There's a new WU-Typ called "ps_sgr_208_test_..." but its a (bad) joke for GPUs:

It took less than 1sec for my slow ATI Radeon 4830 and I got 0,04Credits ?!?

If Travis would increase the WU-Size by 10000 % OK, but this current WU-Typ would increase the WU-Problem

some details :

<core_client_version>6.6.23</core_client_version>
<![CDATA[
<stderr_txt>
Running Milkyway@home ATI GPU application version 0.19e by Gipsel
setting minimum kernel frequency to 2 Hz
allowing 1 concurrent WUs per GPU
CPU: AMD Athlon(tm) 64 Processor 3200+ (1 cores/threads) 2.21133 GHz (611ms)

CAL Runtime: 1.4.283
Found 1 CAL device

Device 0: ATI Radeon HD 4800 (RV770) 1024 MB local RAM (remote 831 MB cached + 831 MB uncached)
GPU core clock: 575 MHz, memory clock: 750 MHz
640 shader units organized in 8 SIMDs with 16 VLIW units (5-issue), wavefront size 64 threads
supporting double precision

0 WUs already running on GPU 0
Starting WU on GPU 0

main integral, 20 iterations
predicted runtime per iteration is 2 ms (500 ms are allowed)
borders of the domains at 0 400
Calculated about 4.7964e+009 floatingpoint ops on GPU, 2.09546e+006 on FPU. Approximate GPU time 0.407227 seconds.
Calculated about 5.8591e+008 floatingpoint ops on FPU (stars).

WU completed.
CPU time: 1.3125 seconds, GPU time: 0.407227 seconds, wall clock time: 1.634 seconds, CPU frequency: 2.21133 GHz

</stderr_txt>
]]>

Looks like this is an issue with the workunit size. I was working with one of our new students last night getting him set up to start up searches and it looks like this test didn't go so well. There shouldn't be too many of these WUs left floating around.

---

YaY summer students :) Don't worry, we've all been there and done that =P

---

one step backwards
two steps forwards

the're is an other WU-Type "ps_sgr_208_1..." :
´
<core_client_version>6.6.23</core_client_version>
<![CDATA[
<stderr_txt>
Running Milkyway@home ATI GPU application version 0.19e by Gipsel
setting minimum kernel frequency to 2 Hz
allowing 2 concurrent WUs per GPU
CPU: AMD Athlon(tm) 64 Processor 3200+ (1 cores/threads) 2.2113 GHz (132ms)

CAL Runtime: 1.4.283
Found 1 CAL device

Device 0: ATI Radeon HD 4800 (RV770) 1024 MB local RAM (remote 831 MB cached + 831 MB uncached)
GPU core clock: 575 MHz, memory clock: 750 MHz
640 shader units organized in 8 SIMDs with 16 VLIW units (5-issue), wavefront size 64 threads
supporting double precision

0 WUs already running on GPU 0
Starting WU on GPU 0

main integral, 160 iterations
predicted runtime per iteration is 101 ms (500 ms are allowed)
borders of the domains at 0 1000
Calculated about 1.53283e+012 floatingpoint ops on GPU, 4.08581e+007 on FPU. Approximate GPU time 18.8037 seconds.
Calculated about 5.61279e+008 floatingpoint ops on FPU (stars).

WU completed.
CPU time: 2.54688 seconds, GPU time: 18.8037 seconds, wall clock time: 46.474 seconds, CPU frequency: 2.21134 GHz

</stderr_txt>
]]>




ps_sgr_208_1_130444_1243929472_0 - 11,5 Credits - These WUs are OK, but I got a lot of the other ps_sgr_208_test WUs

yeah, ps_sgr_208_1 is still running (and fine AFAIK), but i stopped the bad _test one has stopped and there shouldn't be any new WUs for it.

