PDA

View Full Version : DNet



Paladin
01-06-2009, 07:09 AM
The 508 client (Win32) is 30% faster at OGR than the previous 507.

Also there's a separate Win32 CUDA client, but it's only for RC5.

ftp://ftp.distributed.net/pub/dcti/prerelease/

[XC] riptide
01-06-2009, 08:53 AM
How are we doing on 64bit support for speed? IIRC when I ran this a few months ago I was getting shafted like a pig in a spear shop. :D

Paladin
01-07-2009, 08:11 AM
Still no Win64 client. No 508 non-CUDA version available for Linux yet.
Here's the current comparison state:


Sys#1 (WinXP64, 4GB, Q6600@3200mhz)
v.507 OGR 31,637,000
RC5 12,061,000
v.508 OGR 41,349,000*
RC5 12,061,000

Sys#2 (Ubuntu64 8.10, 2GB, E4300@2400mhz)
v.507-64 OGR 31,267,000*
RC5 8,067,000
v.507-32 OGR 24,205,000
RC5 9,035,000*



So for OGR, new versions=better & 64bit linux=more efficient.
For RC5, 32bit clients=better & close call if linux>windows @ same clock.

glennpat
01-08-2009, 02:32 PM
That is really a nice increase. Someday soon I'll have to get a card that supports CUDA. I have my PS3 running OGR-26 with the new level. I checked it yesterday and for some reason it was running slow. I rebooted it and it is back to normal.

Paladin
01-22-2009, 06:56 AM
There's now a beta/rc for Win64 509 client. Also, I tested the Win32-CUDA 508 client and instead of 12M keys/sec on RC5 it got 290M!

[XC] riptide
01-22-2009, 09:47 AM
This might be worth trying then^^

EDIT: MAJOR speed imrpovement on Server 08 64bit. Major. I think I'm up from about 60m nodes/s to about 100million. Now time to try teh CUDA... if it'll work with G80.

EDIT2: OK so I think CUDA version is not 64bit compatibale? CLient just appears are disapears fast.

Paladin
01-22-2009, 06:27 PM
riptide;3596505']EDIT2: OK so I think CUDA version is not 64bit compatibale? CLient just appears are disapears fast.

You have to set the system date to pre 19-Jan-09. If you read the log file it says akin to "this time-limited beta has expired". The servers still send/receive data from it.

[XC] riptide
01-22-2009, 10:45 PM
Ya.. i don't see a logfile... but in any case, I can't set the date behind :( Main rig and such.

[XC] riptide
01-22-2009, 10:47 PM
You have to set the system date to pre 19-Jan-09. If you read the log file it says akin to "this time-limited beta has expired". The servers still send/receive data from it.

I wonder would virtual box do anything for me? Theres a switch in the GUI that alludes to allowing direct access tot eh GFX? hmmm. Must look. I wonder how much Vram I need?


^^^ Scratch that. Stupid idea. So ya I turned back the clock for a few minutes on my machine. It ran. It exceed 400million nodes per second... then I changed the time back... and gone! Anyway to hack teh expiry on the beta?

Paladin
01-23-2009, 01:33 PM
You might try RunAsDate which tricks the time on a program-by-program basis.
http://www.nirsoft.net/utils/run_as_date.html

I tried a simple edit to change the check date from 2009 to 2010, but there is a CRC check somewhere in there to prevent people editing the program. Complex assembly language is too low level for me.

[XC] riptide
01-23-2009, 02:20 PM
You might try RunAsDate which tricks the time on a program-by-program basis.
http://www.nirsoft.net/utils/run_as_date.html



It works! :) Thanks a bunch!

[XC] riptide
01-23-2009, 07:44 PM
Furthermore... is there anyway to to have it NOT pause the work while it flushes teh buffers?

[XC] riptide
01-23-2009, 08:11 PM
I'm using a 8800GTX. I found that the auto selected core picked Core#1 which gave me about 348Mkeys/Sec and about 20% of a single CPU core usage. However... when I ran the benchmark several times, on all cores, I found that Core#9 was 353Mkeys/sec.! And.... only used 1-5% CPU!!!!!! This increased Mkeys/sec is backed up the console readouts aswell as its normally running.

