Posts by Paul D. Buck
log in
1) Message boards : Number crunching : Feed back wanted (Message 11390)
Posted 2535 days ago by Profile Paul D. Buck
Just a side note, cheap MB are not always so cheap. Frys regularly had "specials" where they would sell a MB and CPU set for roughly what the CPU used to sell for off sale. I bought quite a few of those but once replaced the el cheapo MB with an Intel or other more upper end MB and with no other changes, same memory, same CPU, etc. ... saw a 20% improvement in performance. Not saying that this will be true in all cases, but, one thing to keep in mind ...
2) Message boards : Number crunching : Feed back wanted (Message 11362)
Posted 2537 days ago by Profile Paul D. Buck
Yea with the sandy..but the fix is a hardware issue, so untill they release the fixed stuff I would consider them. The GPU will be faster than the CPU in pretty much any case, so the CPU is not what I am really after. The GPU's are the work horses..wish I could afford the high end GPU's.


So why not just get a cheap dual core and a fancy gpu? Why got the 6 core AMD and the i7 at all if what you really want is a fancy gpu. A dual core cpu can run a fancy with no problem, and it is MUCH cheaper. Heck you can even them used or get them for free at your local dump. You can build a nice dual core with a cheap hard drive, you do not need more than a 40 gb one if you all you do is crunch with it, and a dvd drive to load Windows then remove it wen you are done, and some memory and a fan! That is how I have so many pc's up and running myself. I take other people old hardware as payment when I work on their pc's and make crunchers out of them. I have 15 pc's running now with 6 more on the floor ready to go because I do not have the electric power to run them, I have maxed out my available circuits already!! Heck I have 4 usable gpu's on the shelf with no pc's to put them in!!

There comes a point where 20 dual cores simply use more power because of the PSU efficiency and the overhead of multiple MB and what not so that 3 i7 boxes will outperform the larger farm and will use less power. The flip side is that you have fewer slots for GPUs, but, as you have noted we all hit some ceiling with power draw and have to draw the line... the question is where and how ...

In my hay-day I once had 9 machines running but now I am down to 5 and looking to drop one when I can so I would only have 4 ... though, the number of CPUs I would have would stay the same though I would see the loss of one GPU slot ... no matter, the electric bill is far too high anyway ...
3) Message boards : Number crunching : new problem (Message 11334)
Posted 2539 days ago by Profile Paul D. Buck
YOu say that with BOINC off yuo have 79% use of the CPU? Use task manager to find the rnning process. Then look up the process name on the web to find out what it is...
4) Message boards : Number crunching : Have a problem with raising nr of timed out tasks? (Message 11317)
Posted 2539 days ago by Profile Paul D. Buck
Be careful with 6.12.12 though. It is still in testing so there could still be some bugs in it. The latest BOINC version that is cleared for everyone is 6.10.58.

There's never been a BOINC version yet that didn't have bugs, but have been running 6.12.12 on 10 boxes since it was released and it works far better than 6.10.58. Other people seem to be having the same experience with it. 6.10.58 and earlier BOINC clients had severe GPU scheduling problems.

THat is because UCB went to strict FIFO GPU scheduling to solve a task scheduling problem where the system would start task after task after task and then suspend the task to select a different one. It turned out there was a bug causing this that was found but the FIFO scheduling was not removed... even though it caused serious problems like has been noted here (and elsewhere) ...

It is now out and 6.12.12 or the newer .13 seem to work well even with multiple projects thjough I am not convinced it is properly enforcing share still ... but that is another issue for another day ...
5) Message boards : Number crunching : new problem (Message 11316)
Posted 2539 days ago by Profile Paul D. Buck
if you have been returning comp errors then the server decrements the number of tasks it will send you till you only get one a day. If you are getting lots of comp errors the first thing to do is to restart your system. If the GPU gets snarfled, ALL tasks will fail... the only way to reset the GPU is to restart. It is not always obvious when the GPU is snarrfled up because the display may still be fine...

If that does not help, then, your card may be over heating or has gone bad ... others will prolly jump in with other thoughts ...
6) Message boards : Number crunching : BAM question (Message 11315)
Posted 2539 days ago by Profile Paul D. Buck
If I recall correctl that is the number of time your host and BAM have talked ... at least once per day... more if you have been forcing connections to set things up...
7) Message boards : Number crunching : No work for HD6950? (Message 11292)
Posted 2541 days ago by Profile Paul D. Buck
Detach and attach seems to have unstuck it.


