Join PrimeGrid
Returning Participants
Community
Leader Boards
Results
Other
drummers-lowrise
|
1)
Message boards :
Number crunching :
Linux running on an AMD x86_64 or Intel EM64T CPU 1.35 (ATI13ATI)
(Message 29377)
Posted 4480 days ago by Skip Da Shu
Anybody got this working? Is there any "how-to" avail on what needs to be installed?
<core_client_version>6.10.58</core_client_version>
<![CDATA[
<message>
process exited with code 127 (0x7f, -129)
</message>
<stderr_txt>
../../projects/www.primegrid.com/primegrid_tpsieve_1.35_x86_64-pc-linux-gnu__ati13ati: error while loading shared libraries: libOpenCL.so: cannot open shared object file: No such file or directory
</stderr_txt>
]]>
|
2)
Message boards :
AP26 - AP27 Search :
CUDA application released
(Message 19753)
Posted 4846 days ago by Skip Da Shu
New server code is compiling as I write this message. It is the latest and greatest (yeah, right) SVN version, with CUDA23 plan class modified to require only 250MB RAM, so less powerful cards should get GPU work soon.
I will also update web code to the same revision, hoping that it fixes CPU work issues that some people have been having.
Yup, one of my 9600GSO (reports as 383MB) cards picked up some is running them now.
|
3)
Message boards :
AP26 - AP27 Search :
CUDA application released
(Message 19752)
Posted 4846 days ago by Skip Da Shu
One other thing you might try...
CAN you install the older nvidia driver from the repositories... or in other words does the restricted hardware drivers applet recognize the NVID card being present at the desktop?
If it does there are some other options.
PS: And as I recall I had to do this on a live keyboard/mouse/monitor instead of via VNC or SSH.
|
4)
Message boards :
AP26 - AP27 Search :
CUDA application released
(Message 19689)
Posted 4848 days ago by Skip Da Shu
RHEL is Red Hat Enterprise Linux, or CentOS. I was unaware of the re-nicing feature in Ubuntu. Thanks for posting a possible fix for this "feature" that they have added. I don't see the logic behind schedtool being added to the OS. It apparently breaks the CUDA blocking sync feature. Anyone running Ubuntu, if you can try this, and post results.
And once again, re: Windows slowdown, I am aware of this, and it was the only way to get the app to run in Windows without using an entire CPU core, at this time. CUDA's CPU-thread blocking sync is "broken" in Windows, like Ubuntu, with a fully loaded CPU. If your GPU processes AP26 faster than the CPU, then it may be worthwhile for you to disable a core in BOINC and the GPU should run at full speed.
Thanx, I didn't know the acronym.
I also would be real interested for someone else to try it. It'd be nice if the fix is really this simple.
Not sure I'd really call it part of the OS as if you install boinc from Berkeley you get a different set of scripts that don't have this. It's all a part of the scripts that come from the Debian boinc_client package. Which generally I like as it adds the Debian standard command line options for status, stop, start, restart (that works). Last time I was using the Berkeley scripts the reload/restart didn't actually work (in the Dotsch/UX distro).
I'll post a reference to this post on the Debian boinc_devs list in case any of them want to take a look into it. I'm not sure of their current position on GPU support though.
The script code that actually does it is a bit over my head. Maybe it'll mean something to you. From the run level 2 started /etc/init.d/boinc-client script:
schedule()
{
log_begin_msg "Setting up scheduling for $DESC and children:"
if ! is_running; then
log_progress_msg "$NAME not running"
else
if [ ! -x "`which ionice 2>/dev/null`" ]; then
log_progress_msg "ionice not found,"
else
if ionice -c 3 -p $pid 2>/dev/null; then
log_progress_msg "idle,"
else
log_progress_msg "ionice failed,"
fi
fi
if [ ! -x "`which schedtool 2>/dev/null`" ]; then
log_progress_msg "schedtool not found"
else
children=`ps --ppid $pid -o pid= | tr '\n' ' '`
(schedtool -n 19 -D $pid $children >/dev/null \
&& log_progress_msg "idleprio") || \
(schedtool -n 19 -B $pid $children >/dev/null \
&& log_progress_msg "batch") || \
(schedtool -n 19 -N $pid $children >/dev/null \
&& log_progress_msg "normal")
fi
fi
log_end_msg 0
}
|
5)
Message boards :
AP26 - AP27 Search :
CUDA application released
(Message 19684)
Posted 4849 days ago by Skip Da Shu
I'll address some of the issues I see posted:
* Not enough VRAM message: >384 is a BOINC app plan CUDA 2.3 requirement, and since the application requires 2.3, 256MB cards like my 8500gt will have to run via app_info.xml. The CUDA AP26 program only uses about 100mb VRAM while running. My 9600GSO cards report as 383MB so it'll take the server side change u suggested before they get any GPU work from Prime.
