Collatz GPU optimisations

Forum rules
davidbam
General Bitchin'
General Bitchin'
Posts: 6371
Joined: Wed Aug 15, 2018 1:15 pm
Location: Huntly, Scotland
Contact:

#21 Re: Collatz GPU optimisations

Post by davidbam »

Alez wrote: Tue Nov 20, 2018 11:22 am If the utilisation is at 100% with one unit per card then no point in running 2 per card
THIS !!!

I remembered it all wrong. With the app optimisation on Collatz, there is no benefit from running 2 WU per GPU. I think that thought was a hangover from trying it on a sprint.

a Collatz WU is now completing in 6 mins +/- a few seconds so comfortably over 13 million / day Collatz from one machine. It has 9 threads on WCG as a bonus
I think this is fool-proof but could you just try it for me please? • There are 10 types of people in the world; those who understand binary, and those who don’t
Image
User avatar
Alez
[ TSBT's Pirate ]
[ TSBT's Pirate ]
Posts: 10363
Joined: Thu Oct 04, 2012 1:22 pm
Location: roaming the planet

#22 Re: Collatz GPU optimisations

Post by Alez »

Sorted then, 1 unit at 6 mins and 2 at 13 mins. Better to run the single units and all working as it should be. Very nice points haul per day from only 2 cards.
Those two cards are matching the output I have from 4 and a bit cards :clap:
................. Of course this is how arms races start :whistle:
Image
The best form of help from above is a sniper on the rooftop....
davidbam
General Bitchin'
General Bitchin'
Posts: 6371
Joined: Wed Aug 15, 2018 1:15 pm
Location: Huntly, Scotland
Contact:

#23 Re: Collatz GPU optimisations

Post by davidbam »

Cheers - I am keeping my eyes peeled for another GTX1080 as they go for a fair bit less than GTX1080Ti. The Ti doesn't seem to be dropping much in price - possibly due to the bad press that that its successor seems to be getting.

Doubtless the points-per-£ ratio will all change in a few months though
I think this is fool-proof but could you just try it for me please? • There are 10 types of people in the world; those who understand binary, and those who don’t
Image
davidbam
General Bitchin'
General Bitchin'
Posts: 6371
Joined: Wed Aug 15, 2018 1:15 pm
Location: Huntly, Scotland
Contact:

#24 Re: Collatz GPU optimisations

Post by davidbam »

my next problem is very slow WU on GTX960 - 2.5 hours or thereby so approx 320K per day.

I've tried the optimisation parameters from https://boinc.thesonntags.com/collatz/f ... postid=769 and they give a computation error. I note this was for Windoze; is it different under Linux?
I think this is fool-proof but could you just try it for me please? • There are 10 types of people in the world; those who understand binary, and those who don’t
Image
User avatar
Alez
[ TSBT's Pirate ]
[ TSBT's Pirate ]
Posts: 10363
Joined: Thu Oct 04, 2012 1:22 pm
Location: roaming the planet

#25 Re: Collatz GPU optimisations

Post by Alez »

davidBAM wrote: Wed Nov 21, 2018 7:33 pm my next problem is very slow WU on GTX960 - 2.5 hours or thereby so approx 320K per day.

I've tried the optimisation parameters from https://boinc.thesonntags.com/collatz/f ... postid=769 and they give a computation error. I note this was for Windoze; is it different under Linux?
That is wrong. The GTX 960 is only 2GB and not massively more powerful than my GTX 750ti, but that card manages an un-optimised task in approx 46 mins on average.
The comp error is probably due to the card not handling the optimisation especially sieve_size=30 probabily overflows the 2 GB memory. I'd expect the card to run a task in approx 30 mins.

try

Code: Select all

verbose=0
kernels_per_reduction=48
threads=9
lut_size=17
sieve_size=27
reduce_cpu=0
cache_sieve=1
sleep=0
if that crashes go down to sieve_size=26 and threads=8
Image
The best form of help from above is a sniper on the rooftop....
davidbam
General Bitchin'
General Bitchin'
Posts: 6371
Joined: Wed Aug 15, 2018 1:15 pm
Location: Huntly, Scotland
Contact:

#26 Re: Collatz GPU optimisations

Post by davidbam »