[XC] riptide
01-24-2009, 04:35 PM
Results are in. http://stats.distributed.net/team/tmember.php?project_id=8&team=31529

[XC] riptide
01-25-2009, 01:24 AM
Further tweaking..... I selected the option to have the packets in memory (no Disk I/O), which connects to the server to upload every time. It wasn't what i wanted... BUT it results in another fairly significant gain in rate. It appears that the Disk/File handling of the client is flawed. I get better rates because completed WU's upload to the server faster than written to the out buffer on disk!!!! Thats my take on it anyway. Furthermore.... I downloaded one of the personal proxies from their webpage. Acts a bit like SobQueue. So the aforementioned upload to server after every WU is now uploaded to a proxy running on my PC! Decreasing upload time even further...hence increasing the performance of the client.! I reckon that all the above tweaks increased performance by about 10%.

Paladin
01-26-2009, 06:22 PM
Dynamic in buffer set to 24 hours, disable check-pointing, no disk i/o, and finally disable keyserver connectivity seemed to be the only way to get full speed. With these settings you can still initiate a manual fetch/flush from the tray icon and no worries if it runs out between times you flush since it'll just generate it's own random work.

Haven't been able to get the normal client running on the other 3 cores at the same time though.

Paladin
01-26-2009, 07:31 PM
New version of both the Win64 and Win32-CUDA clients posted on the FTP site today. Didn't notice a diff in the regular one, but the CUDA is a little faster plus grabs larger work units so it's not doing i/o all the time. :up:

[XC] riptide
01-26-2009, 10:32 PM
New version of both the Win64 and Win32-CUDA clients posted on the FTP site today. Didn't notice a diff in the regular one, but the CUDA is a little faster plus grabs larger work units so it's not doing i/o all the time. :up:

Cool! Must grab that now!

PS: I'm right behind you... :hehe: Hurry TF up. :D

[XC] riptide
01-27-2009, 12:44 AM
New client is working well. Getting 382-384Mkeys/sec

skycrane
01-27-2009, 05:45 AM
good gaud riptide, i can remember back in the day.... when you were lucky to get a million, thats 1,000,000 keys/sec i think i might just have to get another gpu or 2 :)

[XC] riptide
01-27-2009, 06:07 AM
good gaud riptide, i can remember back in the day.... when you were lucky to get a million, thats 1,000,000 keys/sec i think i might just have to get another gpu or 2 :)

Ya. Its amazing really. Wouldn't mind seeing you on teh team also. I might have to get another GFX. Have to keep teh folding up too.

skycrane
01-27-2009, 07:29 AM
yea, i might be able to bring a few keys a day to our output :)

do you know how well it works on a xp64 ?? and what sort of gpu do i need? ive got 2 88gt, 2 gso's, and a 98gx2. witch ones do you think will work the best ?

[XC] riptide
01-27-2009, 12:25 PM
yea, i might be able to bring a few keys a day to our output :)

do you know how well it works on a xp64 ?? and what sort of gpu do i need? ive got 2 88gt, 2 gso's, and a 98gx2. witch ones do you think will work the best ?

GPU easy question. Highest shader count. That'll be teh 98GX2 at 128 x 2 I expect. I presume the client will recognise the X2 cores?

Works pretty well on my 64bit Server.

skycrane
01-27-2009, 03:10 PM
ill keep the gx2 on folding, the 88gt ill put on here, and im not sure what to do withthe gso's, maybe keep them on gpugird... i think im going to need more gpus, mobos, and cpus..... the wifey is going to kill me... lol

skycrane
01-28-2009, 06:06 AM
how do you know if there is any work being done on your gpus ??

ive got a few clients running on a few cpus, and one gpu, how can you tell if it is working on the gpu ?

