mfaktc: a CUDA program for Mersenne prefactoring

[rebirther]

Is it possible to get the real Multicore feature in the near future?

[James Heinrich]

Originally Posted by rebirther

Is it possible to get the real Multicore feature in the near future?

Which cores are you talking about? Certainly mfaktc will use all available CUDA cores, and (assuming you're not CPU-sieving, because why would you) shouldn't put any significant demands on any CPU core.
Are you referring to high-end GPUs not being fully utilized by a single mfaktc instance?

[kriesel]

Originally Posted by rebirther

Is it possible to get the real Multicore feature in the near future?

That depends on what you mean by possible, by get, by real multicore feature, and by near future.
It's possible to get such a thing if one has the ability, dives into the task, and successfully codes and builds it.

Did you mean multicore & multithreaded on the cpu side to better support the really fast gpus, one gpu per program instance, one instance per gpu, fully loading the fast gpu?
Something else?

TheJudger hasn't posted in this thread that I saw since July 28, but it looked promising then that a v0.22 was in development, and a CUDA 11 build done and A100 performance shown.. https://www.mersenneforum.org/showpo...postcount=3280

If/when he makes a release is to be determined, as is the list of any new features.

[pinhodecarlos]

No, what he means is one instance of the client been able to detect all GPU cards. Boinc users have more than one GPU per computer, sometimes 4 or even more.

What happens at the moment is the client only detects one GPU and it can allocate to this GPU to run 2 or 3 tasks in parallel to get the most from the GPU but it won’t detect if the system has more GPU’s on it to spread out work. Limitation is on the client side, not on BOINC side.

Edit: http://srbase.my-firewall.org/sr5/fo...stid=5765#5765 (client only recognises one device available on the computer)

[kriesel]

Originally Posted by pinhodecarlos

No, what he means is one instance of the client been able to detect all GPU cards. Boinc users have more than one GPU per computer, sometimes 4 or even more.

What happens at the moment is the client only detects one GPU and it can allocate to this GPU to run 2 or 3 tasks in parallel to get the most from the GPU but it won’t detect if the system has more GPU’s on it to spread out work. Limitation is on the client side, not on BOINC side.

Edit: http://srbase.my-firewall.org/sr5/fo...stid=5765#5765 (client only recognises one device available on the computer)

Sounds to me like an issue with the BOINC client. (Note I've never used BOINC.)

Mfaktc takes its marching orders what gpu to use from its command line input. I and many others are running multiple instances on each of multiple gpus per system, in almost any GIMPS gpu application. Syntax for specifying device is different for mfakto because of OpenCl's platform/device-on-platform approach (AMD, NVIDIA, Intel drivers are separate platforms). I have run multiple instances of or mixed combos on a single gpu, of CUDALucas, CUDAPm1, mfaktc, and gpuowl. (Haven't run much mfakto, except on igps, where one is enough.)

Modifying mfaktc from single instance on single gpu, to multi-gpu x multi-gpu-load in a single application, somewhat analogous to prime95's multiple workers support, and adding automatic Mfaktc-compatible-gpu detection and enumeration, would be a large undertaking. The prime95 program has the advantage that all the cores in a cpu are the same. There's no such given regarding gpus installed in a system. Most of my multi-gpu systems are heterogenous.
There are some rough edges on prime95's coping with cpu models' fft length limitations. Something analogous is expected for an mfaktc detection and configuration code section; what Compute Capability, what GPUSieveSize, etc does a given gpu model support? The list is frequently increasing with new models.

The optimal tuning varies depending on the gpu model/exponent/bitlevel combination. And that tune is currently resident in mfaktc.ini; it is derived by manual tuning iterations and manually entered, not by automated program benchmarking as in prime95.

[pinhodecarlos]

You first sentence is that your opinion or Oliver’s? Edit: you edited your post so my question regards your third paragraph.

Edit 2: it is not BOINC limitation since there are a bunch of projects running this way, multi-GPU detection.
Edit 3: you’re giving me reason, client limitation, each instance needs to tune its parameters in function of the GPU type.
Edit 4: can we get in touch with Oliver to understand from his point of view this feasibility for multi-GPU enhancement or was this already discussed somewhere here?

Really thank you for your support.

[kriesel]

From the link you gave in 3414:

ERROR: unknown option '--device' (not working, devs are informed)

<app_config>     <app>         <name>TF</name>         <max_concurrent>1</max_concurrent>         <gpu_versions>         <gpu_usage>1</gpu_usage>         <cpu_usage>0.01</cpu_usage>          </gpu_versions>     </app>     <app_version>         <app_name>TF</app_name>         <plan_class>opencl_ati_101</plan_class>         <cmdline>-d 00</cmdline>     </app_version> </app_config>

-d 00 means OpenCL style device specification, platform 0, device-on-platform 0.
I've seen -d 01 reference the same device as -d 00 in mfakto. See also mfakto help
I've seen 3 platforms report in on an OpenCL querying utility, AMD, Intel, NVIDIA on one system together, but don't have anything configured that way now. It was fragile. https://stackoverflow.com/questions/...es-on-a-system

On my i7-8750H/UHD630 with GTX1050Ti, the UHD630 is -d11 in mfakto.

set gpuappinstance=mfakto-uhd630
mfakto -d 11 >>mfakto.txt

echo At %date% %time% %computername% %gpuappinstance% exited >>\\server\mersennes\appwatch\%computername%-%gpuappinstance%.txt
set gpuappinstance=

-d 0 would be the form for a CUDA capable gpu, mfaktc, first gpu. mfaktc help

config.txt example for gpuowl v6.11, first gpu

-device 0 -user kriesel -cpu asr3/gtx1080 -yield -maxAlloc 6500 -proof 9

(old version gpuowl help) gpuowl -h will list the OpenCL available devices on a system in gpuowl device numbering, after the options and before the fft lengths, in any reasonably recent build.

