Welcome to MilkyWay@home

Posts by JerWA

21) Message boards : Number crunching : ATI Radeon HD5970 performance on boinc (Message 33438)
Posted 21 Nov 2009 by ProfileJerWA
Hmmm ....... the GFLOPS count is pretty much halfway between a 5850 @ 2088 GFLOPS and a 5870 @ 2720 GFLOPS. Good use of a single PCIe slot and (in the UK) makes good financial sense, but it'll be interesting to see how the card crunches.

Which makes perfect sense really, since it's got all the 5870 texture units, but the 5850s clock speeds.

I'm more interested in which brand it is and what (if any) utility software they included, given that the card is clocked to run in a 300w envelope but is sufficiently cooled to run 400w power. Lots of review sites showing clock and voltage tweaks (unlocked from factory). The main "gotcha" is that you can push power draw beyond the PCIe specs, meaning you could theoretically do damage if the motherboard and PSU aren't up to the task.

AnandTech has a pretty good article about it, and what to look for, if you're brave enough to push this card to it's designed limits. Here's the page talking about it: http://www.anandtech.com/video/showdoc.aspx?i=3679&p=4

I will say that the AnandTech article is a bit confusing. They spend every page from 4 to the end of the article saying how disappointed they were with overclocking, but every single test (except for 1 that's CPU bound and thus pointless) showed a noticeable performance gain when the card was overclocked.
22) Message boards : Number crunching : I've got an Idea.... (Message 33435)
Posted 21 Nov 2009 by ProfileJerWA
*takes a number*

So, uh... is there where I whine about how Collatz and MW are both down at once, and have been for days? Oh, and I just finished installing a new shiney HD4890 (I know, I know, 4800 series is so last month) which is now idling. It's a crime, really.

So... is this the place?

23) Message boards : Number crunching : Why No Work? (Message 31593)
Posted 26 Sep 2009 by ProfileJerWA
Yeah, normally 24 WUs would be a lot, but when that's less than a half hour of actual work, not so much.

I can understand why though, since new WUs are generated from the results of previous WUs, at least if I understand it correctly? I wonder what would be involved in increasing the number of WUs currently "live." I imagine there's a reason he hasn't already.
24) Message boards : Number crunching : Sapphire Radeon HD 4670 512MB PCI-e and Milkyway problem (Message 31584)
Posted 26 Sep 2009 by ProfileJerWA
Here's the official list of cards that support Stream, and whether they are single or double precision.

25) Message boards : Number crunching : Why No Work? (Message 31581)
Posted 26 Sep 2009 by ProfileJerWA
Here we go again. MW out of work and Collatz blown up.
26) Message boards : Number crunching : ATI 4770 work ... or not? (Message 31232)
Posted 20 Sep 2009 by ProfileJerWA
Beyond, if you haven't already, bump up to the 9.9 drivers. For me they've been near bullet proof, a marked improvement from all of my previous 9.x drivers.
27) Message boards : Number crunching : BOINC 6.10 support with ATI (Message 29651)
Posted 23 Aug 2009 by ProfileJerWA
Tried 6.10 myself and it ran nicely but would not request GPU work, even with debts being manually set and forced updates. So, back to 6.4.7.
28) Message boards : Number crunching : ATI GPU app 0.19f fixes the ps_sgr_208_3s errors (Message 28790)
Posted 3 Aug 2009 by ProfileJerWA
app_info.xml in the project directory for milkyway

avg_ncpus sets how many "CPUs" each WU takes.
max_ncpus sets how many total "CPUs" MW can take.

Default is .1 and 1 = 10 running WUs, even if only 1 is getting resources (via n1).

To change that behavior, adjust avg_ncpus. I want 4 WUs marked "running" and 3 actually running so I use no n setting (since 3 is default), avg_ncpus .33, and max_ncpus 1.

That marks 4 running, 3 are actually getting resources. I do that so the stupid manager never "sleeps" the project.

If you really only want 1 running, and 1 marked running, set n1 and avg_ncpus 1.

PS: You can edit the file any time you want but changes only take effect when you restart the BOINC manager.

