I fully expect this not to work, because it's my first attempt at releasing an GPU app for linux.

---

Is a BOINC 7 version needed ? Seems not to get send to my 6.10.58 hosts. Will switch boincs debugging on to be able to give more precise feedback.

---

_\|/_
U r s

Yes, I haven't enabled BOINC 6 yet, and I haven't decided what the minimum BOINC version should be.

---

Thanks for info. That explains why :
[debt] ATI GPU: no eligible projects

Will update BOINC version.

A little later :

BOINC at 7.0.28 : First two tasks are running!

---

_\|/_
U r s

Well, what were the results when you tested it before you released it? It worked then, right?

How to explain MB7 stderr log under AP6 result ??

http://setiweb.ssl.berkeley.edu/beta/result.php?resultid=11610941

How to explain MB7 stderr log under AP6 result ??

http://setiweb.ssl.berkeley.edu/beta/result.php?resultid=11610941

1. Check date and remember what did happen to the servers right after that date.
2. MBv7 (old result) log under AP6 result means database failure.

---

_\|/_
U r s

Well, what were the results when you tested it before you released it? It worked then, right?

It works right now !

resultlist hostid 50380
resultlist hostid 51991

---

_\|/_
U r s

Reporting: two wus ended with this error :

<message>
process exited with code 249 (0xf9, -7)
</message>
<stderr_txt>
OpenCL queue synchronized
SETI@home error -7 Garbled encoded workunit
in seti_parse_data()
File: seti.cpp
Line: 971

Just some buggy wu, if looking at all the wingmen.
WU 4969949
WU 4969948

---

_\|/_
U r s

Yeah, there were a bunch from March 5th like that. I thought they had all errored out by now.

---

About CPU usage while GPU(s) are in use :

HostID 50380 (openSUSE 12.2, Catalyst 12.8, 2x HD7750, GPUload: ca.66%) default setting was : (0.994 CPUs and 1.00 ATI GPUs)

After one day of continous crunching i calculated (0.289 CPUs and 1.00 ATI GPUs) for default settings.


HostID 51991 (Kubuntu 12.04, Catalyst 12.4, 2x HD6670, GPU load: ca.95%) default setting was : (0.993 CPUs and 1.00 ATI GPUs)

After one day of continous crunching i calculated (0.052 CPUs and 1.00 ATI GPUs) for default settings.

---

_\|/_
U r s

About CPU usage while GPU(s) are in use :

HostID 50380 (openSUSE 12.2, Catalyst 12.8, 2x HD7750, GPUload: ca.66%) default setting was : (0.994 CPUs and 1.00 ATI GPUs)

After one day of continous crunching i calculated (0.289 CPUs and 1.00 ATI GPUs) for default settings.


HostID 51991 (Kubuntu 12.04, Catalyst 12.4, 2x HD6670, GPU load: ca.95%) default setting was : (0.993 CPUs and 1.00 ATI GPUs)

After one day of continous crunching i calculated (0.052 CPUs and 1.00 ATI GPUs) for default settings.

How about the need of full core reserving under Linux ?

How about the need of full core reserving under Linux ?

Its definitely needed. But i can't test how many have to be reserved for highend GPUs. For one or two midrange GPUs one physical core is more than enough.

---

_\|/_
U r s

How about the need of full core reserving under Linux ?

Its definitely needed. But i can't test how many have to be reserved for highend GPUs. For one or two midrange GPUs one physical core is more than enough.

so, 0.052 CPUs and 1.00 ATI GPUs will probably not reserve any core?

How about the need of full core reserving under Linux ?

Its definitely needed. But i can't test how many have to be reserved for highend GPUs. For one or two midrange GPUs one physical core is more than enough.

so, 0.052 CPUs and 1.00 ATI GPUs will probably not reserve any core?

Cores are only reserved by boinc if the cumulative CPU requirement exceeds 1 (or 2/3/...) It's probably a >1 and not a >=1. With 0.51 two instances running will reserve one core but one instance won't. With 1 you may or may not get the desired behaviour of reserving one core per instance.

I don't know if it's even possible to choose numbers so that the first instance reserves one core, the second doesn't reserve an additional one but the third would.

And the sweet spot will very likely be system specific.

Do you _need_ a core reserved or does it only improve running?
Users may not like having their CPUs blocked that way, if not absolutely necessary.

Remember that the only lever Eric has available to pull is

<cpu_frac>x</cpu_frac>
the fraction of total FLOPs that are done by the CPU. This is used to calculate CPU usage and estimated FLOPS.

The server calculates <avg_ncpus> and <max_ncpus> from there - it'll be different for every host.

David's formula doesn't allow for requiring a CPU other than for productive science work.

Remember that the only lever Eric has available to pull is

x

That is true, and it really makes things difficult when that fraction is varying from GPU to GPU, which is often the case with OpenCL which will divide calculations between CPU and GPU depending upon the capabilities of the GPU.

I've reduced cpu_frac for ATI plan classes by 2/3. Let me know how it works out

---

Remember that the only lever Eric has available to pull is

<cpu_frac>x</cpu_frac>

That is true, and it really makes things difficult when that fraction is varying from GPU to GPU, which is often the case with OpenCL which will divide calculations between CPU and GPU depending upon the capabilities of the GPU.

<edit>I've reduced cpu_frac for ATI plan classes by 2/3. Let me know how it works out</edit>

Should i bring up something that is closer to a "baseline" then ?
AMD5050e+HD6450 or PentiumM+HD6450 ?

---

_\|/_
U r s

Sure.

---

This wus results seem to be a false negative, because of my wingmen WU 5010528

I've rerun this wu and got a match between stock linux x86_64 CPU and GPU, result: no overflow and no signals.

Maybe worth to have a look what goes wrong with this WU on windows apps.

---

_\|/_
U r s