Welcome to MilkyWay@home

Posts by ExtraTerrestrial Apes

21) Message boards : Number crunching : Multi core support (Message 53704)
Posted 17 Mar 2012 by ExtraTerrestrial Apes
Post:
No worries, we're here to help!

MRS
22) Message boards : Number crunching : Multi core support (Message 53701)
Posted 17 Mar 2012 by ExtraTerrestrial Apes
Post:
I don't think these GFLOP number are correct, or at least the amount of credit and computation time you assign are not right. If you look into your profile under "computers belonging to this account" and there on the machine and then "workunits", you'll see that for n-body (all cores) you get about 8 credits per minute, whereas the separation WUs (one core each) take about 10 ks for 160 credits.

Overall that's a maximum of ~10300 RAC for separation tasks and 11000 RAC for n-body tasks. That's a nice agreement, actually. However, AMD GPUs (not your's, only some high end models) can crunch these separation WUs within a minute, so you may consider running only n-body on your CPU to contribute efficiently.

MrS
23) Message boards : Number crunching : ATI HD 5870 not being recognized (Message 53687)
Posted 15 Mar 2012 by ExtraTerrestrial Apes
Post:
The "HD5870 mobile" is actually a HD5770, so BOINc is almost correct. You could also give the new POEM OpenCL app a try. From my point of with a far better investment of computation time than Collatz. Works best with an app_info and multiple WUs per cards, though.

MrS
24) Message boards : Number crunching : Ridiculously high computation speed - what's going on here? (Message 53678)
Posted 14 Mar 2012 by ExtraTerrestrial Apes
Post:
You're right: although you've got fine CPU, 11 TFlops can not be extracted from this hardware even under optimal circumstances. My best bet would be that the FLOPs estimation of the WU is wrong.

MrS
25) Message boards : Number crunching : CPD going down... (Message 53601)
Posted 9 Mar 2012 by ExtraTerrestrial Apes
Post:
No change in credits or run times for me either.

180s is really slow for a Cypress. And your run times vary greatly, which is unusual for a well-running MW. Are you leaving a CPU core free for MW?

MrS
26) Message boards : Number crunching : Do the WU get less credits now since thanksgiving crash (Message 53444)
Posted 26 Feb 2012 by ExtraTerrestrial Apes
Post:
Sorry, I assumed you were talking about the "separation" type tasks. I'm not running the nbody-stuff, so I can't tell whether credits have changed there.

Otherwise your calculation is in principle correct. You'd want to use "run time" rather than "cpu time", as nbody uses both of your CPU cores. And to a varying extend. For the longer WUs "run time" approaches half of "cpu time" - which would be perfect scaling on a dual core. For the smaller WUs (which you used for the estimate) the speed-up due to the 2nd core is less. So, as Mikey said, your actual RAC will depend on the mix of WUs you're getting.

And your CPU is also running separation tasks. Averaging over 4 of them yields 31500 s/WU. That's 24*3600/31500 = 2.7 WUs/day per core, or 5.4 WUs per day for your CPU. Each is worth 160 credits, so you'd get an RAC of 877 running only these. The more of them you get, the lower your actual RAC will be.

By the way: my GPU crunches through these separation tasks in 50s, the current top cards achieve ~30s. Your CPU could be much more useful elsewhere. And your GTX460s need 8000 - 9000s for these WUs - not exactly efficient either.

If I were you I'd opt out of separation tasks and use the GPUs elsewhere (GPU-Grid, Einstein, SETI, POEM etc.).

MrS
27) Message boards : Number crunching : 79XX Dont Run II (Message 53442)
Posted 26 Feb 2012 by ExtraTerrestrial Apes
Post:
To summarize that thread:

- Tahiti is available with 4 load voltage bins (as you listed)
- current Afterburner shows the voltage correctly, whereas other tools might just get 1.175 V from a database
- the chips are binned by leakage and, as it seems, high leakage chips get lower voltages to keep their power consumption approximately in line with the others
- under load the real GPU voltage drops, just as for CPUs.. but this doesn't matter for us
- most cards are 1.175 V default

I take from this that the "typical power consumption" I listed previously is still correct, as the lower voltage parts make up for that with their higher leakage. They would consume a bit more power if run at the same voltage, though.

