Posts by mmonnin

1) Message boards : News : GPU app - beta version for linux nvidia (Message 2348)
Posted 1 Apr 2019 by mmonnin
Post:
I see credit was recently doubled to 8,000 (and the average runtime also seems to have increased with it, though not quite doubled) and now decreased to 800. I'm not complaining, just curious about the reasoning.


The difference in average times on the CPU went from 1.5 hours for the DS15x11 WUs to 3 hours for the DS14x12 WUs (these are averages over 100 WUs, all on the same machine). This is the reason for the initial doubling. So it is approximate, and of course there is no guarantee the GPU times will scale the same, but it should be close.

Then, as discussed in another thread, the credits were way too high for CPUs. To bring them back in line, I had to scale them back down by 10. The cpu WUs are still paying almost 2x the original (from several weeks ago).

Here's the problem: The original crediting system could be majorly abused by GPUs, so we had to change it. CreditNew paid next to nothing, so now we are paying a fixed amount per WU. This is not ideal either, as GPUs and CPUs should be treated differently. I think the best solution will be to use CreditNew, but hack the validator code to scale up the payout for GPUs. That will take work though, and I'd rather spend my energy finishing the GPU apps.

How does the current credit/hour for the GPU compare to other projects?


One of the lowest credit projects now for GPUs. 261k for a 1080Ti yesterday but not sure if that was all at 800 or some older 4k/8k tasks. SETI uses CreditNew and will be better than 1 credit/sec so like 100k or something. Asteroids is probably lower but A@H is more of a CPU project IMO. Collatz, MW, E@H, GPUGrid, Enigma, Amicable, PrimeGrid, and Moo produce more.

If the GPU app can do 20x more work then per second it should receive 20x the credit compared to your own CPU app.
2) Message boards : Number crunching : GPU Error (Message 2328)
Posted 29 Mar 2019 by mmonnin
Post:
I got a little sneaky solution for me.

I simply deleted this files as root:

/etc/OpenCL/mesa.icd and intel.icd

reboot. It's getting nomore errors.

At administration:
Could u send me some GPU-wus , please ?
It always "says" job cache full.


That's the BOINC client saying job cache is full so it won't contact any project server for more work.
3) Message boards : News : GPU app - beta version for linux nvidia (Message 2294)
Posted 26 Mar 2019 by mmonnin
Post:
It was paying around 20 per run sec on my GTX 1080 which is better than many GPU apps but not crazy. Now it's only paying 0.08 per run sec, 1/250 of what it was which is crazy low. It's worse paying than Asteroids on a GPU. Einstein pays 5-7 per run sec.


Maybe I'll try CPU and see if the points are better than GPU payout. :(

edit: If there is no significant point (not number of WUs completion) gain over CPU app, it certainly defeats the purpose of creating GPU app in the first place. I know some of us crunchers are crazy about points rather than the number of WUs completion. Like Asteroid, maybe pay a fix credits for each WU completion which is what is important to NumberField project, rather than the point. In general, the points are to attract crunchers. My one cent.


Tasks vary in length so fixed credit is unfair as well. The only other way is if the FLOPs calculated during the task was known for each task. The prior method was open to abuse since it was a function of the GPUs FLOPs and time.
4) Message boards : News : GPU app - beta version for linux nvidia (Message 2277)
Posted 26 Mar 2019 by mmonnin
Post:
lol, at least you guys are getting tasks. I just keep getting "got 0 new tasks".

I disabled cpu tasks, just to test the GPU versions. I am running driver 418.56 AND I have a compute 3.0 card. Are there any tasks or am I configured incorrectly?


It seems like the 760 should be able to receive tasks. The app is Linux only so the 210 won't work. And it's to old.
5) Message boards : News : GPU app - beta version for linux nvidia (Message 2274)
Posted 25 Mar 2019 by mmonnin
Post:
Only errors here.
GTX645; newest drivers 418
Anyone can help?

PG PPS sieve works fine.


2nd person for a 6xx card to not work with 418. The card might just be too old.
6) Message boards : News : GPU app - beta version for linux nvidia (Message 2265)
Posted 25 Mar 2019 by mmonnin
Post:
Yesterday was triple the output of the Formula BOINC Sprint's best day.


