Having recently installed Debian's AMD64 release on a partition of my system, I wanted to find out what difference a 64-bit vs. a 32-bit environment would make.
The platform is a fairly recent system with an ASUS A8V DeLuxe MB, and an AMD64 3500+ CPU @ 2200 MHz.
I have run various environments on this platform : a plain 32-bit environment, a 32-bit environment run under a 64-bit kernel, and a plain 64-bit environment. I have also run E@H under wine in a 32-bit environment ever since I learnt of the poor performance of E@H for Linux. Currenty, I use a x86_64 kernel for all environments, and it doesn't impact the WU CPU-secs in any noticeable way.
To run the BOINC benchmarks, I have downloaded and compiled boinc_public-cvs-2005-06-15 (BOINC-4.45) in each the environments, and run 'client/boinc -run_cpu_benchmarks' in the compilation directory. GCC 3.3.5 was used in both cases, BOINC was compiled as per the README file instructions.
Here are the results : -------------------------------
2005-06-15 23:13:43 [---] Starting BOINC client version 4.45 for i686-pc-linux-gnu
2005-06-15 23:13:43 [---] Data directory: /usr/src/other/boinc_public
2005-06-15 23:13:43 [---] Version Change Detected (4.27 -> 4.45); running CPU benchmarks
2005-06-15 23:13:43 [---] No general preferences found - using BOINC defaults
2005-06-15 23:13:43 [---] Remote control not allowed; using loopback address
2005-06-15 23:13:45 [---] Running CPU benchmarks
2005-06-15 23:14:43 [---] Benchmark results:
2005-06-15 23:14:43 [---] Number of CPUs: 1
2005-06-15 23:14:43 [---] 1299 double precision MIPS (Whetstone) per CPU
2005-06-15 23:14:43 [---] 3065 integer MIPS (Dhrystone) per CPU
2005-06-15 23:14:43 [---] Finished CPU benchmarks
2005-06-15 23:14:43 [---] Resuming computation and network activity
2005-06-15 23:14:43 [---] request_reschedule_cpus: Resuming activities
2005-06-15 23:14:43 [---] Insufficient work; requesting more
------------------------------------------------------
2005-06-16 01:27:58 [---] Starting BOINC client version 4.45 for x86_64-unknown-linux-gnu
2005-06-16 01:27:58 [---] Data directory: /usr/src/other/boinc_public
2005-06-16 01:27:58 [---] No general preferences found - using BOINC defaults
2005-06-16 01:27:58 [---] Remote control not allowed; using loopback address
2005-06-16 01:28:00 [---] Running CPU benchmarks
2005-06-16 01:28:57 [---] Benchmark results:
2005-06-16 01:28:57 [---] Number of CPUs: 1
2005-06-16 01:28:57 [---] 1667 double precision MIPS (Whetstone) per CPU
2005-06-16 01:28:57 [---] 3728 integer MIPS (Dhrystone) per CPU
2005-06-16 01:28:57 [---] Finished CPU benchmarks
2005-06-16 01:28:57 [---] Resuming computation and network activity
2005-06-16 01:28:57 [---] request_reschedule_cpus: Resuming activities
2005-06-16 01:28:57 [---] Insufficient work; requesting more
------------------------------------------------------
Since BOINC-4.45 for Windows will not install without an IExplorer (>=5), I cannot provide benchmarks for the Windows version of BOINC in my environment. [ Only Wine is installed, there isn't any Windows software on my box ].
The benchmarks above show that a given software can get a performance gain when executing in a 64-bit environmnent on a 64 bit platform, if it has been compiled for execution on such a platform. The 20-30% performance gain in WhetStones or DhryStones cannot be extrapolated from BOINC to E@H, but there would be a gain.
But this gain would be no match for what could be achieved by a proper port of the Einstein application to a Linux environment : on this platform, a current WU takes about 21Ksec in the Wine environment compared to 36Ksec in the native Linux environment.