And I still stand by "if the card typically draws ~200 W, there shouldn't be a need to allow 300 W instad of 250 W".

MrS
28) Message boards : Number crunching : 79XX Dont Run II (Message 53395)
Posted 23 Feb 2012 by ExtraTerrestrial Apes
Post:
Without that +20% you will have problems.

Has anyone actually tested this? I highly doubt you'll draw more than 250 W running MW, even overclocked, as long as you stay at or below 1.175 V (stock). In games the card hardly draws more than 200 W.

Plus I'm running my HD6950 overclocked and with unlocked shaders, yet don't need to increase the default Power Tune limit of 200 W. Heck, the entire machine is drawing < 300 W from the wall with an i7 @ 4 GHz and full BOINC load.

Edit: yes, 38s is fast! Makes me a little jealous ;)

MrS
29) Message boards : Number crunching : Should I bother? (Message 53394)
Posted 23 Feb 2012 by ExtraTerrestrial Apes
Post:
Let's add a little.. perspective to the enthusiasm ;)

Contributing is great and such. However, going with Mikeys example of 10 bazillion workunits per day versus 0.5 an interesting question arises: what if your single WU was the all important one.. and you don't crunch it? What happens? Simple: the next PC asking for a WU gets it and now this one will discover the big bang. It's not that the discovery will not be made, it just takes longer.

And if many people choose not to contribute, at some point it will take so much longer to crunch through WUs that the all-important one will not even be issued to anyone.

So what you want to do is not "contribute somewhere, no matter how small". It's "contribute efficiently, donate where it counts". With Milkyway the situation is special in that we've got "separation" workunits, which can be run on GPUs or CPU. The GPUs are far far better at this stuff, that's why your CPU is looking ancient. On the other hand, if you run something which can not currently be done by GPUs, then your CPU is not that old and slow. At Milkyway you could run "nbody" workunits.

To do this go to "your account/Milkyway@Home settings" and disable "Milkyway@Home", but leave "MilkyWay@Home N-Body Simulation" enabled. Or any other projects which can only use CPUs.

MrS
30) Message boards : Number crunching : 79XX Dont Run II (Message 53380)
Posted 22 Feb 2012 by ExtraTerrestrial Apes
Post:
You mean Cat 12.3 preview? That would be way faster than with previous drivers, isn't it?

MrS
31) Message boards : News : Separation updated to 1.00 (Message 53371)
Posted 21 Feb 2012 by ExtraTerrestrial Apes
Post:
Update: error rate down to 0 out of 3350+ with the new app. Yay!

MrS
32) Message boards : Number crunching : Do the WU get less credits now since thanksgiving crash (Message 53227)
Posted 16 Feb 2012 by ExtraTerrestrial Apes
Post:
Do the WU now collect a smaller credit for completion.

They don't. Long term RAC is highly dependent upon the project being available all the time. RAC is averaged over.. I think about a month. So any downtime will stay in the stat for that long. In order actually reach your maximum RAC you'd need about a month of undisturbed crunching (on your side and on MWs).

You can check the speed yourself if you devide the seconds per day by your time per WU. Multiply by the credits per WU and you get your maximum RAC. Mine hasn't changed since a long time.

MrS
33) Message boards : News : Separation updated to 1.00 (Message 53172)
Posted 14 Feb 2012 by ExtraTerrestrial Apes
Post:
My error rate is down to ~0.1% from a starting value of 0.7 - 0.8%.
Good job here:
The occasional validate errors from empty / truncated stderr should stop

Edit: using the same command line parameters as with 0.82 I'm still seeing the same crunching speed as reported above.