Unfortunately, I don't think those numbers are accurate, just based on the numbers of results coming back. It's correlated to the obscenely high GPU credits being awarded. Maybe it's time I change to "Credit New", from the current "Credit based on runtime"?


I can only see credit, not # of results, so that's what I mean by output.

CreditNew provides lower credit for CPUs than the current system here so those CPU credits will be lowered too. But something will have to be done.
7) Message boards : Number crunching : GPU Error (Message 2262)
Posted 25 Mar 2019 by mmonnin
Post:
ok, see no quick way to install cuda 10.1 , version 9.1 is latest on Ubuntu 18.10.

ok, i'll try to install then Cuda 8 , very later this evening, maybe tomorrow and see.


Driver 418 is available if you add the nvidia repository.
8) Message boards : Number crunching : GPU Error (Message 2254)
Posted 25 Mar 2019 by mmonnin
Post:
Ha well it only ran for 1 second and credit is a function of time and FLOPs.
9) Message boards : News : GPU app - beta version for linux nvidia (Message 2251)
Posted 25 Mar 2019 by mmonnin
Post:
Yesterday was triple the output of the Formula BOINC Sprint's best day.
10) Message boards : Number crunching : GPU Error (Message 2240)
Posted 24 Mar 2019 by mmonnin
Post:
nope, only the task report says:


Now starting the targeted Martinet search:
Error code 700: an illegal memory access was encountered file polDiscTest_gpuCuda.cu line 415.


I had a couple of those as well. Lowered my OC a bit and it went away.
11) Message boards : News : GPU app - beta version for linux nvidia (Message 2223)
Posted 24 Mar 2019 by mmonnin
Post:
There is only once per day external update to FreeDC/BOINCSats but FB can read the team points every hour which worked for the Sprint. Marathon may just look once per day.
12) Message boards : News : GPU app - beta version for linux nvidia (Message 2221)
Posted 24 Mar 2019 by mmonnin
Post:
There is no way to separate it without reading every users tasks, which is impossible with hidden computers.
13) Message boards : News : GPU app - beta version for linux nvidia (Message 2219)
Posted 24 Mar 2019 by mmonnin
Post:
Just low to mid 70% GPU util so I'm running 2x. Now mid 90%.

<app_config>
  <app>
    <name>GetDecics</name>
     <gpu_versions>
      <gpu_usage>0.5</gpu_usage>
      <cpu_usage>0.1</cpu_usage>
     </gpu_versions>
  </app>
</app_config>


Interesting... I forgot you could control those things through the client. I just did the same thing. GPU is running a little hotter and my browser is a little sluggish... I might have to go back to running 1x.

So is 70% considered low? Should I try to optimize the utilization? During development, having too many blocks or threads per block caused crashes due to resource problems. I eventually settled on some values that seemed to work well, but maybe they are not optimal.

But I am pressing on with the OpenCL version first; I will worry about optimization later.


Hotter is to be expected. My Ti is in a dedicated system so slow browsing won't bother me.

I'd consider it low, so I added another task even at the expense of another lost CPU thread feeding the GPU exe.

Some other projects have options the user can configure, I'm guessing something like the blocks and threads per block that you mentioned. Collatz, Amicable and even SETI's user apps can have custom settings that can push GPUs harder. Amicable has a kernel size setting on BOINC's user preferences. Collatz and SETI can look at text files in the project folder.

Memory usage in Nvidia X server is showing about 1.5GB with 2 tasks so plenty of resources left. 567mb per task. 0-1% PCI-E usage. Just 115-118 watts via nvidia-smi on a 1080Ti. Same OC uses 185-195 watts with E@H.
14) Message boards : News : GPU app - beta version for linux nvidia (Message 2217)
Posted 23 Mar 2019 by mmonnin
Post:
Just low to mid 70% GPU util so I'm running 2x. Now mid 90%.

<app_config>
  <app>
    <name>GetDecics</name>
     <gpu_versions>
      <gpu_usage>0.5</gpu_usage>
      <cpu_usage>0.1</cpu_usage>
     </gpu_versions>
  </app>
</app_config>
15) Message boards : News : GPU app - beta version for linux nvidia (Message 2212)
Posted 23 Mar 2019 by mmonnin
Post:
First, YEAH!!!!! A new GPU project! Cool and thanks!!!!

