troublesome

[Bryan]

I don't know if this is pertinent information or not. Mumps, a huge cruncher for SUSA, is a big time Linux user. I know that he has had problems with older AMD cards and the newer Linux kernels. Apparently there were "hooks" in the older linux distributions that the driver relied upon to make them work. Those hooks are no longer available so the old drivers won't work on the 18.x kernels. AMD is no longer supplying drivers for the older cards that will work with the new kernels. I know he has some machines running linux 14.0.x because he can't get the old cards to work on later versions of Linux.

[davidbam]

Thanks both.

I'll try apt install on one of the machines and report back. I didn't even know that you could get stderr messages on each WU - but I do now

[Dirk Broer]

I don't know if this is pertinent information or not. Mumps, a huge cruncher for SUSA, is a big time Linux user. I know that he has had problems with older AMD cards and the newer Linux kernels. Apparently there were "hooks" in the older linux distributions that the driver relied upon to make them work. Those hooks are no longer available so the old drivers won't work on the 18.x kernels. AMD is no longer supplying drivers for the older cards that will work with the new kernels. I know he has some machines running linux 14.0.x because he can't get the old cards to work on later versions of Linux.

The developer that worked at the so-called 'FGLRX' code has left AMD and AMD gave in to the Linux community, that screams for open drivers. Sadly the developers for that code do not see BOINC as one of their objectives and instead only seem to focus on recent gaming support. They also lack the drive to make older AMD cards work.
Basically we're at the same point as OS/2 video support at the end of the previous century: when you want a card to behave for certain software, you have to program the driver for it yourself.
I had an IBM colleague who could do that, write his own drivers for OS/2...the difference is that he made new cards work on an 'old' OS (OS/2 Warp 4 was not exactly old in 1998-99), while we want 'old' cards to work in a new OS.

[davidbam]

Okay, the library was already there (different machine, same AMD card)

Code: Select all

$  sudo apt install ocl-icd-libopencl1
[sudo] password for david: 
Reading package lists... Done
Building dependency tree       
Reading state information... Done
ocl-icd-libopencl1 is already the newest version (2.2.11-1ubuntu1).
ocl-icd-libopencl1 set to manually installed.
0 to upgrade, 0 to newly install, 0 to remove and 3 not to upgrade.

The first PrimeGrid jobs have completed, validated and scored Different sub-project though (Genefer 15 v3.19 (openclatiGFN15))

I'll try a few other combos when I regain the will to live (I've had a long day on non-boinc things) as, to be perfectly honest, Collatz is working away without any input from me at all. Clearly though, if PrimeGrid arises in a sprint, I'd like to have GPUs working too

[davidbam]

Oh!! Primegrid pays over double what Collatz pays !!

I have regained the will to live

[Bryan]

Oh!! Primegrid pays over double what Collatz pays !!

I have regained the will to live

And Nvidia will blow the doors off of AMD on either project

[davidbam]

Credits per ££ is another matter - mind you an RX580 is only quarter of the price of an nVidia GTX1080 at the moment.

Oh!! Primegrid pays over double what Collatz pays !!

This is true only on R7-350. WU from the same Genefer 15 v3.19 project take roughly the same amount of time to run on an RX580 so they must not be able to take advantage of its superior hardware (and hence pay poorly)

[Bryan]

Oh!! Primegrid pays over double what Collatz pays !!

A stupid question, are you aware that you can optimize your card on Collatz and usually dramatically increase your credits?

I don't know about AMD cards, but there is no way that PG gives higher credits on outpay Collatz on Nvidias. My GTX 1080Ti does a little over 1M/day on PG and 10M/day on Collatz.

[davidbam]

Not a stupid question at all. No, please tell me more

[Bryan]

I was editing my previous message and you replied so you may not have seen this:


I don't know about AMD cards, but on Nvidias there is no way that PG gives higher credits than Collatz. My GTX 1080Ti does a little over 3M/day on PG and 10M/day on Collatz. The highest paying project on PG is PPS Sieve.

In the Collatz project folder, on your BOINC client (thesonntags.com), there is a file named config.xml. You place or change the settings in that file so the executable performs differently for your particular GPU. It can make a HUGE difference.