* Slowdown in certain Linux distros with fully loaded CPU: This appears to be an issue with CUDA and certain Linux versions. RHEL, for example, runs the application as it should, with no slowdown. It appears some distros are not scheduling the GPU thread as it should. I hope this issue will be fixed in the upcoming CUDA 3.0 toolkit.
I Don't know what distro RHEL is (thought I bought a tent and some mosquito repellent there once) but COULD this be the re-nicing that Debian/Ubuntu uses (schedtool, since about v8.10 in Ubuntu) to minimize the impact BOINC has on the desktop?? Since I have that rescheduling turned OFF on my dedicated Unbuntu/Ubuntu based crunchers I may not be seeing this or it's small enough that it's not bothering me.
Don't GTX-260 times of around 8 minutes with all CPU cores working suggest that I do not have this issue? Around 8:35 for a GTS-250 w/ all other cores running.
And then:
9800GT w/ all CPU cores running 10:05 and 10:34
9800GT w/ no CPU cores running 10:30
Something to try for those that are seeing this under Ubuntu/Xubuntu/Kubuntu/Mint/Debian using the standard Debian boinc-client scripts (/etc/init.d/boinc-client):
sudo gedit /etc/default/boinc-client
Change the SCHEDULE from the default of "1" to "0" and see if that fixes the issue after a restart.
# Set this to 1 to enable advanced scheduling of the BOINC core client and
# all its sub-processes (reduces the impact of BOINC on the system's
# performance).
SCHEDULE="0"
Then restart w/:
sudo /etc/init.d/boinc-client restart
If it does, then you have a decision to make based on the usage of the machine in question. The difference between having it on or off isn't earth shattering at the desktop.
If it doesn't resolve it... oh well, it was just a guess on why I don't think I'm seeing it on a variety of quad Intels (Q66, Q67, Q94, Q9550s) under various Ubuntu versions/varients.
Zai Jian
|
6)
Message boards :
AP26 - AP27 Search :
CUDA application released
(Message 19681)
Posted 4849 days ago by Skip Da Shu
The runtime estimate is an order of magnitude too low. First estimate was 59 sec for a full task. Reality on this system is 13 minutes. The low estimate plays havoc with other subprojects; a recently started CWSieve task now displays 31 hours to completion (it's more like 2.5 hours). Please fix this. Mine straightened out the estimates after it finished 1 WU.
|
7)
Message boards :
AP26 - AP27 Search :
CUDA application released
(Message 19680)
Posted 4849 days ago by Skip Da Shu
Under Ubuntu 9.10 (amd64) a AP26 GPU WU needs nearly one hour to complete on a GTX260 with 4 AP26 WUs running concurrently on the quad core CPU. If it runs alone it only needs about 8 minutes to complete. I currently see this behaviour with the released app too.
Ditto on a v9.04, same set up. 1st two: 7:55 and 8:04 w/ various other stuff (NFS mostly) running on the other CPU cores.
|
8)
Message boards :
Number crunching :
Any plans for GPU crunching (under Linux)?
(Message 16658)
Posted 5005 days ago by Skip Da Shu
Was just wondering if PrimeG has any plans to develop GPU apps (Linux in particular) anytime this year? Did a search on the forums and didn't find any chatter about it.
|
9)
Message boards :
Proth Prime Search :
PPS Sieve
(Message 16355)
Posted 5023 days ago by Skip Da Shu
I'm also interested in the Proth Prime Search sieve. Would you make it available in my PrimeGrid preferences.
Also, is the following correct for the app_info.xml
<name>pps_sr2sieve</name>
<user_friendly_name>Proth Prime Search (Sieve)</user_friendly_name>
I have a working sr2sieve on my linux running PPC64.
Thanks
Greg Trigg
Ditto
|
10)
Message boards :
Problems and Help :
64-bit OS crunching speed
(Message 9805)
Posted 5352 days ago by Skip Da Shu
See if "kerry beagle" is running. If so, turn it off.
It Is not :-/ Anyway, thanks for the hint.
However, behaviour is somewhat strange.
Checking with atop I see both processors running the primegrid_psp_sieve job at averaged 99% each.
Little of CPU distributes for firefox, Xorg and boinc_client etc.
So I guess there is nothing 'hijacking' CPU power.
Is there a way to make absolutely sure that the CPU is not throttled down by any daemon or for whatever reason...?
Regards
Jochen
try doing sudo head /proc/cpuinfo
See what it reports the CPU MHz as. This will AT LEAST be the full CPU frequency.
Now stop the BOINC daemon with sudo /etc/init.d/boinc stop .
Then do the 'head' again and see if it's dropped.
Like all things Linux there's probably 7 more ways and mine may or may not be effective as I don't have any APIC power controls enabled on my Xubuntu crunchers.
|
Next 10 posts
|