---

Here's another one:

Summer Student! wrote:

Task ID 71872834
Name ps_sgr_208_1_145929_1243931910_0
Workunit 70903844
Created 2 Jun 2009 8:38:31 UTC
Sent 2 Jun 2009 8:40:50 UTC
Received 2 Jun 2009 10:20:49 UTC
Server state Over
Outcome Success
Client state Done
Exit status 0 (0x0)
Computer ID 22807
Report deadline 5 Jun 2009 8:40:50 UTC
CPU time 487.5811
stderr out

<core_client_version>6.4.7</core_client_version>
<![CDATA[
<stderr_txt>
Running Milkyway@home version 0.19 by Gipsel
CPU: Intel(R) Core(TM)2 Quad CPU Q6600 @ 2.40GHz (4 cores/threads) 2.39991 GHz (196ms)

WU completed. It took 487.581 seconds CPU time and 491.775 seconds wall clock time @ 2.39995 GHz.

</stderr_txt>
]]>

Validate state Valid
Claimed credit 2.03907685949227
Granted credit 11.50045
application version 0.19

---

Seejay **Proud Member and Founder of BOINC Team Allprojectstats.com**

Last Problem:

The WU-Size is too small (even smaller than the other WUs)

We (ATI GPU User) need WUs with 10000% of the current one or 40000% if its gonna be calculated in Single Prec.

Last Problem:

The WU-Size is too small (even smaller than the other WUs)

We (ATI GPU User) need WUs with 10000% of the current one or 40000% if its gonna be calculated in Single Prec.

workunits on milkyway_gpu will be around 100-500 (or more) times the work of milkyway workunits and double precision gpus will still use double precision.

---

Last Problem:

The WU-Size is too small (even smaller than the other WUs)

We (ATI GPU User) need WUs with 10000% of the current one or 40000% if its gonna be calculated in Single Prec.

workunits on milkyway_gpu will be around 100-500 (or more) times the work of milkyway workunits and double precision gpus will still use double precision.

So we only have to wait for the Projectsite (feeder, trans, etc)

Last Problem:

The WU-Size is too small (even smaller than the other WUs)

We (ATI GPU User) need WUs with 10000% of the current one or 40000% if its gonna be calculated in Single Prec.

workunits on milkyway_gpu will be around 100-500 (or more) times the work of milkyway workunits and double precision gpus will still use double precision.

So we only have to wait for the Projectsite (feeder, trans, etc)

And working applications for the project. Workunits are not going to be the same as the ones here.

---

Gipsel is ready - only waiting for your (Travis) final decision how you gonna calc (SP or DP on GPU; SP on GPU, DP on CPU ....)

Gipsel is ready - only waiting for your (Travis) final decision how you gonna calc (SP or DP on GPU; SP on GPU, DP on CPU ....)

I thought it's already decided that most GPUs will use only SP. I plan to just try one or two summation approaches and ask, if the results are okay. I may also test different amounts of SP vs DP calculations on the CPU to see which CPU parts can get away with SP when the GPU does also only SP. I hope I get the time to do it.

Looks like the code is somehow complete now, except the checkpointing.
But I would really prefer the integration layout as in my proposal in the code discussion section, as it features less overhead and communication over the PCI-Express port. Especially with faster cards (look to the future!) it becomes much more important than for the GPU Travis uses for testing (afaik mobile 9600GT). Furthermore the delay for launching a kernel may be different between nv and ATI. But I guess it doesn't matter, if the ATI version uses a different layout than the CUDA version, as long as one gets a comparable precision.

One thing Travis has still to think about is the decision for which WUs DP is used and how this is flagged to the app(s). For nvidia cards most people would like to calculate only SP WUs (even if the card could do DP), but also with DP capable ATI cards some people may prefer the faster SP. One solution would be using two applications (with a DP opt-in checkbox on the website, must be per computer!). But I think mixing SP and DP for the GPU project is not that straightforward.

Hi there

i just get one wu working on mi gpu and timings are geting done for 32 hours !!! ?

Wu Nümber > ps_s25_17_273723_1243976097_0

sorry for mi pore inglisch writting

Hi there

i just get one wu working on mi gpu and timings are geting done for 32 hours !!! ?

Wu Nümber > ps_s25_17_273723_1243976097_0

sorry for mi pore inglisch writting

I had some ps_s25_17 WUs and they ran without any problem (it is just a normal 27 credit WU). Maybe the BOINC client and its prediction has some kind of a hiccup. A restart will most likely sort it out.

yes, i restart mi computer and the wu go on for 5 minutes again
Merci