MrS
34) Message boards : News : Separation updated to 1.00 (Message 53118)
Posted 12 Feb 2012 by ExtraTerrestrial Apes
Post:
Why is your CPU load so low? I'm leaving one logical core free on my i7 and that nails GPU utilization at 99%. Otherwise performance suffered too much for my taste (0.82).

MrS
35) Message boards : News : Separation updated to 1.00 (Message 53113)
Posted 12 Feb 2012 by ExtraTerrestrial Apes
Post:
Been running 6.10.58 for a long time on Win before I switched to 6.12.

MrS
36) Message boards : News : Separation updated to 1.00 (Message 53109)
Posted 12 Feb 2012 by ExtraTerrestrial Apes
Post:
Results

Config:
HD6950, unlocked shaders (HD6970), 900 MHz GPU, 625 MHz memory, Win 7 64, i7 2600K @ 4.0 GHz, BOINC 6.12.34

Before:
0.82 CAL + app_info, 1 WU at a time, parameters: --process-priority 3 --gpu-disable-checkpointing
Times: 53 - 54s GPU, 2.4s CPU

Now:
1.02 OpenCL, no app_info, using AVX code path
default: 56s GPU, 3.3s CPU
setting refrseh rate to 10 Hz: 55s GPU, ~2.9s
setting refrseh rate to 1* Hz: 55s GPU, ~2.9s

That's about 3.5% slower. Previously the error rate was ~0.7%, so if these were fixed now (can't tell yet) overall throughput would be less than 3% lower. That's about 7800 credits/day less.. but running without an app_info if nicer.

MrS

(*) No worries, the IGP is driving my display ;)
37) Message boards : Number crunching : Very Long Crunching Time on Celeron (Coppermine) (Message 52835)
Posted 6 Feb 2012 by ExtraTerrestrial Apes
Post:
That old Celeron wasn't that crippled. It's got 128 kB L2 instead of 256 kB for the full Coppermine, and an FSB of 100 MHz compared to 100 or 133 MHz for the Coppermine. Everything else is similar.

Since it's only got SSE1 you might want to use it in a project which doesn't benefit from SSE2+. Otherwise it will always seem absymal compared to modern hardware.

MrS
38) Message boards : Number crunching : Very Long Crunching Time on Celeron (Coppermine) (Message 52825)
Posted 5 Feb 2012 by ExtraTerrestrial Apes
Post:
Raw MIPS are very unrelated to real world performance :p

MrS
39) Message boards : Number crunching : 79XX Dont Run II (Message 52771)
Posted 1 Feb 2012 by ExtraTerrestrial Apes
Post:
In the end I decided to refuse my 7970 delivery and send it back. It's not worth spending over £400


Yeah, buying it for BOINC is not worth it right now. You can pull the trigger whenever the software is ready, and at that point it will certainly not be at £400 any more.

MrS
40) Message boards : Number crunching : 79XX Dont Run (Message 52717)
Posted 28 Jan 2012 by ExtraTerrestrial Apes
Post:
@Crossfire: does that matter at all? It's for games and only used if AMD gives you a profile for the game. This should have nothing to do with GP-GPU, where you usually run 1 pogram per GPU (unless specified otherwise). If XFire was working as intended you'd have to see halved times per WU, but only running 1 WU at a time per divce, compared to several WUs at once in the normal mode.

@Testing: with the recent server upgrade and insta-purge being gone we've finally got a good method of testing stability: at the bottom of "show tasks for your computer" it shows valid results, invalids and errors.

For me about 3000 tasks are kept in this record, which is statistically relevant. At about 1600 WUs/day it takes 2 days for this statistic to update completely after I change something. I suggest you guys use this instead of "only 1 or 2 errors in x hours". That number is too small to judge stability.

On my Cayman I observed a failure rate of 0.7 % at 900 MHz @ 1.10 V. That's what I settled at in painful hand-tuning. Now I increased the GPU clock to 905 MHz and observe an increased failure rate to 1.4 %. That's a significant increase and certainly not worth it. Trying 1.11 V now.

MrS


Previous 20 · Next 20

©2019 Astroinformatics Group