Oh, forgot one more thing: None of this may help. The manager may just decide it needs more cores to meet debt requirements, and not run any CPU work anyways, even if you lock MW to just 1 WU at a time. Not sure about the 5.x managers, but that's what the 6.x managers do when they get confused. Watching a CPU core idle because the manager is trying to run multiple instances of MW. <sigh>
29) Message boards : Number crunching : boinc 6.6.36 version (Message 28789)
Posted 3 Aug 2009 by ProfileJerWA
And yes, tons of running WUs despite your settings for avg and max ncpus, all with their timers counting even when not getting resources is how 6.6.36, 37, and 38 all behave. It breaks it even worse than it already is, because not only is the scheduler still trying to schedule it against CPU time, but now it thinks every WU is going to take hours instead of minutes to run. Seems like nothing but negative progress since 6.4.7 which is the last version I've found to behave at all predictably.

Not saying it's better, just saying it's broken in a way that's predictable and can be worked around. Unfortunately it's very touchy about how projects schedule WUs, so I have to tweak it for every project I'm mix and match with MW. Was running it with Enigma last month for instance, and they like sending batches of very short deadline units mixed with normal deadlines, causing the manager to panic "high priority" stuff breaking all scheduling and resource sharing. Sooo, had to set very short queues, and script a debt reset to positive debt values for MW which was also running 205 resource share vs 50 for Enigma, and forced 30 minute project updates. Took awhile to figure that mess out, but once I did 80k/day smooth as could be. Now I'm on Yoyo and we're back to it being a mess again. Ah well :-)

Official CAL support isn't due until 6.10 or 6.12 "near the end of the year" at the earliest.
30) Message boards : Number crunching : server crash (July 29) (Message 28578)
Posted 30 Jul 2009 by ProfileJerWA
I still don't have any MW stuff on BOINCStats, but I figure since it wiped my account completely it'll take until the next major update to sort itself on the stats sites. Good work on the quick recovery.
31) Message boards : Number crunching : Any way to hack "non CPU intensive" client side? (Message 28264)
Posted 23 Jul 2009 by ProfileJerWA
32) Message boards : Number crunching : Any way to hack "non CPU intensive" client side? (Message 28256)
Posted 23 Jul 2009 by ProfileJerWA
Was digging through the client stuff and noticed this flag, meant for projects that are primarily not using CPU. Sound familiar? Unfortunately it is currently project wide, i.e. if you've got CPU apps and GPU apps they'd all get flagged non CPU intensive and bad things would ensue.

I was just curious if anyone has looked at the possibility of hacking this client-side, if it's something that gets pushed in the configs or whatnot. Since we're only running 1 flavor of app at a time, if we could do it locally as part of the GPU app install/setup (even if we have to do it manually), I think that'd go a long way towards sorting the ATI scheduling issues until we can get an official fix that has to come from the BOINC client folks. Likewise, since we'd be fooling the client locally, it wouldn't negatively impact the CPU app users.

Just trying to find a shorter term fix than "in 6.8, no 6.10, maybe 6.12 around the end of the year..." which is starting to read like never. Are we out of luck or is this something we can hax0r? :-D
33) Message boards : Number crunching : ATI GPU app 0.19f on Vista ulimate 64 + 4870 x2 (Message 27092)
Posted 5 Jul 2009 by ProfileJerWA
Are you still using 6.6.36?

I've tried everything from 0.3 to 0.01 and only 0.01 made a difference, and that wasn't a good change (all MW WUs trying to run at once, to the point of shutting down all but 1 CPU app). Still don't quite understand what this value is really doing in terms of scheduling.

I've got mine set to .2 so that 5 WUs will flag "running," but it doesn't seem to have any impact on scheduling at all. I am finding out, however, that Enigma (which I'm only running because it's the KWSN project of the month) doesn't play well with MW (or QMC, since every single QMC WU that ran with Enigma running hit 100% and then failed computation). The wrapper doesn't update progress and timers very often and it tends to make the manager panic a lot that WUs are going to run-over their predicted time. I've had to "kick-start" MW several times tonight.
34) Message boards : Number crunching : ATI GPU app 0.19f on Vista ulimate 64 + 4870 x2 (Message 27033)
Posted 4 Jul 2009 by ProfileJerWA
Nice! Today I learned that trying to install the BOINC Manager without network connectivity = fail. Had to wait for my Internet connection to come back online before I could get any version of the BOINC Manager to load my existing projects and work. If I temp'd the data directory away it loaded up fine, and would even do so with my projects (all cleared, "initializing") but no version would load with my current work load until my 'net was back.

