2nd test. This time on my i7 machine running WinXP with dual GTX260's, nvidia drivers 190.38, cuda 2.3 DLL's and BOINC 6.6.37. It has picked up 5 ABP1 wu, two of which are running as I type this. Normal crunch time for these is approx 8 hrs 20 mins on this machine. I will have to report progress later.
2nd test. This time on my i7 machine running WinXP with dual GTX260's, nvidia drivers 190.38, cuda 2.3 DLL's and BOINC 6.6.37. It has picked up 5 ABP1 wu, two of which are running as I type this. Normal crunch time for these is approx 8 hrs 20 mins on this machine. I will have to report progress later.
The 4th wu on the list has completed (it was the 1st one processed). Run time was approx 4 hrs 15 mins. Other work units running or waiting to run.
I notice in the output it says
Quote:
Using CUDA device #1 "GeForce GTX 260" (804.82 GFLOPS)"
Much as i'd like the card to be able to do 804Gflops it can't. Boinc reports the GTX260 (216sp) as being 96Gflops.
Also Bernd or Oliver is it possible to put the approx elapsed time in the output. GPUgrid do in their tasks. BOINC likes to report cpu time which isn't really relevant to cuda tasks.
Also Bernd or Oliver is it possible to put the approx elapsed time in the output. GPUgrid do in their tasks. BOINC likes to report cpu time which isn't really relevant to cuda tasks.
Cheers,
MarkJ
Actually, recent versions of BOINC (which CUDA testers will be using) do report both CPU and elapsed time - I know, I've just spent some time forging plausible entries (based on initial estimates) for the Celeron 400 MMX result I forgot to report. Something like 111 hours 40 minutes...... (see other thread).
But the BOINC server database doesn't have anythere to store the elapsed time, and these BOINC websites refuse to believe any CPU time report that is lower than the time between "Sent" and "Received" (as for my reissued task).
That's nothing to do with the particular version of the BOINC web code in use here ar Einstein - we're having exactly the same problem at AQUA, where their multi-threaded applications finish with CPU time up to four times higher than elapsed time.
2nd test. This time on my i7 machine running WinXP with dual GTX260's, nvidia drivers 190.38, cuda 2.3 DLL's and BOINC 6.6.37. It has picked up 5 ABP1 wu, two of which are running as I type this. Normal crunch time for these is approx 8 hrs 20 mins on this machine. I will have to report progress later.
Given the problems GPUgrid have been having with the 190.38 drivers I expected similar failures. Much to my suprise all 5 finished successfully. Well done for a 1st beta-test version.
I have decided to drop out of this Beta. The reasons are as follows:
o Using BOINC 6.6.28, XP Pro 32 SP3 on this system
1. E@H and SETI won't play nice with each other on gpus. This is probably a BOINC problem rather than the E@H CUDA app, but it's a show-stopper for me. When both are active, the E@H cpu tasks will run, but the gpu tasks never start, or if started, do not resume after SETI grabs the gpu. The only way to get the E@H gpu tasks to run is to suspend SETI and this is not acceptable.
2. When the E@H gpu task runs it usurps both the gpu and a full cpu. Although E@H throughput is increased, overall BOINC throughput takes a hit.
For me to resume the Beta you need to find some way that SETI and E@H can run nicely together. Most other projects using gpu run almost all processing on the gpu and only use the cpu to load data to the gpu. I might be willing to try the CC override to specify an extra cpu to 'fool' BOINC into running an extra WU on the cpu while E@H runs a gpu WU, but you need to get the cpu usage of those WUs lowered to a point where the extra WU gets useful cpu cycles as well.
People at GPUGrid are reporting errors with the GTX260 and 190.38 drivers (one even narrowed it down to the 192 version, has a 216 which works OK). Adding W7 and x64 to the mix adds to the probability that it's a driver issue.
The CUDA app will run only ABP1 tasks, right? If so, the project prefs need to be updated to allow one to select **ONLY** ABP1. I want to run ONLY the CUDA GPU tasks, and NOT any CPU tasks.
