I'm still learning current apps behavior and looking through completed results found this one:
Run time (sec):8,912.38
CPU time (sec):8,761.12
Peak working set size (MB):451.89
Peak swap size (MB):15881.12
Peak disk usage (MB):4.47
Состояние проверки:Valid
Предоставлено очков:2,000
Приложение:Gravitational Wave search O2 Multi-Directional GPU v2.07 (GW-opencl-nvidia)
x86_64-pc-linux-gnu
So, ~15GB of SWAP space was used, really?
Is this normal/usual for this app?
It's NV GPU app, btw.
EDIT: and here is result for CPU app of same type from same host:
Run time (sec):36,855.66
CPU time (sec):28,208.21
Peak working set size (MB):2158.96
Peak swap size (MB):2161.07
Peak disk usage (MB):4.48
Состояние проверки:Valid
Предоставлено очков:1,000
Приложение:Gravitational Wave search O2 Multi-Directional v2.07 (GWnew)
x86_64-pc-linux-gnu
Big memory usage, but not 15GB still...
I'm quite curious why GPU app takes so much of SWAP
You'll probably have to get a
)
You'll probably have to get a reply from some of the old-timers that really understand the applications. Maybe Bernd will chime in.
I'm guessing that because the gpu is processing in parallel, that when it comes time to push the interim calculated results back to the cpu for final FP64 processing, that the arrays used in the gpu are much larger in size than the serial arrays used in the cpu application.
Just wild arse guesses on my part. Take with salt.
I'm testing with 3 per gpu,
)
I'm testing with 3 per gpu, 12 at a time total. Main memory usage is about 8 Gb total including OS, browser etc. The virt column says 22g but my guess is that NV-drivers map some of GPU memory to virtual main memory for the transfers.
So you think those 15GB are
)
So you think those 15GB are reserved but not committed ones?
Would be interesting to try the same under Windows, where I would have additional means for analysis.