Multiple GPU wu's

[NudgeyNR]

Is there a way to make collatz run multiple GPU wu's ?

[Dingo]

In the BOINC/projects/boinc.thedonntags.com_collatz folder add a file called app_config.xml
In that folder paste the following;

Code Select Expand

<app_config>
<app>
  <name>collatz_sieve</name>
  <max_concurrent>0</max_concurrent>
  <gpu_versions>
     <gpu_usage>0.5</gpu_usage>  // 1.0 = 1 instance, 0.5 = 2 instances, 0.33 = 3 etc
     <cpu_usage>0.6</cpu_usage>
  </gpu_versions>
</app>
</app_config>

After you have added the file stop and start BOINC and then look in the boinc manager event log to see that it is OK.


I do not run app_config.xml on Collatz as I found that running one at a time is better overall but you can try different options.

[JugNut]

Hey Nudgey,  yea you can and it does help a little but it's much more productive to tune your settings for best results.  A bit of reading and testing is required to see which setting work best for each GPU you have but it's by far the best way to go. It's possible to increase production close to 2 fold maybe even more depending on the GPU your using. It's a large increase over running multiple Wu's. 

The settings file can be found in the collatz folder (boinc.thesonntags.com_collatz)  which is can be found in the boinc data directory.
The file name will be similar to the GPU app that your using.  EG: For AMD cards it's a text file called.. collatz_sieve_1.30_windows_x86_64__opencl_ati_gpu.config
For Nvidia Cards it's collatz_sieve_1.30_windows_x86_64__opencl_nvidia_gpu.config

If you like you can try a few settings you find in this thread. https://boinc.thesonntags.com/collatz/forum_thread.php?id=8#69
These are the setting I use.

For the AMD radeon VII .

verbose=1
kernels_per_reduction=48
threads=8
lut_size=18
sleep=0
cache_sieve=1
reduce_cpu=0
sieve_size=29

For all my Nvidia cards.

verbose=1
kernels_per_reduction=48
threads=8
lut_size=17
sleep=1       <-- try this setting with 1 or 0 to see which is best for you.(only needed for Nvidia cards )
cache_sieve=1
reduce_cpu=0
sieve_size=29


The two most important settings are...
lut_size=  <-- Try increasing this number +1 at a time until your times start getting slower or ---
                      you get a compute errors. Then back down 1 number. 
sieve_size=  <- same with this setting as with the above. 

Give it a try it's not as hard as it sounds.  Plus the extra credit is unbelievable. 

First download some work and then run a few tasks to see what your baseline is.  Then suspend all work and add the new settings to the above mentioned .config file.(edit the .config file using notepad). 
After the settings have been inserted and saved click on the "options" menu in boinc manager and click the menu item that says "read config file" then suspend just one task and see how it runs.
If it fails with a computation error then try lowering both the lut_size and sieve_size settings by 3 if those settings work then try raising each setting +1 at at time until you have the best setting possible. If it fails again try lowing those two setting one number lower and try again ect. 

It's also possible to see the settings others are using by simply looking at their tasks output on the collatz website. So if you find someone with a similar GPU than you have that seems to be doing a lot better than you you can try his settings.  Well you can unless that person hides their devices(in the top computer ranking the host is marked as Anonymous)

This is the first guy I could find who was not anonymous in the top lists with a 1080 ti like yours .  https://boinc.thesonntags.com/collatz/results.php?hostid=846624&offset=0&show_names=0&state=4&appid=  Click on any of his work units in the  "task" column.  It should look like this...  https://boinc.thesonntags.com/collatz/result.php?resultid=43732560  If you scroll down the page you'll get to the part we are after. It's in the box called Stderr output.

You can see the settings he used in his  .config file

verbose             1 (yes)
kernels/reduction   48
threads              2^8 (256)
lut_size              17 (1048576 bytes)
sieve_size           2^30 (51085096 bytes)
sleep                  1
cache_sieve        1 (yes)
reducecpu           0 (no)

Which translates almost exactly to the settings he used in his .config file.

verbose               1
kernels/reduction 48
threads               8
lut_size              17
sieve_size          30
sleep                 1
cache_sieve       1
reducecpu          0

I hope this helps.  Post if you have any difficulties.

[Dingo]

I have the same setup as JugNut but had to change my config file recently when I updated the driver and it stopped working with an error. (Error: GPU steps do not match CPU steps. Workunit processing aborted.)   It is a bit of chnage the config file close boinc and startup again and see if the work unit still runs.  I have had to go down to the following on my 1080Ti;

Code Select Expand


verbose=1
kernels_per_reduction=48
threads=8
lut_size=17
sieve_size=25
sleep=0
cache_sieve=1
reduce_CPU=0

If you go this way a bit of trial and error with the "threads" and the "sieve_size" values are needed.

[NudgeyNR]

19:06 down to 7:22 - a massive difference and no errors yet.

Thanks for help with the much needed changes in the config file.

[JugNut]

19:06 down to 7:22 - a massive difference and no errors yet.

Thanks for help with the much needed changes in the config file.

Nice!!! Well done


@ Dingo: Sometimes when I fire up collatz after not crunching there for a while I start getting constant compute errors. All I do is reduce the  "sieve_size=" down from 29 to 27. Then after it starts crunching properly I wait a further 10 - 15 mins then change it back 29 and then it will works as it should once again.  Well so far so far it has.  I've only seen that behaviour on my GTX 970's but I imagine there could be a lot of GPU's that have similar problems.  It's almost like it needs to get warmed up first.

