Actually its worse than 1.08 (The recalc stage takes longer) for this host (https://einsteinathome.org/host/13160863) which negates the CUDA advantage. I'll try to create an app_info to target 1.08
XLAL Error - XLALFstatCheckSFTLengthMismatch (/home/boinc/lalsuite/EinsteinAtHome/source/lalsuite/lalpulsar/lib/ComputeFstat.c:1895): Length of input SFTs (1800 s) must be less than 1730.29 s for CW signal with
frequency = 1237, spin-down/up = -2.7e-09, binary asini = 0, period = 0, ecc = 0,
in order to stay below an Fstat-mismatch of 0.01.
there was a clear change at 40% where the GPU memory bandwidth use dropped to almost nothing, and the VRAM use dropped below 1GB. then at 50%, the task restarted, reloaded the VRAM to ~2.4GB, then the task resumed with high GPU and memory bus utilization. I assume it will do the same thing at 90-100%. very nice.
I hesitate to say this, but
)
I hesitate to say this, but version 1.11 does not change anything significant compared to 1.08/1.10.
Host 12801270.
Actually its worse than 1.08
)
Actually its worse than 1.08 (The recalc stage takes longer) for this host (https://einsteinathome.org/host/13160863) which negates the CUDA advantage. I'll try to create an app_info to target 1.08
DF1DX wrote: I hesitate to
)
I know. 1.12 should do better.
BM
Just FYI,First 1.11
)
Just FYI,
1.11 https://einsteinathome.org/task/1587604598 and others crash on startup.
1.12 https://einsteinathome.org/task/1587618661 and others crash at the 49.5 percent mark.
Host https://einsteinathome.org/host/12791541.
Hm. 1.12 does now do the
)
Hm. 1.12 does now do the "recalc" on the GPU, but it seems that this code doesn't work (yet).
BM
Beta 1.13: Error while
)
Beta 1.13: Error while computing
yeah V1.13 error right
)
yeah V1.13 error right away.
_________________________________________________________________________
Yep - new version underway.
)
Yep - new version underway.
BM
nice. 1.14 is running now.
)
nice. 1.14 is running now. need to check that it makes it past the 49.5% transition and to completion.
so all stages should be on GPU now? no long CPU-only portion?
_________________________________________________________________________
seems like it's
)
seems like it's working.
there was a clear change at 40% where the GPU memory bandwidth use dropped to almost nothing, and the VRAM use dropped below 1GB. then at 50%, the task restarted, reloaded the VRAM to ~2.4GB, then the task resumed with high GPU and memory bus utilization. I assume it will do the same thing at 90-100%. very nice.
_________________________________________________________________________