[XC] riptide
01-28-2009, 07:17 AM
how do you know if there is any work being done on your gpus ??

ive got a few clients running on a few cpus, and one gpu, how can you tell if it is working on the gpu ?

Cos it says it is!

skycrane
01-28-2009, 07:48 AM
lol, thanks.

umm then i would have to say a big definate NO my lappy is putting up a respectiabe 50 mill :) but my others are no where near that, what are you running to get the 450 mill? does the gpu use a full core to make that ?

[XC] riptide
01-28-2009, 07:57 AM
lol, thanks.

umm then i would have to say a big definate NO my lappy is putting up a respectiabe 50 mill :) but my others are no where near that, what are you running to get the 450 mill? does the gpu use a full core to make that ?

Well its hard for me to tell how many shaders cores the program is using.... but my temps are as high as if I was folding.

Heres another thing, make sure that the dnet process in tasnkmanger is set 'above normal'., if you have other stuff runnning aswell. It'll still only use a few % of CPU. You have the correct client ya? When you start it up does it tell you its detected the GPU?

skycrane
01-28-2009, 08:05 AM
no ill have to look at that when i get back home in a few hrs

ok, so you can tell by the temps. ill have to check that

no i dont think i have the correct client then, just the regular client. ive been looking all over the site, and i cant find the gpu one.

[XC] riptide
01-28-2009, 08:11 AM
no ill have to look at that when i get back home in a few hrs

ok, so you can tell by the temps. ill have to check that

no i dont think i have the correct client then, just the regular client. ive been looking all over the site, and i cant find the gpu one.

http://www.distributed.net/download/prerelease.php

2nd down from top.

skycrane
01-28-2009, 09:10 AM
thanks, ill give them a whirl when i get home.

do you know if it will work in xp64 ? as long as i have cuda 2.0 ?? it says its for x86/cuda but nothing about 64bit.

[XC] riptide
01-28-2009, 09:35 AM
Well it works in my 64bit OS. So ya!

skycrane
01-28-2009, 02:34 PM
well this sucks... its only putting out about 300 mill on my 9800gx2 :( lol

[XC] riptide
01-28-2009, 02:42 PM
well this sucks... its only putting out about 300 mill on my 9800gx2 :( lol

Well thats not good. Is it identifying the 2nd core? Show me the console logs.

BTW!!! Some guy reported on Free-DC that these new WU's that have 64 packets are only now giving 1point each? WTF???? Thats a major problem. I'm going back to the previous version.

http://www.free-dc.org/forum/showthread.php?p=132118#post132118

skycrane
01-28-2009, 02:45 PM
no i just found it, ive got gpu grid on one core, dnet on the other, and 2 instances of folding all going on at once... :) maybe that had something to do with it lol

[XC] riptide
01-28-2009, 03:02 PM
no i just found it, ive got gpu grid on one core, dnet on the other, and 2 instances of folding all going on at once... :) maybe that had something to do with it lol

rofl. you had GPU folding aswell? hahah. I'm gonna put your name in that pic 2 posts above and make it your avatar! :ROTF::ROTF::clap:

skycrane
01-28-2009, 04:45 PM
well why not ??? id figure id make it work for me :)

[XC] riptide
01-29-2009, 12:24 AM
well why not ??? id figure id make it work for me :)

Theres no problem... just don't expect huge scores. What you should really do it check out the what the scores are with each poroject on the card. Then see if what they are with both, and see if there is an advantage to be had. But I doubt it.

[XC] riptide
01-30-2009, 05:06 AM
Stats are back and corrected. Check them out. Nice entrance skycrane.

http://stats.free-dc.org/stats.php?page=team&proj=rc572&team=XtremeSystems

skycrane
01-30-2009, 06:10 AM
thanks riptide, i dont think ill get to the #1 in this one... :( lol

[XC] riptide
02-02-2009, 04:39 AM
New Beta out. Last one expired. beta3 for Cuda.

[XC] riptide
02-27-2009, 04:46 PM
http://www.distributed.net/download/prerelease.php

New Betas, Beta4 for Cuda.