Played with 6.6.36 a bit, but the fact all the MW units "run" the timer up w/o actually running messes things up a whole bunch. Estimating 40+ minutes on every WU means I was constantly having to force the next batch of MW work to start as the manager didn't want to start such long WUs. Back on 6.4.7 and tampering with settings more hehe.
35) Message boards : Number crunching : ATI GPU app 0.19f on Vista ulimate 64 + 4870 x2 (Message 26980)
Posted 3 Jul 2009 by ProfileJerWA
If I had to guess it's project debt. The MW vs BOINC Manager fight affects every project. You cannot change debts without it impacting all other projects on a machine. And since the manager is trying to schedule it against CPU time that it isn't using things get wonky.

Since I started MW these are all the new and interesting things I've seen BOINC do (after running for 3 days w/o a single glitch when I first joined):

- Stop requesting work for unrelated projects (caused by: debt for that project jumping to a negative six digit number despite all other projects still having normal debt #'s)

- Stop running anything on 1 core of my 4 core CPU (caused by: suspect it's debt again, as the manager tries to "make up" debt to MW by giving it CPU time it doesn't want, which is why I say it's a balancing act)

- Estimate MW work units at everything from 40 seconds to 32 minutes to completion (wish I knew the cause, happens once every 24-32 hours or so)

- Decide that 2 day queue = 2 WUs, or 4, or maybe 8, or every WU it will let me have, in no particular order, and constantly changing minute to minute (doesn't bug me overly much but MAN does it spam the MW server a lot... my machine = Created 22 Jun 2009 13:19:56 UTC, Number of times client has contacted server 6961)

- Mark MW work as running that isn't getting any resources (think that was a CAL failure)

- Change from a 60-minutes-per-project cycle to a 2-seconds-per-project cycle, fixable only by restarting the manager

I suspect that the majority of the weirdness is the fact that BOINC Manager sees MW as a CPU app and tries to schedule it accordingly. Meanwhile, actual CPU apps are using all of your CPU time. Therefore you have projects picking up and losing more "total" debt than can be accounted for in the amount of time it's running. Even worse, the MW GPU app DOES use some CPU time, just relatively little, so CPU apps get shorted a few seconds too. It's a mess, but it sure is fun. :-)
36) Message boards : Number crunching : ATI GPU app 0.19f on Vista ulimate 64 + 4870 x2 (Message 26970)
Posted 3 Jul 2009 by ProfileJerWA
PoorBoy I think I remember you mentioning that you've got MW on a low resource share, right? My script will fix the debts, but then the manager is going to try to schedule it like a CPU app which means it's competing for resources it doesn't actually need. Set the project resources higher than your other stuff so that the manager will want it to be running more often. It's still going to dig itself into a credit debt problem, but at least it will keep running until the next time the script fixes debts hehe.

All of my projects are at the default 100 resource share, MW is at 190 now (also tweaking it to find the sweet spot where it doesn't drop to 3 CPU apps).
37) Message boards : Number crunching : ATI GPU app 0.19f on Vista ulimate 64 + 4870 x2 (Message 26897)
Posted 2 Jul 2009 by ProfileJerWA
I noticed the project has implemented a WU cap as well, as I'm getting pretty regular "No more work for you" messages that I wasn't before.

Seems to be running well now, for the most part. Unfortunately it does not recover from any unexpected changes, so that's always an automatic required manual intervention. This is as much (or more?) about the Manager 6.4.7 as it is MW.

For instance my home network lost Internet access for most of the day yesterday. Since MW will only queue about 20 minutes of work at max (and rarely keeps the queue that deep) it promptly ran dry and just sat there all day. When my Internet connection was restored, MW ran fine, but the weird debts caused by the lack of access made the manager run only 3 CPU tasks no matter what else was running (I think because it was trying to fill MW debt for a change). Had to exit and restart the manager then it fixed itself.

Also noticed that one of my projects that was running fine for awhile has now fallen on its head. Checked the client state file and that project is sitting at something wild like -138000 debt now. No clue why. Ah well, guess it needed a break hehe. As long as all 4 CPUs are churning and MW is keeping the GPU fires lit I'll let it figure itself out. And that's mostly what I've accomplished with my weird, tangled, overly complex setup. So until they release a better manager and/or just give us back control over our own resources I'll leave it be.