For whatever reasons, the E@H application doesn't perform in the Linux environment and the team doesn't consider this a priority. So many Linux users run BOINC projects under Wine, but the workaround may have its own problems soon : the BOINC/E@H application seems to break in the newer Wine release (20050524), with a recurring "exited with zero status but no 'finished' file" condition which prevents any work to be done.
E@H for Linux - a non-topic ?
Copyright © 2024 Einstein@Home. All rights reserved.
BOINC benchmarks i386/x86_64 -- E@H for Linux , the ugly duck
)
Thanks for the intresting post. I hope the developers get the message re. a more efficient Linux E@H application. Hopefully now that they are beta testing a Mac OSX application they will address the Linux situation next. Cheers, Rog.
RE: Thanks for the
)
I agree. A better linux port would be very helpful!!!!! *hint to developers*
RE: Since BOINC-4.45 for
)
Microsoft provides the installer in a separate package for Win95, you don't need IE5. Get it here.
You might have problems running BOINCMGR, if so you can run boinc "standalone". In a dos box run:
boinc -redirectio
RE: RE: Since BOINC-4.45
)
Thanks for the hint, Walt.
Now, I'll have to go through the routine to starve BOINC/E@H before installing the new version. I'll report when I'm done.
-rg-
Even trying out some of the
)
Even trying out some of the more obvious optimisations will not improve the BOINC performance much ...
-----------------------------------------------------------------------------
export CC=gcc-3.4 CXX=g++-3.4
./_autosetup
CFLAGS='-O3 -mtune=athlon-4 -mfpmath=sse -funroll-loops -pthreads' \
CXXCFLAGS='-O3 -mtune=athlon-4 -mfpmath=sse -funroll-loops -pthreads' \
./configure --disable-server --enable-client
make CC=gcc-3.4 CXX=g++-3.4 clean all
client/boinc -run_cpu_benchmarks
-----------------------------------------------------------------------------
2005-06-27 23:11:09 [---] Starting BOINC client version 4.45 for x86_64-unknown-linux-gnu
2005-06-27 23:11:09 [---] Data directory: /usr/src/other/boinc_public
2005-06-27 23:11:09 [---] No general preferences found - using BOINC defaults
2005-06-27 23:11:09 [---] Remote control not allowed; using loopback address
2005-06-27 23:11:11 [---] Running CPU benchmarks
2005-06-27 23:12:08 [---] Benchmark results:
2005-06-27 23:12:08 [---] Number of CPUs: 1
2005-06-27 23:12:08 [---] 1709 double precision MIPS (Whetstone) per CPU
2005-06-27 23:12:08 [---] 4281 integer MIPS (Dhrystone) per CPU
2005-06-27 23:12:08 [---] Finished CPU benchmarks
2005-06-27 23:12:08 [---] Resuming computation and network activity
2005-06-27 23:12:08 [---] request_reschedule_cpus: Resuming activities
2005-06-27 23:12:08 [---] Insufficient work; requesting more
2005-06-27 23:12:25 [---] Received signal 2
2005-06-27 23:12:25 [---] Exit requested by user
2005-06-27 23:12:25 [---] request_reschedule_cpus: exit_tasks
-----------------------------------------------------------------------------
Einstein was running at the lowest priority at that time, so BOINC might get a bit more "stoned" when benchmarking an idle system ;)
But tweaking the C??FLAGS this way leads nowhere. Has anyone a better idea ?
-rg-
RE: the E@H application
)
Well - I think it does. The "problem" has been well discussed on the forums here, and the conclusion seems to be that the einstein app simply sacrifices accuracy for speed on Windoze.
So maybe there really is no problem: the two platforms complement each other.
For quick answers, send wus to Win-boxes, for *correct* answers, send them to Linux machines. Sort of like libs and conservatives :-) .
Greetings, Mr Ragnar Schroder
RE: Well - I think it
)
Thanks, this is enlightening, and I am sorry I have missed those posts.
Now, assuming this IS so, what should Linux users do ? Crunch under Wine, or natively ? Which benefits the project most ?
Regards.
-rg-
RE: Which benefits the
)
Returning valid results.