Welcome to MilkyWay@home

Posts by Stojag

1) Message boards : Number crunching : Tesla K40 & System Upgrade Thoughts? (Message 61398)
Posted 15 Mar 2014 by Stojag
Post:
To be honest, i have no experience using SLI and there are no other posts in this forum about using SLI for Milky Way.

A quick google search about BOINC and SLI brought me to this post
http://boinc.berkeley.edu/dev/forum_thread.php?id=3592

They said it is better to disable SLI when doing SETI work using CUDA
The same might be correct for Milky Way.
2) Message boards : Number crunching : Tesla K40 & System Upgrade Thoughts? (Message 61395)
Posted 14 Mar 2014 by Stojag
Post:
I'm astonished how much money you are willing to put into your System.

But a K40 card is around 5000$. If you don't need the very special workstation features you can go with the new GTX Titan Black for just 1000$ that has the same performance.
With the leftover 4k you can build another Computer with those specs.

Milky Way Workunits need Double Precicion Calculations. Most Nvidia cards are poor at DP calcs (including yours and mine GTX 770), except for the TITAN cards.

The Titan Card is about 8 to 10 times fast than your GTX770, so you will complete much more WUs
3) Message boards : Application Code Discussion : AVX2 support (Message 61324)
Posted 4 Mar 2014 by Stojag
Post:
Although I don't know if Mikyway apps are using any kind of vector extention ops,i wonder if it's possible to get support for the new AVX2 operations.
Any Haswell CPU Core can do 16 DP FLOPs. For example my Xeon E3-1230 v3 (Quad-Core @3.7GHz) is therefore capable of outperforming my Nvidia GTX770 in terms of DP calculation performance.

Link for all people interesseted
4) Message boards : Number crunching : No credits, verify errors for all work (Message 61257)
Posted 28 Feb 2014 by Stojag
Post:
What kind of WUs are you crunching? CPU, GPU, MilkyWay@Home, MilkyWay@Home N-Body Simulation, Milkyway@Home Separation, Milkyway@Home Separation (Modified Fit) ?
My WUs seem to work all. Have go no errors so far.

[edit] I saw you are using ATI card. Mine is a Nvidia. Hope someone with an ATI card can verify your problem, or help you.
5) Message boards : Number crunching : Benchmark results - times wanted for any hardware, CPU or GPU, old or new! (Message 61252)
Posted 28 Feb 2014 by Stojag
Post:
If summarized some facts about the different Milkyway@Home Separation (Modified Fit) GPU WUs.
First comes the WUs name, then the typical runtime and the credits.

99% utilisation:
de_modfit_15_3s_modbpl_128_wrap_1_XXXX 608-610s (213.76 credits)
de_modfit_86_3s_modbpl_128wrap_rev_4_1_XXXX 306-308s (106.88 credits)
ps_modfit_15_3s_modbpl_128_wrap_1_XXXX 570-620s but mostly 608-610s (213.76 credits)
ps_modfit_16_bplmodfit_128_wrap_1_XXXX 608-610s (213.76 credits)

average for those "long" 213.76 credit WUs is 609s.

60% utilisation:
de_modfit_15_3s_128_wrap_1_XXXX 537-538s (213.76 credits)
de_modfit_86_3s_128wrap_rev_4_1_XXXX 271-272s (106.88 credits)
ps_modfit_15_3s_128_wrap_2_XXXX 536-540s (213.76 credits)
de_modfit_15_3s_128_wrap_1_XXXX 536-538s (213.76 credits)

Another iteresting fact is, that 213.76 credit WUs with 99% utilisation have an average CPU time of about 70 seconds whereas 60% utilisation WUs just have about 5.5s CPU time.
6) Message boards : Number crunching : Need Help! (Message 61248)
Posted 27 Feb 2014 by Stojag
Post:
thats what i use at the moment

its striped down to configure GPU WUs CPU and GPU utilisation

<app_config>
<app>
<name>milkyway</name>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>
<app>
<name>milkyway_separation__modified_fit</name>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>

</app_config>
7) Message boards : Number crunching : Benchmark results - times wanted for any hardware, CPU or GPU, old or new! (Message 61247)
Posted 27 Feb 2014 by Stojag
Post:
Oh, did you check to see if your CPU was going full turbo running MW?