PS: I forgot to mention, the "script" is just a plain text file. Use notepad to create it, as Wordpad will probably leave weird formatting that may cause problems. You can right-click make new Text file, then just rename it .cmd after you edit it.
38) Message boards : Number crunching : ATI GPU app 0.19f on Vista ulimate 64 + 4870 x2 (Message 26894)
Posted 2 Jul 2009 by ProfileJerWA
Uhm, I DID make it a script. You really think I manually reset it hourly? LOL. Sorry I wasn't clear about that part, I just assumed nobody would want to babysit it that much and would figure out a way around it. My fault. :-) So, here's the fix:

Here's my "script," you can name it whatever you want, mine is MWDebt.cmd

Save it in your BOINC manager directory.

CD "D:\Program Files\BOINC"
boinccmd --set_debts http://milkyway.cs.rpi.edu/milkyway 0 500

Change "D:" to whatever drive your BOINC manager is installed on. Do not skip this step.

Change "D:\Program Files\BOINC" to the location of your manager, not the projects! Do not skip this step.

The 0 is short-term debt, the 500 is long-term debt. I used 0 0 for awhile but it seems to dig itself into a hole too fast. I've been playing with the values a little bit to see if it makes things more predictable.

On Windows, go to Control Panel, Administrative Tools, Task Scheduler. Create a new task to run the .cmd file you made (i.e. MWDebt.cmd). Make it trigger once a day and then run every hour. Haven't looked into how to do that on any OS but Windows 7 yet (I have, just a long time ago), but I'm sure Vista has very similar options. You could also just make it run every hour, it's just that Win 7 has a specific "repeat every" option that makes it even easier to just have it trigger once and repeat every hour hehe.

The reason you shouldn't skip the drive and directory steps is that the default BOINC Manager security will only allow boinccmd when run in the same context as the manager (i.e. same path). If you run it from another path it will require user credentials, which unless you've setup remote access is going to be some huge random string of letters and numbers.

If you need more help setting this up to try just let me know.

Edit: Just changed my debts to 0 short, 500 long after noticing the project stalling out again and at -1800 debt even with hourly debt resets. Sheesh!
39) Message boards : Number crunching : ATI GPU app 0.19f on Vista ulimate 64 + 4870 x2 (Message 26806)
Posted 1 Jul 2009 by ProfileJerWA
I've done all that numerous time, you shouldn't have to do it though. I have Milkyway set to run 99.20% of the time on my i7 yet if I didn't constantly coax it into running more MWay Wu's the ATI Cards would sit Idle 1/2 the day or more ...

Right now I've just stopped running anything but the ATI Cards on those Box's, the CPU's can sit Idle if that's what it takes to keep the ATI Cards running.

I didn't invest the Money I did in the ATI Cards with more coming to just see them sit Idle because their waiting to run ...

Done all what numerous times? If you're reseting debt you have to do it hourly to keep up with the project. It's not a one time fix, it's something you have to do constantly in order to keep the project going. And no, you shouldn't have to do it, but that's the BOINC managers fault not MWs. The app is GPU but not CUDA, the manager doesn't understand any GPU apps but CUDA, therefore it schedules it like a CPU app. The fact it runs all the time despite the CPU not being used constantly digs itself further into "negative" debt, i.e. telling the manager to schedule the other projects before it.

My system with just a single OEM spec 512MB HD4870 ran itself to -18000 debt in just a few days and wouldn't run MW work at all because it was waiting for the other projects to "catch up."

As for 99% that doesn't matter, because it's going to run 100% of the time and constantly dig itself into negative debt when the scheduler can't figure out why it's not able to catch up that .1% on other projects. My workaround so far is working near perfect for me. I'm losing about 2% CPU time when the manager can't figure out what to do and I end up with 3 CPU apps and MW running, but the projects are all running now, all queuing work, and MW isn't running dry or hanging up.
40) Message boards : Number crunching : ATI GPU app 0.19f on Vista ulimate 64 + 4870 x2 (Message 26792)
Posted 30 Jun 2009 by ProfileJerWA
Check your project debts. In the main BOINC data directory you should find a client_state.xml file with all the projects current status in it. Betting MW is sitting at some large negative number for debt, which is why it isn't running. The BOINC Manager treats it like a CPU app, and gets progressively more angry about it running more than the normal CPU apps (since it's using GPU time instead). It runs enough work on my system that I have to reset debts hourly to keep it running predictably.

Previous 20 · Next 20

©2023 Astroinformatics Group