See post number 71 HERE and you may find exactly what is needed. If not, then scroll through the thread and see if you can find it.

When you put that file in place "config.xml" the very next WU that is launched will pick it up and use those settings. If you want to try changing something then make the change and the next WU to start will use those changes. No need to stop and restart BOINC. You can make changes on the fly and then look at the completion times of the WU and see if it got better or worse.

[davidbam]

Right! Sounds like a rainy-day project to get that optimised. Many thanks for the link

[Bryan]

Right! Sounds like a rainy-day project to get that optimised. Many thanks for the link

You're more than welcome ... good luck with it

We've got to keep in mind that you are tilling new soil. Some of this stuff is 2nd nature to us old farts and we don't necessarily think to pass on the knowledge as early as we should. Our bad!

[Alez]

For my 2 x 980's I'm running this and twice as fast as standard

["app_config.xml"]

<app_config>
<app>
<name>collatz_sieve</name>
<max_concurrent>2</max_concurrent>
<gpu_versions>
<gpu_usage>1.0</gpu_usage>
<cpu_usage>0.5</cpu_usage>
</gpu_versions>
</app>
</app_config>

use notepad to make this file and save as other. For a single card change <max_concurrent>2</max_concurrent> to <max_concurrent>1</max_concurrent>

For config

verbose=0
kernels_per_reduction=48
threads=8
lut_size=17
sleep=0
cache_sieve=1
reduce_cpu=0
sieve_size=30

This file is collatz_sieve_1.30_windows_x86_64_opencl_nvidia_gpu.config andwill be empty. Linux will obviously be slightly different name and you will be looking for the AMD version but in var/lib/boinc/projects/boinc.thesonntags.com_collatz

For different cards look HERE and pick closest.

edit using gedit. In terminal use gksudo gedit and then navigate to file. If you dont have gedit then first off sudo apt-get gedit wlll do the trick. Gdit is standard in ubuntu but not my prefered Lubunu.

[davidbam]

Thanks, trying this now - on rx580

[davidbam]

Yup - big difference on the rx580 / ubuntu, thanks. Run-times cut by a third, in fairness points awarded cut slightly too tho. Provided they have enough WU to run 24/7, I should earn about 3.3M per day per card. I'm a happy bunny.

Still waiting for the first WU to arrive for the nVidia GPU. I do have it enabled in preferences so not quite sure why.

ETA: This is why (from event log)... Thu 01 Nov 2018 11:41:08 GMT | collatz | Not requesting tasks: don't need (CPU: job cache full; NVIDIA GPU: job cache full). Be patient David

[davidbam]

Three days later and Collatz hasn't sent me any work for the nVidia.

Event log says "collatz | Not requesting tasks: don't need (CPU: job cache full; NVIDIA GPU: job cache full)". Thing is, I am getting GPU work from other projects which have a lower resource share though.

[Alez]

The problem you are facing is the points given by Collatz. BOINC bases which projects you get units based on resource share you set but mainly RAC. Because Collatz pays so well, it will take ages for other projects to catch up and so Collatz will get no work. I basically only set one GPU project I want to get units for, at the moment Collatz. If I want to run Primegrid then I set Collatz to no new work and set Primegrid to receive new work. After all these years the BOINC manager is still dumb as a box of rocks in reality because it creator tried to have all projects give the same credit. As they don't the logic just doesn't work.
The best option is to set a GPU project you wish to crunch and add a second GPU project as a backup ( if you wish ) with a share of 0. That way it will only get work if your chosen project has no work. Alternatively, if you check your systems regularly as I do, then simply set one project at a time. I do the same for CPU's. I may run different projects at the same time, but I do it on different systems so I have control of what I crunch.
Alternatively you can do it with resource share but you end up setting ridiculous share percentages like 10,000 for Collatz to get any work. Micro-managment works best.

[davidbam]

you end up setting ridiculous share percentages like 10,000 for Collatz to get any work

SPOOKY - that is exactly the value I have set

[Alez]

you end up setting ridiculous share percentages like 10,000 for Collatz to get any work

SPOOKY - that is exactly the value I have set

Not my first rodeo I've seen many different versions of the manager since I started doing this and they all suffer from the same issues due to dogma.

