Posts by petri33

1) Message boards : Number crunching : Linux CUDA 'Special' App finally available, featuring Low CPU use (Message 1885347)
Posted 6 hours ago by Profile petri33Project Donor
Post:
Try pfb = 32, or try taking it out. Those are definitely high numbers.
What is your command line?
Clocking?


I don't overclock my cards, it's just stock timings.
I tried pfb 32 and it seems to be worse. The command line has only the pfb setting.
I'll try backing it off to pfb 8, or just take it out like you suggest.


Set GPU fan to 100% and see if that helps.
2) Message boards : Number crunching : SETI/BOINC Milestones [ v2.0 ] - XXIX (Message 1884976)
Posted 2 days ago by Profile petri33Project Donor
Post:
Congratulations to all those on their recent milestones.

I myself have reached a RAC of 121K.

Cheers.


You seem to take good care of your computers. A nice and steady output!
3) Message boards : Number crunching : SETI/BOINC Milestones [ v2.0 ] - XXIX (Message 1884964)
Posted 2 days ago by Profile petri33Project Donor
Post:
New record for me also !

RAC have reached 2.4 millions ...

But you only need 7 Petri computer's like to beat me ! ;)


WOW!


How scientist explain the famous 'wow' signal :
http://earthsky.org/space/wow-signal-explained-comets-antonio-paris


An interesting article.

But how would you explain a 'negative Gaussian'. Ten or fifteen yesars ago I had a Seti screen saver running and there was a dip in the random noise. The dip moved to the right and up in the picture on the screen. Something blocked the signal or background noise and ...
4) Message boards : Number crunching : SETI/BOINC Milestones [ v2.0 ] - XXIX (Message 1884958)
Posted 2 days ago by Profile petri33Project Donor
Post:
New record for me also !

RAC have reached 2.4 millions ...

But you only need 7 Petri computer's like to beat me ! ;)


WOW!
5) Message boards : Number crunching : SETI/BOINC Milestones [ v2.0 ] - XXIX (Message 1884924)
Posted 2 days ago by Profile petri33Project Donor
Post:
So close to hitting a new record ..

Average credit 359,990.38

EDIT: Looks like it just did it
Posts: 1409
Credit: 238,199,020
RAC: 360,408
6) Message boards : Number crunching : Linux CUDA 'Special' App finally available, featuring Low CPU use (Message 1884798)
Posted 3 days ago by Profile petri33Project Donor
Post:
The Downloads page has a similar bug.

Sort rows for your preferred column. And BANG the next update trashes the sort order.

I do not care, but I know it is there.

The BUG.

--
petri33
7) Message boards : Number crunching : GUPPI Rescheduler for Linux and Windows - Move GUPPI work to CPU and non-GUPPI to GPU (Message 1883675)
Posted 8 days ago by Profile petri33Project Donor
Post:
Hi,

The BLC/GBT/Guppi work units complete in under 2 or 3 minutes on NVIDIA with the 1080 and the 1080Ti running 'special' app. The old CUDA 50 app is slow. The same applies to Arecibo vlars. So I don't reschedule.

Petri
8) Message boards : Number crunching : Panic Mode On (107) Server Problems? (Message 1883438)
Posted 9 days ago by Profile petri33Project Donor
Post:
Hi,

I had my GPUs running out of work on Tuesdays and on other weekdays too. 'Not sending work' - 'No tasks available' bla bla blah.

I took a look at the boinc client source code and found that there was a had coded limit of 1000 WUs per host. I removed that.
Next I made the client to tell the servers that I have four times the GPUs I actually have. Now I have 100 CPU + 4 * 4*100 GPU tasks in the cache constantly.

No problems any more. The cache survives most of the Tuesday outage(s) and the servers refusing to send Arecibo vlars to NVIDIA (special app) hosts.


. . I wish I had your skills, it would be nice to feel buffered against maintenance days.

Stephen

:)

I wish the project was able to support such caches for everybody and cheating the code was not required.

Meow.


