM usage as high as 310MB.
Unfortunately there is a bug somewhere between the OS and the NVidia driver such that this App doesn't run reliably on Mac OS 10.6 (SnowLeopard) (it keeps loosing the CUDA context). A bug report has been filed to NVidia, I believe that they're working on it.
Until this issue has been fixed distribution of this App is limited to Macs running Mac OS 10.5 (Leopard), i.e. Macs running 10.5, CUDA driver 3.2 and BOINC >= 6.10 should get it automatically.
More precisely the problem is present in Mac OS 10.6.6 and absent in Mac OS 10.5.8, (confirmed by NVidia). Actually we didn't bother to look at previous patch levels of these OS versions. Could be that the App works with some previous 10.6 version, if you can install the CUDA 3.2 driver there.
Some results about fullCPU
)
Some results about fullCPU and nv270 apps:
fullCPU:
1 app: 3000 seconds 40% GPU load
2 apps: 4200 seconds 70% GPU load
nv270:
1 app: 4500 seconds 30-40% GPU load
2 apps: 6500-7000 seconds average 40-45% GPU load
3 apps: one of three restarts every 3-5 seconds, haven't saved logs, something about "app exited with zero code but haven't returned result, restarting app"
I know, that I get 2 additional CPU WUs running with nv270 app, but two 27000 seconds apps won't cover performance loss in GPU WUs.
what version of app do you
)
what version of app do you use? 1.08 has same performance for fullCPU and for nv270
ÐºÐ°ÐºÐ°Ñ Ñƒ Ñ‚ÐµÐ±Ñ Ð²ÐµÑ€ÑÐ¸Ñ Ð¿Ñ€Ð¸Ð»Ð¾Ð¶ÐµÐ½Ð¸Ñ? на 1.08 ÑкороÑÑ‚ÑŒ Ð¾Ð´Ð¸Ð½Ð°ÐºÐ¾Ð²Ð°Ñ Ð½Ð° обоих вариантах.
3 Ð¿Ñ€Ð¸Ð»Ð¾Ð¶ÐµÐ½Ð¸Ñ Ñ€ÐµÑтартуют, Ñкорее вÑего, из-за нехватки памÑти. у Ð¼ÐµÐ½Ñ Ñ‚Ð°ÐºÐ¾Ðµ наблюдаетÑÑ Ð½Ð° гигабайтной 560
RE: We already have a set
)
Excellent work - the 0.20 CPU version is now very close to the fullCPU version in performance. Thanks for the fix!
RE: what version of app do
)
I use 1.08. About 3 apps: i don't remember exact values, but nidia-smi reported that there is more than enough memory to run 3 apps.
1 wu does not use fixed
)
1 wu does not use fixed amount of gpu ram, but this amoun floats during calculation
3 wu takes from 800 up to 1040 mb that is why u got restarting.
reduce it's number to 2
RE: This may add even more
)
The 1.08 nv270 application is performing quite a bit better. The performance is almost the same as the full CPU version with a difference of about 1-minute per task at the most. Thanks for the update and work on this application.
My observations: on dual AMD
)
My observations: on dual AMD Opteron 2218 + GF 9800GX2 (host 15037; ancient combination, I know), it takes around 40% of CPU to keep GPU happy. I can't say anything about GPU utilization (due to known bug in nvidia-smi 270), however nicely high GPU temeperatures (around 80°C) indicate utilization in vicinity of 100%.
I've observed G-RAM usage as high as 310MB.
Metod ...
There is now a BRP3 CUDA App
)
There is now a BRP3 CUDA App (1.08) for Mac OS.
Unfortunately there is a bug somewhere between the OS and the NVidia driver such that this App doesn't run reliably on Mac OS 10.6 (SnowLeopard) (it keeps loosing the CUDA context). A bug report has been filed to NVidia, I believe that they're working on it.
Until this issue has been fixed distribution of this App is limited to Macs running Mac OS 10.5 (Leopard), i.e. Macs running 10.5, CUDA driver 3.2 and BOINC >= 6.10 should get it automatically.
BM
BM
what about macos 10.6.6 ?
)
what about macos 10.6.6 ?
More precisely the problem is
)
More precisely the problem is present in Mac OS 10.6.6 and absent in Mac OS 10.5.8, (confirmed by NVidia). Actually we didn't bother to look at previous patch levels of these OS versions. Could be that the App works with some previous 10.6 version, if you can install the CUDA 3.2 driver there.
BM
BM