I had one box that I wanted to configure for only mini-collatz and it kept ignoring the project prefs and downloading the normal collatz WUs. I had to detach and re-attach to get it to work properly.

Collatz on the 6950 is slower than the 5870 while MW is the other way around.


MW uses double precision floats whereas Collatz uses integers. Collatz uses a lot of texture memory lookups so texture access and memory speed play a big part as well.

Actually I am guessing that it is the number of calculation units in the case of Collatz being more critical ... tough I could very well be wrong. I would have preferred to get a 6970 the two the store had on Saturday where were were there were already gone (sob). Well, I am guessing that my path ahead would be to get my first 6970 then the second to make a pair, then I will get a second 6950 to parir and replace the 4870s I have ... something like that. Though I will admit that the performance is not as bad as I reared and the cost is about $50 less so a suite of 4 of these would give me 2 5870s, 4 6950s and an odd man out ... though I am hoping to squeeze in my first 6 core this year as well ... so maybe that will retire the quad core and I will only have 6 GPUs .... unless I get a MB with 3 GPU slots ...

Probably not ... I do need to cut back on the power usage ...
8) Message boards : Number crunching : No work for HD6950? (Message 11286)
Posted 2542 days ago by Profile Paul D. Buck
Detach and attach seems to have unstuck it.

Interestingly enough, though I know that this has been discussed by the HW gurus... Collatz on the 6950 is slower than the 5870 while MW is the other way around. Have not tried DNETC and PG on both systems.
9) Message boards : Number crunching : No work for HD6950? (Message 11283)
Posted 2543 days ago by Profile Paul D. Buck
I only have the one venue and my other ATI systems are pulling work... mini colatz is off in that I dont do collatz on the CPU at all (what a waste that would be... I do other CPU only projects on the cpus)...
10) Message boards : Number crunching : No work for HD6950? (Message 11277)
Posted 2543 days ago by Profile Paul D. Buck
I just got a new ATI 6950 and it pulls work from MW and PG, but I cannot seem to get Collatz work for it... or is it just me?

I reset debts this am and also raised resource share up to 100 so I should have as much collatz on hand as MW ... but I get nada ...

BOINC Ver is 6.12.12 if that matters, 64 BIt windows (W01 in my computer list)...
11) Message boards : Number crunching : Minor Update for the new AMD Radeons/FireGL/FireStream needed (Message 11050)
Posted 2558 days ago by Profile Paul D. Buck
And besides that, someone over at berkeley should think about exporting the ATI/AMD GPU identification to a dll...that would make more sense than hardcoding it into the client source....

Don't you mean a TXT file?

THen all you would have to do is update the text and away you go ...
12) Message boards : Number crunching : MAC OS X 32-bit CUDA App (Message 10017)
Posted 2649 days ago by Profile Paul D. Buck
I posted my woes with the latest version ... not sure if it were me or the tasks or what. All I know is that at the end of the day I was getting the out of memory error, had other issues so took care of those (failing disk drives) and when I came back decided to upgrade to the latest CUDA version and it does not seem to work... no nibbles or suggestions yet so not sure what to do ...
13) Message boards : News : CUDA 3.1 Application Available for OS/X (Message 9989)
Posted 2653 days ago by Profile Paul D. Buck
Well, not sure it is working ...

<core_client_version>6.10.56</core_client_version>
<![CDATA[
<message>
process exited with code 49 (0x31, -207)
</message>
<stderr_txt>
Collatz CUDA v2.04 OSX CUDA 3.1
Start 2365811052613944977768
Checking 206158430208 numbers
Calculate 256 items per loop
Loop 32 times per reduction
Sleep 1 ms while waiting
Device GeForce 8800 GT
Memory 511 MB
Error in initial cudaThreadSynchronize: unspecified driver error
Error in cudaUnbindTexture: unspecified driver error
Error in cudaFree(d_outp): unspecified driver error
Error in cudaFree(d_inp): unspecified driver error
Error in cudaFree(d_red): unspecified driver error
Error in cudaFreeHost(h_red): unspecified driver error
Error in cudaStreamDestroy(stream): unspecified driver error
Error in cudaEventDestroy(start): unspecified driver error
Error in cudaEventDestroy(stop): unspecified driver error
22:40:57 (199): called boinc_finish

</stderr_txt>
]]>

I am using the latest driver and cuda kit... sadly, most of the instructions in the read me did not seem to be applicable in the sense that I do not have a lib64 directory and have no idea where the referenced files come from to copy to this non-existant directory.

