PDA

View Full Version : New client version: 2.9102-508



alpha
01-14-2009, 03:29 AM
There is a new client version available on the download page (http://www.distributed.net/download/clients.php) for lots of operating systems. Changelog:


new: x86: OGR-NG mmx core
new: cellbe: OGR-NG SPU C core
new: cuda: RC5-72 core for nVidia CUDA video cards (#4030)
new: haiku: Added support for Haiku operating system
fix: beos: Fix bug in network code preventing server communication
new: cellbe: OGR-NG SPU ASM core
imp: x86: Identify AMD "Unknown" processors (#4106)
fix: x86: Pentium M naming (#4075)

Make sure you benchmark new cores and select the fastest. :thumbs:

alpha
01-14-2009, 04:17 AM
A quick benchmark of the new client shows I'm at ~68 Mnodes/sec up from ~50 Mnodes/sec. That's a 36% increase in performance, nice!

IronBits
01-14-2009, 04:27 AM
Mine didn't auto detect CUDA. :(

Death
01-14-2009, 05:59 AM
afaik you must have cuda 2 drivers installed, not cuda 2.1
maybe that's an issue?

IronBits
01-14-2009, 06:53 AM
erm, don't think so cause in the 'other directory' where I have the previous 'cuda' version, it works fine.

What core do I need to use to force CUDA?
It was 9 in the windows 'cuda' version...

[Jan 14 08:00:47 UTC] Automatic processor type detection found an Intel Core 2/Extreme/Xeon processor.
[Jan 14 08:00:47 UTC] RC5-72: using core #10 (GO 2-pipe alt).

Digital Parasite
01-14-2009, 07:04 AM
The new speedup in the MMX core (#2) means that my Intel systems are selecting that over the asm-generic core now, nice speed improvement. This should help the project finish even faster which looks like it will be done in record time.

alpha
01-14-2009, 08:40 AM
IB: if you go to the dnetc interactive config and choose "performance related options" and then "core selection", it should list all the cores available. If a CUDA core isn't listed, maybe something has gone wrong somewhere? :confused:

DP: Yeah, I've tried two different boxes now and seen a 30-35% improvement in both cases, and the new fastest core was selected automatically, in both cases. As you say, this must be the fastest (dnet) OGR project to date! Onwards with OGR-27.

IronBits
01-14-2009, 09:29 AM
There is no CUDA core selection in this new client. :cry:

Death
01-15-2009, 06:19 AM
try
-multiok=1

it should work with both cpu and gpu.

please post your video card model, and some log.

Windows 32bit
1 [x86/CUDA] v2.9102.508b2 (beta) 2008-12-24 http | ftp
1. nVidia CUDA video card accelerated client. RC5-72 only. CUDA 2.0 drivers must be pre-installed (not CUDA 2.1 or 1.1). Beta expires in 28 days. Visit bug 4030 for details. See readme.cuda

distributed.net client for nVidia CUDA-compatible GPUs
document revision $Id: readme.cuda,v 1.6 2008/12/24 23:15:37 jlawson Exp $

Welcome to the distributed.net client.

This document covers information specific to the client for the nVidia
CUDA-capable video card GPU. Refer to other enclosed documentation or
browse the online FAQ at <http://faq.distributed.net/> for
non-platform-specific documentation.

1.0 Getting started
2.0 nVidia CUDA specific notes
3.0 Troubleshooting and Known issues

1.0 Getting started ------------------------------------------------

Just unpack/unzip/untar the client in a directory of your choice and
fire it up.

If you have never run the client before, it will initiate the
menu-driven configuration. Save and quit when done.

Then, simply restart the client. From that point on it will use the
saved configuration.

The configuration options are fairly self-explanatory and can be run
at any time by starting the client with the '-config' option.
A list of command line options is available by starting the client
with '-help' or '--help'.

A complete step-by-step guide to running your first client is
available at <http://www.distributed.net/docs/tutor_clients.php>

2.0 nVidia CUDA specific notes ------------------------------------

This client requires nVidia CUDA-capable hardware and appropriate
drivers to be installed. For a list of GPU hardware that supports
CUDA, see <http://en.wikipedia.org/wiki/CUDA#Supported_GPUs>

When installing the CUDA drivers and toolkit, be sure to download
the versions for CUDA 2.0 (and not version 2.1 nor 1.1). Visit
<http://www.nvidia.com/object/cuda_get.html>

At the moment, our CUDA clients only provide support for RC5-72.
Due to the nature of OGR, it is difficult to parallelize in a way
that can make use of the CUDA architecture.

Our CUDA clients also only execute crunchers on the GPU. In order
to utilize the CPUs on your computer, you will need to download
and run another instance of the standard client from a separate
subdirectory.


2.0 Troubleshooting and Known issues ------------------------------

We recommend using the nVidia driver version 178.xx (for use with
CUDA 2.0), since we have received reports of the client hanging
with driver 180.xx (which is intended for CUDA 2.1).


On Windows Vista platforms, the CUDA client will not be able to
access the GPU when run as a Service due to operating system
limitations. The current workaround is to run the client
interactively as a logged-in user instead.


On Linux platforms, you may see errors when running the CUDA
client on a text-mode system without X11 (init level 3 not 5).
This usually occurs because the nVidia module has not been
loaded or the /dev/nvidiactl control file has not been
initialized. To make your RHEL/CentOS system automatically
prepare these things without starting X11, see
<http://forums.nvidia.com/lofiversion/index.php?t52629.html>
Similar solutions will be necessary for other distributions.

You may also encounter problems on Linux platforms if you try
to run the CUDA client as a non-root user that does not have
permission to access the /dev/nvidiactl control file.
Typically membership to the "video" group is used to manage
access, so you will need to ensure that your UNIX user is a
member of it. If uncertain, view the file permissions of
the /dev/nvidiactl control file.


If you are unable to execute the dnetc binary because of a missing
libcudart.so library, you will need to visit the nVidia website
and download the "CUDA toolkit 2.0" first. After installing, you
may need to either set your LD_LIBRARY_PATH environment variable
to its lib directory prior to running dnetc.