I think the same. The project should make the cache relative to RAC. My cache lasts now 1h 20 min for shorties and 7h for guppi work. If I had only one CPU and no GPUs I should have a cache of max 2 WUs per real core.
9) Message boards : Number crunching : So who is going to be a guinea-pig this time?? (Message 1883431)
Posted 9 days ago by Profile petri33Project Donor
Post:
A Good point Zalster. Thanks.
EDIT: and I stated that in the first p.s.
10) Message boards : Number crunching : Welcome to the 18 Year Club! (Message 1883430)
Posted 9 days ago by Profile petri33Project Donor
Post:
I'm wearing the 15yr T-shirt at the moment. I Joined in 99 and lost my original account. That is the reason I look a bit younger by the numbers.
11) Message boards : Number crunching : Panic Mode On (107) Server Problems? (Message 1883429)
Posted 9 days ago by Profile petri33Project Donor
Post:
Hi,

I had my GPUs running out of work on Tuesdays and on other weekdays too. 'Not sending work' - 'No tasks available' bla bla blah.

I took a look at the boinc client source code and found that there was a had coded limit of 1000 WUs per host. I removed that.
Next I made the client to tell the servers that I have four times the GPUs I actually have. Now I have 100 CPU + 4 * 4*100 GPU tasks in the cache constantly.

No problems any more. The cache survives most of the Tuesday outage(s) and the servers refusing to send Arecibo vlars to NVIDIA (special app) hosts.
12) Message boards : Number crunching : So who is going to be a guinea-pig this time?? (Message 1883424)
Posted 9 days ago by Profile petri33Project Donor
Post:
Apologies to the OP for the hijacking of this thread...

It's always been a buy beware sort of thing.

EVGA tends to be pretty good with RMA items so there is that. But they don't have a true 4 x16 pci board. But they seems to be the best (at least in my opinion)

That being said, I run ASUS boards (what??!!) Yes, you read that right. I do this because they have the X99e WS boards. The closest we have ever gotten to a true 4 X16 PCIe board. They have a new board out X99e 10ws which replaces their X99e 6ws that replaced the x99e ws usb3.1 that replaced.... you get the picture....

The problem with ASUS is customer service. It's horrible... Of the 5 boards I have bought from ASUS, 1 was DOA out of the box, the second died within a few months. When they work, they work great. But they don't always work..Really unsettling when you look at the cost of these boards.

After EVGA, ASUS, ASRock would be my 3rd choice. Gigabyte boards are ok, but not in the quality (in my opinion as the first 3)


I'm a happy runner of ASUS X99-E-WS and a 6/12 core Intel chip. The CPU runs 6 MB tasks and the 6 additional HT cores serve the GPUs that do the main work.

p.s. To run seti you do not need to have a full blown XXX8-16 * 4 PCIe configuration. The PCIe bus usage is about 1% - 4%. The SoG and my 'special' do their best at hiding the memory transfer latencies, i.e. they do other work a*) when the GPU to RAM memory transfer is occuring. The b*) RAM to GPU mem transfer is a one time event. *a) Happens often. It is hidden by launching work to the GPU before the memory transfer and doing CPU work when waiting for the memory transfer to complete. *b) Is run one time only.

p.p.s. Save your money to best available GPUs and the cheapest 4 core procesor and a MoBo with four 2x wide slots. Add a decent cooling and make sure your cards run near the performance limits set by the GPU hw and drivers. On NVIDIA that means overclocking the idle state since the GPU calculation work load runs a couple of performance states lower.

p.p.p.s I have had my ASUS MoBo replaced twice. The earlier models did not have (or I did not connect) the auxiliary PCIe power connector and I had the burnt up pin 22/23 or something problem.

Petri
13) Message boards : Number crunching : Linux CUDA 'Special' App finally available, featuring Low CPU use (Message 1881779)
Posted 18 days ago by Profile petri33Project Donor
Post:
Thanks Petri, I mainly asked because of your post of the fix for the outdated error messages to V7. I'll monitor the thread looking for the latest when you think it is ready for release.

Thanks for the explanation about the -pfb setting. I went back to the docs for the x41Z app and refreshed my memory. I used to run -pfb=16 for the CUDA50 app on the 970's. I'll think I'll give that a try for the 970s on the special app.


