I recompiled beignet from the latest git, but it didn't fix the problem.
Quote:
Process creation (../../projects/einstein.phys.uwm.edu/einsteinbinary_BRP6_1.52_x86_64-pc-linux-gnu__BRP6-Beta-opencl-intel_gpu) failed: Error -1, errno=13
execv: Permission denied
You didn't get the permissions right. Since you are apparently running distro version of BOINC you need to make the app executable by user 'boinc' so either make the app's owner 'boinc' and give execute permission to owner or give execute permission to others.
Quote:
I read somewhere, that boinc needs root privileges to run tasks with beignet, but had forgotten about it. I'll post my results tomorrow.
Root privileges or setting kernel command line parameters aren't needed unless you don't have X running. Also if you have ATI or NVIDIA GPUs and you use proprietary drivers for those then I suspect you'd need to use workarounds.
I compiled beignet from the latest git. It compiled and worked, but it did not verify against BRP6-cuda32-nv301.
So the latest version of beignet known to work with einstein is 0.9.2.
There's a new Known Issue on the Beignet homepage for Haswells regarding shared local memory (__local) which I think applies to BRP app. Did you try patching the kernel or using the kernel command line parameter?
Questions: Do older versions of beignet work on Haswell without patching the kernel? Cause that's the platform, where I want to use it in production. Haswell 24/7-running i5.
Questions: Do older versions of beignet work on Haswell without patching the kernel? Cause that's the platform, where I want to use it in production. Haswell 24/7-running i5.
I didn't have any luck on i7-4790 with unpatched 3.13 kernel. 4.1 kernel has a command line parameter that should help but it's still in release candidate stage.
With that i5 do you mean the i5-3570 on your computer list? Isn't that an Ivy Bridge?
What was needed was Linux kernel 4.1 and "i915.enable_ppgtt=2" on the kernel command line.
While Beignet 0.9.2 was good to run BRP on Ivy Bridge it isn't as good on Haswell. Some tasks validate just fine but others cause GPU hangs. The only other version that works is 1.1 built from git master. The one from commit e64445f seems to be working.
Performance is not what I expected. i7-3770 completed BRP4 tasks in 14 minutes, now i7-4790 with Beignet 0.9.2 needs 16 minutes and with 1.1 (git-e64445f) 20 minutes.
----
Tasks that cause GPU hangs create results like the following:
So the numbers in second, third and fourth columns are the same on every line. To human that result is (imho) obviously invalid and I was expecting the validator to throw the result out right away but it didn't do that.
I run one such task several times and even had one reboot between runs. The result file was identical on every run. I think that suggests that another host could create a result that's identical to mine. If using Beignet on this project ever catches on I think there is some risk of two such bad results validating each other.
It doesn't actually 'claim' or 'apply for' credit at all. It simply reports how long the processing took, and the rather elderly server software here works out that figure, on the assumption that the "CPU time: 493.70 seconds" was the only processing done. At this stage, the server doesn't take any notice of the fact that the bulk of the work was done on the intel_gpu.
Once your work is validate by a second user, you should get the same standard fixed credit as everyone else - 1,000 credits. That assumes that the results match.
RE: I recompiled beignet
)
You didn't get the permissions right. Since you are apparently running distro version of BOINC you need to make the app executable by user 'boinc' so either make the app's owner 'boinc' and give execute permission to owner or give execute permission to others.
Root privileges or setting kernel command line parameters aren't needed unless you don't have X running. Also if you have ATI or NVIDIA GPUs and you use proprietary drivers for those then I suspect you'd need to use workarounds.
RE: I compiled beignet from
)
There's a new Known Issue on the Beignet homepage for Haswells regarding shared local memory (__local) which I think applies to BRP app. Did you try patching the kernel or using the kernel command line parameter?
I did a "echo 0 >
)
I did a "echo 0 > /sys/module/i915/parameters/enable_cmd_parser" and patched the Kernel (4.0.4) with https://01.org/zh/beignet/downloads/linux-kernel-patch-hsw-support
But I only ran one WU, which turned out invalid.
Hello, thanks for the hint
)
Hello,
thanks for the hint with the execution flag. I just set 777 and it worked.