Thanks - that is marginally better but still estimating over 2 hours :-(

Card reports as having 4Gb btw ...
Wed 21 Nov 2018 23:20:29 GMT | | CUDA: NVIDIA GPU 0: GeForce GTX 960 (driver version 390.87, CUDA version 9.1, compute capability 2.1, 4096MB, 4011MB available, 691 GFLOPS peak)
Wed 21 Nov 2018 23:20:29 GMT | | OpenCL: NVIDIA GPU 0: GeForce GTX 960 (driver version 390.87, device version OpenCL 1.1 CUDA, 4535MB, 4011MB available, 691 GFLOPS peak)

Just noticed ... OpenCL 1.1? Is that correct? And CUDA version 9.1. Wondering if I have the wrong drivers?
I think this is fool-proof but could you just try it for me please? • There are 10 types of people in the world; those who understand binary, and those who don’t
Image
User avatar
Alez
[ TSBT's Pirate ]
[ TSBT's Pirate ]
Posts: 10363
Joined: Thu Oct 04, 2012 1:22 pm
Location: roaming the planet

#27 Re: Collatz GPU optimisations

Post by Alez »

[mention]davidBAM[/mention] I currently don't have any systems running GPU / Linux but 390 is old. I believe 396 works well as do the 415 / 416 beta drivers. Some have claimed problems with 410 and 411 although 410.73 are reported to be very good on GPUgrid ? Make of that what you will. You could try 396 or straight to the beta 415 / 416. I believe 410 is the latest version officially approved for LTS so you should be able to choose it from software center. If not, follow instructions below.
I see your 1080's and opencl 1.2

try updating the drivers:

Code: Select all

sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt-get update
Check which drivers are available in software center
or continue change the 410 to whichever version you want

Code: Select all

sudo apt-get install nvidia-driver-410
sudo apt-get install nvidia-modprobe --reinstall
sudo apt-get autoremove
if you want the absolute latest drivers then install this ppa

Code: Select all

sudo add-apt-repository ppa:xorg-edgers/ppa
check that the opencl hasn't been wiped after new driver install

Code: Select all

sudo apt install ocl-icd-libopencl1
Image
The best form of help from above is a sniper on the rooftop....
davidbam
General Bitchin'
General Bitchin'
Posts: 6371
Joined: Wed Aug 15, 2018 1:15 pm
Location: Huntly, Scotland
Contact:

#28 Re: Collatz GPU optimisations

Post by davidbam »

An update ... I was going fairly well until "sudo apt-get install nvidia-driver-410" errored with all sorts of dependency failures. I worked through these until 410 seemed to be installed but a reboot had boincmgr report no useable GPU.

I think I'll have a fresh try at this from a new install after the Sprint (as that machine is having some success with LHC CPU work). Problem is though - I have about 1000 Collatz jobs bunkered for the 20 fake GPU cards too :lol: :lol: I am going to be very unpopular if I reset that project
I think this is fool-proof but could you just try it for me please? • There are 10 types of people in the world; those who understand binary, and those who don’t
Image
User avatar
Alez
[ TSBT's Pirate ]
[ TSBT's Pirate ]
Posts: 10363
Joined: Thu Oct 04, 2012 1:22 pm
Location: roaming the planet

#29 Re: Collatz GPU optimisations

Post by Alez »

[mention]davidBAM[/mention] do this

Code: Select all

sudo apt-get install libnvidia-compute-410
and then do

Code: Select all

sudo apt-get install nvidia-driver-410
sudo apt-get install nvidia-modprobe --reinstall
sudo apt-get autoremove
hopefully that works and then check the opencl is installed.

or if it has installed, BOINC will not see the GPU unless you install nvidia-modprobe.
Image
The best form of help from above is a sniper on the rooftop....
User avatar
Bryan
Boinc Brigadier
Boinc Brigadier
Posts: 2621
Joined: Thu May 21, 2015 6:18 pm

#30 Re: Collatz GPU optimisations

Post by Bryan »

If you get into a bind and have to reinstall Linux you can save all your BOINC WU. Copy the /var/lib/boinc-client folder to a USB stick. After you get the new Linux installed then install the same version of BOINC. Once you have it running shut it down and overwrite the boinc-client with the one that you saved.
Image
davidbam
General Bitchin'
General Bitchin'
Posts: 6371
Joined: Wed Aug 15, 2018 1:15 pm
Location: Huntly, Scotland
Contact:

#31 Re: Collatz GPU optimisations

Post by davidbam »

Awesome, thanks. I did wonder about that but felt the chances of success were ... well ... limited
I think this is fool-proof but could you just try it for me please? • There are 10 types of people in the world; those who understand binary, and those who don’t
Image
Post Reply Previous topicNext topic

Return to “Collatz Conjecture”