CUDALucas, and CUDAPm1 (forked from CUDALucas) house device selection in an ini file or use the command line. again, numbering is zero-based.

# DeviceNumber is the same as the -d option. Use this to run CUDALucas on a GPU
# other than "the first one". Only useful if you have more than one GPU.

 DeviceNumber=0

Most of this is mfaktc on one system, 4 gpus, and note the number of processes per gpu-number:

>"c:\Program Files\NVIDIA Corporation\NVSMI\nvidia-smi.exe
Sun Oct 25 11:03:19 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 445.75       Driver Version: 445.75       CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 1650   WDDM  | 00000000:03:00.0 Off |                  N/A |
| 43%   58C    P0    43W /  45W |    796MiB /  4096MiB |    100%      Default |
+-------------------------------+----------------------+----------------------+
|   1  GeForce GTX 1650   WDDM  | 00000000:04:00.0 Off |                  N/A |
| 44%   52C    P0    45W /  45W |    796MiB /  4096MiB |    100%      Default |
+-------------------------------+----------------------+----------------------+
|   2  GeForce RTX 2080   WDDM  | 00000000:05:00.0 Off |                  N/A |
| 44%   64C    P2   123W / 125W |   1359MiB /  8192MiB |    100%      Default |
+-------------------------------+----------------------+----------------------+
|   3  GeForce GTX 108... WDDM  | 00000000:06:00.0 Off |                  N/A |
|  0%   58C    P2   125W / 125W |    726MiB / 11264MiB |     97%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      5496      C   ...\2\mfaktc-2047-win-64.exe    N/A      |
|    0   N/A  N/A      7764      C   ...tc\mfaktc-2047-win-64.exe    N/A      |
|    1   N/A  N/A      5224      C   ...\2\mfaktc-2047-win-64.exe    N/A      |
|    1   N/A  N/A      5752      C   ...tc\mfaktc-2047-win-64.exe    N/A      |
|    2   N/A  N/A      1568      C   ...\2\mfaktc-2047-win-64.exe    N/A      |
|    2   N/A  N/A      7436      C   ...tc\mfaktc-2047-win-64.exe    N/A      |
|    2   N/A  N/A      8644      C   ...\3\mfaktc-2047-win-64.exe    N/A      |
|    3   N/A  N/A      7772      C   ...4-g36f4e2a\gpuowl-win.exe    N/A      |
 +-----------------------------------------------------------------------------+

gpuowl is able to enumerate them all, on multiple OpenCL platforms, but uses a specified one in any gpuowl instance.

-device <N>        : select a specific device:
 0  : GeForce RTX 2080- not-AMD
 1  : GeForce GTX 1650- not-AMD
 2  : GeForce GTX 1650- not-AMD
 3  : GeForce GTX 1080 Ti- not-AMD
 4  : Intel(R) Core(TM) i7-4790 CPU @ 3.60GHz- not-AMD

(edit to follow shortly)

[kriesel]

or gpuowl handling multiple platforms multiple devices again,

-device <N>        : select a specific device:
 0  : GeForce GTX 1050 Ti- not-AMD
 1  : Intel(R) UHD Graphics 630- not-AMD
 2  : Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz- not-AMD

(edit time ran out for previous post while I bounced from system to system and dug further)

See also OpenCL-Z and other utilities. Maybe this.

It would have been good, if rebirther had meant in BOINC use in 3411, to specify.

[kriesel]

A small utility to list the OpenCL platforms and devices on them and a bit of description.
See https://www.mersenneforum.org/showpo...74&postcount=6

For CUDA there's nvidia-smi.

[rebirther]

Originally Posted by kriesel

or gpuowl handling multiple platforms multiple devices again,

-device <N>        : select a specific device:
 0  : GeForce GTX 1050 Ti- not-AMD
 1  : Intel(R) UHD Graphics 630- not-AMD
 2  : Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz- not-AMD

(edit time ran out for previous post while I bounced from system to system and dug further)

See also OpenCL-Z and other utilities. Maybe this.

It would have been good, if rebirther had meant in BOINC use in 3411, to specify.

We have tried every possible mapping in BOINC with no success.

[kriesel]

Originally Posted by rebirther

We have tried every possible mapping in BOINC with no success.

Examples? The developers have, or the end users have? How do you know it is not working?
The correct mappings work locally, without BOINC involved. Run times for gpuowl tasks are typically long.
OpenCL mapping is platform is zero-based, but device on a platform is not, apparently.
-d01 is the first platform of the first device in lsgpu and mfakto.

(Maybe some of this thread should be moved to a BOINC thread.)