2nd test. This time on my i7
)
2nd test. This time on my i7 machine running WinXP with dual GTX260's, nvidia drivers 190.38, cuda 2.3 DLL's and BOINC 6.6.37. It has picked up 5 ABP1 wu, two of which are running as I type this. Normal crunch time for these is approx 8 hrs 20 mins on this machine. I will have to report progress later.
Links to wu:
1st wu
2nd wu
3rd wu
4th wu
5th wu
BOINC blog
RE: 2nd test. This time on
)
The 4th wu on the list has completed (it was the 1st one processed). Run time was approx 4 hrs 15 mins. Other work units running or waiting to run.
I notice in the output it says
Much as i'd like the card to be able to do 804Gflops it can't. Boinc reports the GTX260 (216sp) as being 96Gflops.
Also Bernd or Oliver is it possible to put the approx elapsed time in the output. GPUgrid do in their tasks. BOINC likes to report cpu time which isn't really relevant to cuda tasks.
Cheers,
MarkJ
BOINC blog
RE: RE: I notice in the
)
RE: On CPU, a WU needs
)
I'm sorry, I have made a typing error. It should read:
Bye, Grubix.
RE: Also Bernd or Oliver is
)
Actually, recent versions of BOINC (which CUDA testers will be using) do report both CPU and elapsed time - I know, I've just spent some time forging plausible entries (based on initial estimates) for the Celeron 400 MMX result I forgot to report. Something like 111 hours 40 minutes...... (see other thread).
But the BOINC server database doesn't have anythere to store the elapsed time, and these BOINC websites refuse to believe any CPU time report that is lower than the time between "Sent" and "Received" (as for my reissued task).
That's nothing to do with the particular version of the BOINC web code in use here ar Einstein - we're having exactly the same problem at AQUA, where their multi-threaded applications finish with CPU time up to four times higher than elapsed time.
RE: 2nd test. This time on
)
Given the problems GPUgrid have been having with the 190.38 drivers I expected similar failures. Much to my suprise all 5 finished successfully. Well done for a 1st beta-test version.
BOINC blog
I have decided to drop out of
)
I have decided to drop out of this Beta. The reasons are as follows:
o Using BOINC 6.6.28, XP Pro 32 SP3 on this system
1. E@H and SETI won't play nice with each other on gpus. This is probably a BOINC problem rather than the E@H CUDA app, but it's a show-stopper for me. When both are active, the E@H cpu tasks will run, but the gpu tasks never start, or if started, do not resume after SETI grabs the gpu. The only way to get the E@H gpu tasks to run is to suspend SETI and this is not acceptable.
2. When the E@H gpu task runs it usurps both the gpu and a full cpu. Although E@H throughput is increased, overall BOINC throughput takes a hit.
For me to resume the Beta you need to find some way that SETI and E@H can run nicely together. Most other projects using gpu run almost all processing on the gpu and only use the cpu to load data to the gpu. I might be willing to try the CC override to specify an extra cpu to 'fool' BOINC into running an extra WU on the cpu while E@H runs a gpu WU, but you need to get the cpu usage of those WUs lowered to a point where the extra WU gets useful cpu cycles as well.
Seti Classic Final Total: 11446 WU.
Cannot get the CUDA beta test
)
Cannot get the CUDA beta test to run on any of my three 64 bit systems :(
(a) Vista-64, AMD Opteron 275, 9800gtx+ with 190.38: see previous post in the 3.09 thread (which I probably should not have posted there)
(b) Vista-64, Intel Q6700, GTX280 with 190.38
I suspended all projects, then allowed E@H to get new tasks after I added the contents of the beta file.
this is what I saw after rebooting but before allowing new tasks for E@H
got this after the E@H CUDA actually started.
(c) Windows-7 64 bit, AMD Opteron 270, 9800gtx+ with 190.38
I assume the problem is the 64 bit os and will go back to seti and gpugrid for now.
Richard Haselgrove
)
The CUDA app will run only
)
The CUDA app will run only ABP1 tasks, right? If so, the project prefs need to be updated to allow one to select **ONLY** ABP1. I want to run ONLY the CUDA GPU tasks, and NOT any CPU tasks.
Reno, NV Team: SETI.USA