Driver version 4xx.xx (at least for winblows) is the first driver to include support for the 2000 series RTX GPUs, and at least initially, it severely hampered the performance of the 10-series cards. I do not think many of your 10-series users will be willing to try a 4xx.xx driver version anytime soon.

EDIT: I decided to do my homework (AFTER I have already run my mouth :D ), and perhaps just the first few 4xx.xx drivers hurt the 10-series, I guess I will be updating my driver now to find out for myself. :)


I recall hearing that somewhere as well and a reason I installed just 396 on my 1080Ti. Pls report back if it's still the case.

Attempting to install 418 broke my Ubuntu install and I don't recall the exact OC clocks on the Ti but it's performing just as well on E@H with 418 vs 396 after the Ubuntu reinstall. The E@H frequency is constantly changing which changes the compute time slightly so if there is a driver difference it seems small with OpenCL.

Comparing the 11 tasks I received here at NF and with a random Ti at Collatz, the RAC is about 1/3 at Collatz which is known to be a very high RAC project. Once people realize the RAC possibilities here, expect more members to flock here. Especially with the limited # of GPU projects.
16) Message boards : Number crunching : GPU Error (Message 2207)
Posted 23 Mar 2019 by mmonnin
Post:
GTX 1060, only a couple of months old, running proprietary driver 390.116, the latest in the repositories. Thanks for your "insight".


You need at least 418, which are in the nvidia PPA.

sudo add-apt-repository ppa:graphics-drivers/ppa

update and it will be available in drivers.
17) Message boards : News : GPU app - beta version for linux nvidia (Message 2193)
Posted 22 Mar 2019 by mmonnin
Post:
No wonder I couldn't get anything with 396.51. I didn't even know NV drivers were up to 418.
18) Message boards : Number crunching : NF is this week's Formula BOINC's Sprint (Message 2184)
Posted 18 Mar 2019 by mmonnin
Post:
I have some information regarding server load. I think a good measure of that is how many WUs are processed per hour. Here is an average, broken down by day:
Before Sprint (3-13): 3571 WUs/hr
Thurs (3-14): 9329 WUs/hr
Fri (3-15): 5042 WUs/hr
Sat (3-16): 4311 WUs/hr
Sun (3-17): 6593 WUs/hr

On Friday, the current batch that had been running (14x11) ran dry and processing started on the next batch (14x12). The WUs in 14x12 average twice the run time as 14x11, so to compare apples to apples, I did a 2nd calculation where I double the WU rate for the 14x12 cases before adding to the 14x11 rate. This gives the following WU rates (this is what the numbers would have looked like if both batches had the same run times):
Thurs (3-14): 9329 WUs/hr
Fri (3-15): 6871 WUs/hr
Sat (3-16): 7930 WUs/hr
Sun (3-17): 10692 WUs/hr

I think the key point here is that the worst server load is at the start and end of the sprint. I'm not an expert on the strategies behind these competitions, but I assume it's due to bunkering up front and then holding back results until just before the finish. Anyways, the other point of interest is that the worst server load was about 3 times our typical load.


https://stats3.free-dc.org/stats.php?page=proj&proj=num

This shows over 20x normal load by credit at the peak and 9-10x the last two days. If your GPU version is 20x as fast on just a 1050 I would expect there to be much more 20x once a GPU version is released. More people will come to the project with GPUs to cash in on the higher credit. Assuming the same task is worth equal if ran on a GPU vs CPU.

19) Message boards : Number crunching : NF is this week's Formula BOINC's Sprint (Message 2168)
Posted 14 Mar 2019 by mmonnin
Post:
I saw ~380k tasks available a little after the event start and now its over 400k. I don't think FB has as high of participation than the Pentathlon even if there are more teams. Some projects were in their 1st year of being a BOINC project last calendar year and did not know about being listed for FB Sprint. Some tripled their normal daily output and their servers suffered.
20) Message boards : Number crunching : NF is this week's Formula BOINC's Sprint (Message 2164)
Posted 13 Mar 2019 by mmonnin
Post:
FYI, there will probably be some increased server load.
http://formula-boinc.org/sprint.py?lang=&year=2019


Next 20


Main page · Your account · Message boards


Copyright © 2024 Arizona State University