This laptop has some serious heating problems, but I manged to complete one task:
http://einsteinathome.org/workunit/220732263
It's waiting for validation now. The cpu usage was < 1%, but the gnome user interface was very slow during the processing.
Looking forward to see weather it'll be validated correctly.
How's the roadmap for a configuration which doesn't need an app_info.xml?
Bye
MPW
Nice, it was validated
)
Nice, it was validated correctly.
Gonna try this on an Intel BayTrail next.
Questions: Do older versions of beignet work on Haswell without patching the kernel? Cause that's the platform, where I want to use it in production. Haswell 24/7-running i5.
RE: Questions: Do older
)
I didn't have any luck on i7-4790 with unpatched 3.13 kernel. 4.1 kernel has a command line parameter that should help but it's still in release candidate stage.
With that i5 do you mean the i5-3570 on your computer list? Isn't that an Ivy Bridge?
RE: With that i5 do you
)
Of course you're right. Never mind, will try it.
Got Beignet working on
)
Got Beignet working on Haswell.
What was needed was Linux kernel 4.1 and "i915.enable_ppgtt=2" on the kernel command line.
While Beignet 0.9.2 was good to run BRP on Ivy Bridge it isn't as good on Haswell. Some tasks validate just fine but others cause GPU hangs. The only other version that works is 1.1 built from git master. The one from commit e64445f seems to be working.
Performance is not what I expected. i7-3770 completed BRP4 tasks in 14 minutes, now i7-4790 with Beignet 0.9.2 needs 16 minutes and with 1.1 (git-e64445f) 20 minutes.
----
Tasks that cause GPU hangs create results like the following:
[pre]0.819292935458 1462.994140625000 0.192481309175 1.753485441208 17.4516 14.6716 16
0.608097423207 1462.994140625000 0.192481309175 1.753485441208 17.0202 14.0824 16
0.680923461914 1462.994140625000 0.192481309175 1.753485441208 16.7255 13.6822 16[/pre]
So the numbers in second, third and fourth columns are the same on every line. To human that result is (imho) obviously invalid and I was expecting the validator to throw the result out right away but it didn't do that.
I run one such task several times and even had one reboot between runs. The result file was identical on every run. I think that suggests that another host could create a result that's identical to mine. If using Beignet on this project ever catches on I think there is some risk of two such bad results validating each other.
p2030.20150531.G32.86+00.63.C
)
p2030.20150531.G32.86+00.63.C.b6s0g0.00000_3408_0 223767791 11955750 23 Jul 2015 11:51:59 UTC 30 Jul 2015 11:51:59 UTC è®¡ç®—ä¸ --- --- --- --- Binary Radio Pulsar Search (Arecibo, GPU) v1.52 (BRP4G-Beta-opencl-intel_gpu)
p2030.20150531.G68.35+00.83.S.b3s0g0.00000_1536_0 223713556 11955750 22 Jul 2015 2:06:33 UTC 23 Jul 2015 14:08:40 UTC 已完æˆï¼Œç‰å¾…éªŒè¯ 81,587.69 493.70 4.28 ç‰å¾…ä¸ Binary Radio Pulsar Search (Arecibo, GPU) v1.52 (BRP4G-Beta-opencl-intel_gpu)
p2030.20150531.G68.35+00.83.S.b4s0g0.00000_1040_0 223713287 11955750 22 Jul 2015 2:06:33 UTC 22 Jul 2015 6:41:11 UTC 被用户ä¸æ¢ 0.00 0.00 0.00 --- Binary Radio Pulsar Search (Arecibo, GPU) v1.52 (BRP4G-Beta-opencl-intel_gpu)
为什么è¿è¡Œ81,587.69秒，åªç”³è¯·4.28分？
Why did I run 81587.69 seconds but only apply for 4.28 credits?
RE: p2030.20150531.G32.86+0
)
Task 510876799
It doesn't actually 'claim' or 'apply for' credit at all. It simply reports how long the processing took, and the rather elderly server software here works out that figure, on the assumption that the "CPU time: 493.70 seconds" was the only processing done. At this stage, the server doesn't take any notice of the fact that the bulk of the work was done on the intel_gpu.
Once your work is validate by a second user, you should get the same standard fixed credit as everyone else - 1,000 credits. That assumes that the results match.