I hope you all can live with the message having a 'typo'. An to the -pfb N, I think you'll notice an improvement in speed.
And for those who have blocking sync enabled I recommend disabling it if you are after the ultimate speed.

Petri
14) Message boards : Number crunching : Linux CUDA 'Special' App finally available, featuring Low CPU use (Message 1881723)
Posted 19 days ago by Profile petri33Project Donor
Post:
Petri, do you have your latest special app hosted somewhere?

What is a sensible -pfb value for reference GTX970s? I have a command line currently at -autotune and -nobs. I searched in the top 100 computers for other 970 users and I only found Stephen who is using the same defaults and Mr. Kevvy who is using autotune and -pfb=32 setting. I think the -pfp value correlates directly to the number of compute units. I am unclear on what the -pfb parameter sets up in the card. A short explanation would be appreciated. I looked in the docs and it doesn't really explain what the parameters accomplish, just their syntax.


Hi,

I do not have the latest (cuda9) version anywhere yet. It'll come in a week or so. I'll post here the links.

The -pfb can have 8, 16 or 32 on modern HW (7xx, 9xx & 10xx). It is short for Pulse Find Blocks per sm. The default is 4 if I remember correctly. Setting it to 8 or higher speeds up the computation. It is not autotuned yet. I use -pfb 32 on my computer.

Petri
15) Message boards : Number crunching : Linux CUDA 'Special' App finally available, featuring Low CPU use (Message 1881670)
Posted 19 days ago by Profile petri33Project Donor
Post:
Hi,

detected ... v7 ...

That is just an old message and I have changed that to say now v8.

If you have trouble with restarted tasks you could try setting in BOINC manager 'write to disk at most every xxx seconds' to such a high number that a longest running GPU task does not do checkpoints.
My computer finishes GPU tasks in under 300 seconds so I can set 300 to the BOINC.

Petri
16) Message boards : Number crunching : Linux CUDA 'Special' App finally available, featuring Low CPU use (Message 1880774)
Posted 24 days ago by Profile petri33Project Donor
Post:
I'm going to look at the code when home.
17) Message boards : Number crunching : Setting up a Linux machine to crunch CUDA80 for Windows users (Message 1880544)
Posted 25 days ago by Profile petri33Project Donor
Post:
Hi,
This is the best state to run. P-state 0 (nvidia-smi reports it P0 in the picture) . i.e. perf level 3 (as reported by nvidia settings). You have everything ok. I can not get my cards to run at anything but P2 (nvidia-smi says).

And to the other issue 'running a startup script as root' you find help with a little googling: linux startup script root

Petri
18) Message boards : Number crunching : Setting up a Linux machine to crunch CUDA80 for Windows users (Message 1880366)
Posted 26 days ago by Profile petri33Project Donor
Post:
From command prompt run as root or sudo
nvidia-smi -acp UNRESTRICTED

After that all users can set the -ac parameter.
19) Message boards : Number crunching : Setting up a Linux machine to crunch CUDA80 for Windows users (Message 1880359)
Posted 26 days ago by Profile petri33Project Donor
Post:
Hi,

What value did you set for the coolbits?
Mine is set with:
nvidia-xconfig --enable-all-gpus --cool-bits=28
and reboot.

Look 'Enabling overclocking' from this link https://wiki.archlinux.org/index.php/NVIDIA/Tips_and_tricks for more options, like setting max perf level (Option "RegistryDwords" "PerfLevelSrc=0x2222").
20) Message boards : Number crunching : SETI/BOINC Milestones [ v2.0 ] - XXIX (Message 1880313)
Posted 27 days ago by Profile petri33Project Donor
Post:
Hi,

Congrats! to everyone.

I'm celebrating passing +340 000 RAC on a single computer. [EDIT: damn outage]

--
Petri

p.s. Looked at the GPU statistics and is seems someone is running a linux version of a MB sw with a TITAN Xp and it is doing well: 1.00 vs 0.84 against my (1x1080Ti + 3x1080) average. 16 %units difference. Scaling? (20x3+28)/4 ... 22 vs 30 SM units, hmm. I think it is time for the owner to optimize cooling, cool-bits, clocks etc. The scaling should be nearly linear. I should have .7333...


Next 20


 
©2017 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.