While it appeared to be processing v1.52 (BRP6-cuda32-nv270 jobs with high success in the very recent past, it has errored 37 straight v1.54 (BRP6-Beta-cuda55) jobs in about 4 hours, with zero successfully returned.
representative stderr text reads in part:
process exited with code 1 (0x1, -255)
../../projects/einstein.phys.uwm.edu/einsteinbinary_BRP6_1.54_x86_64-pc-linux-gnu__BRP6-Beta-cuda55: /lib64/libc.so.6: version `GLIBC_2.7' not found (required by ../../projects/einstein.phys.uwm.edu/einsteinbinary_BRP6_1.54_x86_64-pc-linux-gnu__BRP6-Beta-cuda55)
What I've seen so far looks like a 20-25% improvement over the CUDA 3.2 version.
I've been busy elsewhere and only recently noticed that all my hosts with nvidia GPUs have been getting 1.53 and 1.54 beta tasks. I had inadvertently left the beta tasks settings on. Because of the cache settings I use, only one host had got to the 1.53 and none had got anywhere near the 1.54.
I've quickly gone around the other 20 hosts involved and 'rebranded' all the 1.53 tasks as 1.54 so that they will actually be crunched rather than crashing. I 'rebranded' around 80 tasks and promoted a few but it's too early to make any guess as to the 'improvement' :-). Maybe tomorrow when some of them have finished.
I was considering attempting to 'rebrand' the 1.52s as well, but decided to leave well enough alone. I surmised that a change in 'plan class' might very well be an issue. I was quite happy to 'rebrand' the plan class as well as the version number but I wondered what would happen if the result came back with a different plan class from what it was issued with. Would such tasks be rejected?
I have no idea about how the Client would react to it, but the sever shouldn't reject the result. Most likely it will keep the original app version info of the task and won't update it when it's reported.
I've shifted all the results and discussion for the new cuda55 app for Linux to this new thread in Crunchers Corner so that Bernd doesn't have to compete with all of us wanting to post interesting performance data and commentary. No doubt there will be further announcements as apps for other platforms are deployed.
A little over two hours ago one of my Windows 7 hosts received CUDA55 Parkes PMPS work. So I'm happy to see a Windows application added to the previous Linux application.
Sadly, first trials on all three of my Windows 7 GPU hosts have given very fast (less than two elapsed seconds) error exits, all with exit status -1073741515 (0xffffffffc0000135).
Here is a link to the task result page for a representative one of these failing 1.54 Windows CUDA55 tasks.
All three of my failing hosts are 64 bit Windows 7 machines. All three run the 344.60 driver version. If others succeed where I have failed it may be worthwhile to explore other configuration details.
Lacking Linux myself, I
)
Lacking Linux myself, I browsed a little among my quorum partners looking for any 1.54 CUDA55.
I happened to stumble on this host
While it appeared to be processing v1.52 (BRP6-cuda32-nv270 jobs with high success in the very recent past, it has errored 37 straight v1.54 (BRP6-Beta-cuda55) jobs in about 4 hours, with zero successfully returned.
representative stderr text reads in part:
../../projects/einstein.phys.uwm.edu/einsteinbinary_BRP6_1.54_x86_64-pc-linux-gnu__BRP6-Beta-cuda55: /lib64/libc.so.6: version `GLIBC_2.7' not found (required by ../../projects/einstein.phys.uwm.edu/einsteinbinary_BRP6_1.54_x86_64-pc-linux-gnu__BRP6-Beta-cuda55)
The errors mostly come from
)
The errors mostly come from Gaurav's cluster, which is running a too old libc version. I'm in contact with him to sort this out.
What I've seen so far looks like a 20-25% improvement over the CUDA 3.2 version.
BM
BM
RE: What I've seen so far
)
That looks really nice. When can we expect a Beta version for Windows?
-----
RE: What I've seen so far
)
I've been busy elsewhere and only recently noticed that all my hosts with nvidia GPUs have been getting 1.53 and 1.54 beta tasks. I had inadvertently left the beta tasks settings on. Because of the cache settings I use, only one host had got to the 1.53 and none had got anywhere near the 1.54.
I've quickly gone around the other 20 hosts involved and 'rebranded' all the 1.53 tasks as 1.54 so that they will actually be crunched rather than crashing. I 'rebranded' around 80 tasks and promoted a few but it's too early to make any guess as to the 'improvement' :-). Maybe tomorrow when some of them have finished.
I was considering attempting to 'rebrand' the 1.52s as well, but decided to leave well enough alone. I surmised that a change in 'plan class' might very well be an issue. I was quite happy to 'rebrand' the plan class as well as the version number but I wondered what would happen if the result came back with a different plan class from what it was issued with. Would such tasks be rejected?
Cheers,
Gary.
RE: Would such tasks be
)
I have no idea about how the Client would react to it, but the sever shouldn't reject the result. Most likely it will keep the original app version info of the task and won't update it when it's reported.
BM
BM
I've shifted all the results
)
I've shifted all the results and discussion for the new cuda55 app for Linux to this new thread in Crunchers Corner so that Bernd doesn't have to compete with all of us wanting to post interesting performance data and commentary. No doubt there will be further announcements as apps for other platforms are deployed.
Cheers,
Gary.
A little over two hours ago
)
A little over two hours ago one of my Windows 7 hosts received CUDA55 Parkes PMPS work. So I'm happy to see a Windows application added to the previous Linux application.
Sadly, first trials on all three of my Windows 7 GPU hosts have given very fast (less than two elapsed seconds) error exits, all with exit status -1073741515 (0xffffffffc0000135).
Here is a link to the task result page for a representative one of these failing 1.54 Windows CUDA55 tasks.
All three of my failing hosts are 64 bit Windows 7 machines. All three run the 344.60 driver version. If others succeed where I have failed it may be worthwhile to explore other configuration details.
Sorry, that got out a little
)
Sorry, that got out a little early. Deprecated the Windows app version for now.
BM
BM
RE: Sorry, that got out a
)
I have a several of the BRP6 betas on my windows machine as well. Should I go ahead and and abort those?
The application failed to
)
The application failed to initialize properly (0xc0000135)
That usually means a needed DLL file can't be found - I'm guessing no, or the wrong, CUDA runtime.
Every Google search for the error code says that it's the Dot Net runtime which is missing, but that won't be the case here.