So, I don't know if it is because I am using 3.2, don't have it installed correctly for some other reason, or the extra instructions are needed and I did not fulfill them and that is the reason. I had run off a bunch of PG tasks, so, CUDA is working somewhat... I guess I will have to re-enable that so I get some work done

{Add}
Note that the PG tasks are suggesting that they are running under CUDA 3.1 successfully for what that is worth. See task for an example. As always my computers are open on all projects so, suggestions are welcome...
14) Message boards : Number crunching : No WUs? (Message 9353)
Posted 2704 days ago by Profile Paul D. Buck
Question abou LTD:

8/25/2010 8:58:44 AM AQUA@home [wfd] overall LTD -461407.97
8/25/2010 8:58:44 AM Collatz Conjecture [wfd] overall LTD -6216315.67
8/25/2010 8:58:44 AM DNETC@HOME [wfd] overall LTD -1187355.19

Why all LTD is "-"?

STD is normalized with upper bound of 0 (or the lower bound is 0, I can never remember) ... stupidly, LTD is normalized about 0 which means as one debt rises above, the other lowers (assuming only two) by the same amount.

You can set an option in the CC Config to zero debts (and it is not a bad idea to do so every month or so, more often if needed) because UCB has never really gotten work fetch working correctly (associated modules RR SIM and Resource Scheduler are also bent). The few of us that bother reporting issues have gotten so tired of the lack of response to the persistent issues that we rarely bother to report issues any more. If you are interested in the issues you can go and get the archives of the posted e-mails on the UCB site ...
15) Message boards : Number crunching : MAC OS X 32-bit CUDA App (Message 9283)
Posted 2710 days ago by Profile Paul D. Buck
There is a video update as part of the OS update that is suggested to install. I made the mistake of doing so... from then on I could not get 3.1.10 CUDA driver or 3.1 Toolkit to install and work with Collatz. Installing 3.0.14 Driver and 3.0 Toolkit seems to be working now... Though I will not know for sure for some time to come (a couple of hours at least when I can see several tasks complete without failure) ...

So, be warned ...
16) Message boards : Number crunching : getting radeon 4290 and gtx 260 to crunch at the same time (Message 8173)
Posted 2791 days ago by Profile Paul D. Buck
Thanks that worked and now both are crunching.

Glad you are going again ... :)
17) Message boards : Number crunching : getting radeon 4290 and gtx 260 to crunch at the same time (Message 8165)
Posted 2791 days ago by Profile Paul D. Buck
I'm running Windows 7 64 bit edition.

TO use the GPUs you cannot have it installed as a service.

Even with only a few tasks on hand, sometimes BOINC takes a long time to start up and no one quite seems to know why. There are times when it seems to me to be a minute or so with the "Connecting..." dialog box sitting there ...

So, reinstall, NOT as a service and you should get the GPU back ... one caution, the ATI and Nvidia control panels are rumored to cause issues with the other guy's card so be advised it is probably not a good idea to use them ...
18) Message boards : Number crunching : Long "to completition" time => running in high priority (Message 8132)
Posted 2793 days ago by Profile Paul D. Buck
Yes. Also 1 GPU task is running parallel on the same machine.

Ok, you may be running into the mixed-mode DCF problem ... right now the DCF is on a per-project basis ... it should be on an application version basis ... until that is done you may see things like this ... if you make the machine purely CPU side the DCF should correct and give a good number ... with the GPU in there you have the CPU side make a correction, then the GPU side makes a counter correction and ... the estimates for one or the other are going to be off ...

One of the reasons I don't mix GPU projects up by running them on the CPU side as well ...

{eidt}
High priority mode only means BOINC is going to favor running that task over others ... it does not actually change system priorities
19) Message boards : Number crunching : Long "to completition" time => running in high priority (Message 8129)
Posted 2793 days ago by Profile Paul D. Buck
Hello,
I have problem with collatz conjecture work units. Every unit which runs on CPU (not GPU) shows very long predicted time to completiton. When it is started is shows about 120 hours (after 2,5 hours from beginning it still shows about 80 hours). In fact it usually takes about 12 hours to complete. Because of this every WU from collatz project runs in high priority. Is it typical situation?
Computer configuration: Phenom X4 955, 4GB DDR3, Win7 32bit.

Are you also running GPU tasks on this same machine?
20) Message boards : Number crunching : CUDA tasks erorr out after suspend / resume (Message 8106)
Posted 2794 days ago by Profile Paul D. Buck
There was a suspend resume thing in the BOINC client IIRC, try the latest version 6.10.56...


Next 20

Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.