CPU times were with full turbo (3,7GHz)

Re 60% util & 99% util tasks (weird btw!), I don't suppose each type was a particular credit WU was it?


Both 213.76 and 106.88 credit WUs have those 60% utilisation problem.
And on the other hand some of the 213.76 and 106.88 credit WUs have 99% utilisation.

So it doesn't depend on the credits.

If it's true that the TITAN get only 18% utiliisation on one WU than there might be some software issue. (Source: http://www.kwsnforum.com/viewtopic.php?p=204096#204096)
8) Message boards : Number crunching : Benchmark results - times wanted for any hardware, CPU or GPU, old or new! (Message 61240)
Posted 26 Feb 2014 by Stojag
Post:
Got some times for my CPU (Intel Xeon E3-1230 v3 @ 3.30GHz,up to 3.7 GHz Turbo speed)

9,709.11, 9,709.04, 9,707.05, 10,117.93, 10,163.91

average 9881,41 seconds

should be the "longer" WUs of the 213.76 WUs as i had some with ~7500 seconds runtime too.
9) Message boards : Number crunching : Benchmark results - times wanted for any hardware, CPU or GPU, old or new! (Message 61239)
Posted 25 Feb 2014 by Stojag
Post:
I've done some utilisation tests now

It seems like WUs with "modbpl" in their name get full utilisation, while the others work far below the limit. (on my NVIDIA card)

99% utilisation:
de_modfit_15_3s_modbpl_128_wrap_1_XXXX
de_modfit_86_3s_modbpl_128wrap_rev_4_1_XXXX
ps_modfit_15_3s_modbpl_128_wrap_1_XXXX
ps_modfit_16_bplmodfit_128_wrap_1_XXXX

60% utilisation:
de_modfit_15_3s_128_wrap_1_XXXX
de_modfit_86_3s_128wrap_rev_4_1_XXXX
ps_modfit_15_3s_128_wrap_2_XXXX
de_modfit_15_3s_128_wrap_1_XXXX

I can't say where the bottleneck is, but it dosen't seem like having half or a full CPU assigned to the GPU makes any difference. Windows task manager shows 0 to 1% CPU usage for the GPU WU. As Mumak said above it is most likely a OpenCL driver issue.
Therefore I think I can't make my WUs times much faster.

10) Message boards : Number crunching : Benchmark results - times wanted for any hardware, CPU or GPU, old or new! (Message 61237)
Posted 25 Feb 2014 by Stojag
Post:
To get the 100% GPU utilisation seems to be tricky.
To make sure no other tasks limit the GPU I shut down all other CPU WUs.
I used the app_config.xml to alter the CPU value on the WUs.

The "de_modfit_15_3s_128_wrap_1_XXXXXX" runs at ~60% utilisation although i assigned 1 CPU + 1 GPU to the WU.
On the other hand the "ps_modfit_15_3s_modbpl_128_wrap_1_XXXX" runs at 99% utilisation with 1 CPU + 1 GPU.


I've also tried to alter the CPU process priority from normal to high but that made no difference.

I'll post some results as soon as possible.
11) Message boards : Number crunching : Benchmark results - times wanted for any hardware, CPU or GPU, old or new! (Message 61233)
Posted 25 Feb 2014 by Stojag
Post:
I'm using a NVIDIA GTX 770 OC(1333 MHz core clock) Driver version 334.89, on Windows7, Boinc 7.2.39 (x64) having a theoretical 170 GFLOPS DP capability

My 5 fastest times are 537.16, 537.19, 537.30, 538.18, 538.19.
Average is 537,60 seconds

I am having a GPU utilisation of just 70% and 50% TDP when on single WUs. But at least it seems as I am the one with the fastest core clock.
CPU results are not ready yet as I am recieving no CPU WUs with modified fit code.
12) Message boards : News : New N-Body Runs (Message 61183)
Posted 22 Feb 2014 by Stojag
Post:
Warum verdrängen diese Aufgaben alle anderen Arbeiten (von 4 Aufgaben wird nur die eine von Milkyway bearbeitet)?
relbag


This actually is german.
He said: "Why do these workunits supress all others? (From 4 workunits only the milky way workunit will be done.)"





©2024 Astroinformatics Group