Posts by Acey Pilot
log in
1) Message boards : News : Phasing out Multi-Size (Message 19208)
Posted 1363 days ago by Acey Pilot
Getting GPU utilization of only 63% with first large work unit. Large work unit GPU config has same settings as Solo Callatz, which is running fine at 99%. I have dual GPUs and have restarted system several times. Any Ideas?
2) Message boards : Number crunching : Optimizing Collatz v6.xx OpenCL and CUDA Applications (Message 18665)
Posted 1429 days ago by Acey Pilot
Yes the links cleared it up for me. Got it working now. Thanks Slicker.
3) Message boards : Number crunching : Optimizing Collatz v6.xx OpenCL and CUDA Applications (Message 18663)
Posted 1429 days ago by Acey Pilot
It has reduced my cuda task run times from about 4,610 seconds to
about 2,100


Give the GPU at least half a CPU in the config file, at present too many have hung up on that CPU time column and are trying to crush it down to a low value - the latter cripples the GPU, don't do it.

That CPU column does not at present represent reality as such at the moment, and the times inside it should not be used as indication of speed et al in any way shape or form. Most certainly do not go through hoops and loops crushing that time down, ignore the value inside the column for the moment. I am seeing some machines crushing it down to 10-15 secs, with a WU crunch time of 4/5/ thousand seconds. A good fast time for a high end 7xxx 6xxx card is around +/- 600 seconds, not thousands of seconds.

Slicker is aware of the issue, it has a long history as to why its like that, but for the present, ignore the value inside that CPU column its highly misleading, do not try and crush it down to a low value - in fact ignore the value until Slicker has time to work on it. For now use 0.5 as CPU in the config files as a minimum. I don't do much CPU WUs so I give it a full CPU per GPU, app_config below

<app_config>


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

Adjust to suit, but don't try and crush down the time in the CPU column, it will cripple the machine, ignore that column for now. Until that column is tweeked, just use app_config as an indicator, and accept the use of circa 0.5 CPU, if the classic 0.01 et al is inserted in the app_config, it will crush many machines.

Target for most mid to high end machines is circa 520 to 700 secs (ish). If a figure of thousands is spotted, the config file is way out and needs adjusting as above.


Which folder of which config file does this go in? For this new guy, what is complete file name for this above? I am already using the solo_collatz_6.06_x86_64-apple-darwin__opencl_nvidia_gpu.config for settings:
verbose=1
items_per_kernel=18
kernels_per_reduction=9
threads=8
sleep=1
build_options=-Werror

I want to allocate one cpu to each gpu as described above. Better yet, is there a global Boinc setting where one can allocate one complete cpu per gpu?

Is the xml file required, and if so does it go into the Collatz folder?

I do gpu WU only.

Thanks for your patience.
4) Message boards : Number crunching : Optimizing Collatz v6.xx OpenCL and CUDA Applications (Message 18648)
Posted 1430 days ago by Acey Pilot
xml files. Take the .txt off of the file if there is one. Also VERY important: when inside editing the file, make sure the formatting is PLAIN TEXT! If it is rich text, your cc_config file will not work even with the .xml file type.

Then restart Boinc or read the config files under advanced. You should then see a difference.

Best of luck.
5) Message boards : Number crunching : Optimizing Collatz v6.xx OpenCL and CUDA Applications (Message 18646)
Posted 1430 days ago by Acey Pilot
Each Collatz 4.07 application is distributed with an empty config file. The config file has the same name as the executable but with the extension ".config".

There are a number of parameters that can be altered to improve speed or video response or to aid in solving issues. They are:

verbose=[0|1]
A value of 1 causes more information about the GPU, OpenCL version, etc. to be written to the log file. If enabled, this should be the first line of the config file so that it will report the other settings in the log file.

items_per_kernel=[10..22]
The number is the power of two 256-bit numbers (e.g. 2^N) that will be calculated per kernel call. Setting this number higher places a larger load on the GPU. Setting the number too high WILL cause the driver to crash and the application to hang. The default is 14, or 2^14, or 16384 items.

kernels_per_reduction=[2..9]The number (2^N once again) of kernels to run before doing a reduction. The default is 8 or 2^8 = 256. A lower number can improve video response. A larger number may result in a higher GPU load. Too high a number will result in CPU as well as GPU utilization.

threads=[5..10]
This contains the number of work groups to run in parallel. Higher is not necessarily faster. This number is device dependent. If set too high, the application will automatically reduce it to a value compatible with the device.
Most AMD GPUs allow up to 256 (a setting of 8). NVidia GPUs may allow 512 or even 1024 (a setting of 9 or 10). OpenCL requires a minimum of 32 (a setting of 5) according to the Khronos specifications.

build_options=[string containing any optional OpenCL build options]
This was added strictly for debugging in order to be able to use "-cl-opt-disable -Werror". If the OpenCL application crashes within 1-2 seconds of starting, you may want to use "build_options=-cl-opt-disable -Werror" and see if that fixes the problem.

sleep=[1..1000]
This controls the number of milliseconds that the application goes into a sleep state while waiting for the asynchronous kernel calls to complete. The default is 1. Setting this higher (e.g. 2-5) will result in better video response but will slow down the application considerably.

The config file will be renamed to collatz.config when it is copied to the BOINC slot folder when an application starts running. Exiting BOINC and editing the version in the project folder will not change the settings of the applications in progress as their config is taken from the slot folder.

A sample collatz.config file looks like:

verbose=1
items_per_kernel=20
kernels_per_reduction=9
threads=8
sleep=1
build_options=-Werror


Since the workunits very somewhat in the number of total steps they produce, I would suggest that you run several and take the average runtime to determine whether one set of values in the config works better than another set.

Note: The values in the sample above work quite well on my HD 6970 and HD 7970 without making either too sluggish.



Just to confirm: So now that we have the Cuda option in our account settings, do these settings still apply? In other words, are these settings only applicable to Open CL work units?
6) Message boards : News : Have your cake and eat it too! (Message 18645)
Posted 1430 days ago by Acey Pilot
While most people wanted only CUDA applications, some had some good reasons for using OpenCL applications which means it really can't be one or the other. As a result, you can use whichever you want or even both by selecting your preferred plan class(es) for each venue.

The Windows and Linux CUDA applications, both 32 and 64 bit, have been released so your preferences should take effect the next time your computer requests work. The default is that all plan classes are selected.




EXCELLENT WORK! THANK YOU SLICKER!
7) Message boards : News : CUDA vs nVidia OpenCL User Poll (Message 18588)
Posted 1433 days ago by Acey Pilot
Solution A or manual choice in the preferences

Slicker, I would like to thank you again for all your efforts !

Best Regards,

Philippe



My thoughts exactly, but in my own case, I find the Cuda is faster and uses less wattage. It is simply more efficient all the way around, but then that is just me.

Thanks Slicker. Regards,
William




Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.