[Bryan]

I rarely run my GPUs but when I do, I setup a 2nd instance of BOINC. One runs the CPUs and the other for the GPUs. That way the BOINC scheduler doesn't have to actually try to think ... it isn't very good at that

[davidbam]

[mention]Bryan[/mention] That sounds interesting. Would you happen to have a link please for how to do this under Linux?

How does boincmgr figure out which to connect to when both are on localhost?

[Bryan]

They run separate and independent. On one you enable your CPU project and on the other you enable GPU. If you attach to the same project, ie PG, then the project will see it as two different computers with both having the same name.

The method for setting up multiple BOINC clients is on TSBT somewhere ... I have trouble finding stuff like that. Maybe Alez can find it

EDIT: I found the one for Windows ... if someone can't find one for LInux then I'll write one up.

[davidbam]

My GTX 1080Ti does a little over 3M/day on PG and 10M/day on Collatz. The highest paying project on PG is PPS Sieve.

I finally got some Collatz work for nVidia by having a machine dedicated to running GPU (no CPU work). Using a GTX960 at the moment but I'll build a machine capable of taking both this and the GTX1080

My question is therefore, with 2 non-identical cards will I see two collatz_sieve_1.40_x86_64-pc-linux-gnu__opencl_nvidia.config files to allow me to optimise them independently ?? Maybe I should wait and see what happens and only optimise them after both are running together

[Alez]

Nope, only 1 collatz_sieve_1.40_x86_64-pc-linux-gnu__opencl_nvidia.config.

[Bryan]

I hear it is difficult to get multiple GPUs that aren't identical to run in Linux. I can't speak from experience on this.

If you can get it working where BOINC recognizes both cards then you can setup 2 BOINC instances with a instance controlling only a single card. Then you can have 2 config files available.

[davidbam]

I should probably learn to crawl before I start trying to run. Either that or sell/buy hardware to reduce variety

[davidbam]

Next question is whether the speed of CPU materially affects the performance of the GPU?

Background : I couldn't get nVidia card working in a Xeon / x79 motherboard (more questions on that later), so I moved it into an old HP machine (Pentium G3250 processor). This works but I am underwhelmed by the first results - it would maybe manage 220,000 credits per day on Collatz. Is it maybe related to the number of PCI lanes supported by CPU ??

Out of interest, an ageing ATI/AMD r7-350 in the HP machine manages about 340,000 per day on Collatz

[Bryan]

Yes, the speed of the CPU can affect the results a GPU gets. As you mentioned the PCI lanes etc but it is more the processing speed.

Most GPU projects also use the CPU. The CPU gets things ready and then passes it to the GPU to crunch. When the GPU finishes crunching that batch it contacts the CPU to get the next batch. This happens over and over until the WU is totally finished.

I have 2 GTX 1080Tis in different machines. One is in a I7-3930K running at 4.2G w/ PCIe 2.0. The 2nd is in a E5-2683 V4 server running at 2.7G w/ PCIe 3.X. The GPU in the I7 machine always outperforms the other GPU even though it has PCIe 2 vs PCIe 3.

Also remember that on most GPU projects you need to reserve some CPU threads otherwise the GPU winds up waiting to be serviced by the CPU.

[davidbam]

This file is collatz_sieve_1.30_windows_x86_64_opencl_nvidia_gpu.config andwill be empty. Linux will obviously be slightly different name

Is anyone crunching Collatz on nVidia under Linux please? I am beginning to think that Collatz doesn't use CUDA & that Linux drivers don't support OpenCL on nVidia.

I'd be DELIGHTED to be proved wrong though

[davidbam]

scratch that - I have that running myself on one machine.

Config file is called : collatz_sieve_1.40_x86_64-pc-linux-gnu__opencl_nvidia.config proving that opencl is used

It is only doing about 225K credits per day on Collatz though (unoptimised and fitted in a fairly old machine)

ETA : I can't remember which boinc client I loaded as I see there is a boinc-client-nvidia-cuda. Back to the drawing board