When you tune Collatz properly it really works the GPU hard probably more than any other project.  But the beauty of this method is you can tune it to where your happy with how the GPU is running.
If you think it's running to hot then just lower the settings until you get to a place you like.  I wish all projects were like this.    LOL The massive credit doesn't hurt any either.

[NudgeyNR]

After making the changes yesterday everything appeared to be running ok, the GPU was running a little hotter, then right as the cricket started black screen of nothing and reboot. 

Several hours passed in which i tried numerous ways to get windows to run. Windows would get to the login screen then after about 5secs the pc would simply reboot over and over and over again, with no errors at all.

The last thing i tried was uninstalling the GPU drivers, but ofcourse being winblows 10 i was unable to uninstall them even from safe mode.

So in the end i have reinstalled windows. Hopefully nothing goes wrong now.

[JugNut]

Eekk that terrible news perhaps next time(if your game) try with a sieve_size=26 and see how how that goes then work up +1 at a time until your sure of stability. Perhaps try no higher 28 Yea it really works your GPU hard the further you tune it. 

I sure hope you get up & running again soon. 


EDIT: I'll do a few today to see how it fares.  Maybe the latest app is slightly more sensitive?
EDIT2 : I use DDU to uninstall sticky drivers. While a lot of people swear by it I heard others say it didn't work at all for them so I guess YMMV? If you do use it then use the the way in which it was made to work, in safe mode.  https://www.wagnardsoft.com/
EDIT 3: I can not get any work at the moment. I'll try again later.

[chooka03]

I can sympathize with you Nudgey. I've done a couple of recent wipe and fresh installs of Windows. It's all the updates after reinstalling that's the painful part.

I've learnt to keep nothing on C: drive.

Goodluck mate 

[JugNut]

Yes it's a pain alright   

@ Nudgey: Well i've started running off a few collatz some with a sieve_size=28 but most with 29 and so far they seem to be working as expected.
What does that say? Well not much at all really, it only says that everyone's systems are different. 
But while the first 1080 ti was running cool at 66 deg the second one was running hotter than i'd usually like.(close to 80 degs) while the hot card is a much higher clocked model thats still to high for 24/7 crunching.
In the end I suppose only through testing can you figure it out?  But even if you use a sieve_size=26 or 25 as dingo does there's still huge improvement to be had

I hope things work out for you.  In the end i'm sure you'll find a reliable setting.

If your interested this is a link to my collatz work units. https://boinc.thesonntags.com/collatz/results.php?hostid=137377&offset=0&show_names=0&state=4&appid=

PS: As you see by the link for some reason you get lower credit per WU from tuning,  the higher the tune the less credit you receive.  I would have thought it would have been the other way around.  This isn't new behaviour it's always been that way but I thought i'd point it out anyway.  It sure is a head scratcher as to why?  But regardless because WU is done so much faster your overall credit is improved by a ton. 

Come to think of it that probably raises another line of thought?  There must be a point where increasing the sieve_size becomes detrimental to credit. In other words even if you crunch each work unit a little bit faster you lose enough credit for it to no longer be worth it.
Just a thought..

[tazzduke]

Greetings

Have thrown 2 x WIN pcs onto Collatz, to see if I can even get a single workunit, so far nil, nada, nothing.  I only want to do a couple of workunits to dial in my GPUS.

Might have to email the admin to see if I can get an answer lol.

Cheers

[JugNut]

One of the reasons I wasn't getting work at first was because of a security error from collatz in the message logs.  After adding & removing I received work about an hour later.  After receiveing a good hand full(about 50) they started coming pretty regularly.

Well until a wind storm hit here that is.  In my area(sydney west) there was lots of local damage done to trees & roofing plus we had a couple of blackouts just to make things extra interesting. Actually the blackout has done something odd to the power in my house where now about a quarter of the houses lights have stopped working,  at the moment I have no idea whats gone wrong.  The circuits haven't been tripped so something else must of happened.  I will see if I can figure out what went wrong tomorrow in the daylight. Luckily my crunching room still has power.
My crunchers were not impressed about the blackouts but at least they all came back up without to much of a fight with the exception of some corrupted files that is.

Anyhow I hope you get some work soon tazzduke.
All the best.

[NudgeyNR]

No luck so far, my pc keeps resetting whenever i mod the settings.

[JugNut]

I guess you've tried with a lower sieve_size= of say 24 , 25 or 26?  As I mentioned previously some of my GPU's need to start off with a lower sieve_size= until they warm up a bit.

Also as mentioned the only two setting that make a huge difference are..
lut_size=
sieve_size= 

Try lowering both one value lower.

You could also try a collatz project reset.  That way you you'd know you have all new apps & files. It's at least possible that one of the files was damaged during the crash? It's worth a try at least.

How is the card crunching without the added settings?

[NudgeyNR]

With my GPU (1080ti) set to default clocks and collatz to default settings the wu's are completing between 8-9mins.

Had to increase GPU fan to 70%, temps came down from 74c to 60c allowing auto OC to increase to much higher speeds.

Not sure if it is worth changing settings for the extra degrees seen on the GPU, to get the wu's to be better than 8mins.