Linux vs Windows performance.
6 Feb 2006 11:37:28 UTC
Topic 190720
(moderation:
I have home computer which run windows as well as linux. For linux I have: Measured floating point speed 787.48 million ops/sec, the same machine under Windows: 1363.58 million ops/sec.
Integer speed is the same.
I see this picture with other linux and windows based compters (all 3 or 3.2Ghz P4): for linux float speed is half of windows's speed.
Is this measurement problem or compilation problem?
Linux vs Windows performance.
)
The problem is your boinc client. If your boinc client is optimized(you can find optimized clients built for Linux), it will show you a more accurate floating point speed.
The einstein program itself though is already optimized, so you should notice about the same performance between windows and Linux.
If I remember right, the calculated speeds are only used for calcuating granted credit, and in fact I think having the lower floating point speed in Linux, but still getting the work done at the same rate as windows means you will be given more points. However, almost everyone else uses Windows, so usually your score is thrown out in favor of the Windowss crunchers' numbers, since their numbers match each others, but not yours.
such things just should not be writ so please destroy this if you wish to live 'tis better in ignorance to dwell than to go screaming into the abyss worse than hell
_Very_ strange for me,
)
_Very_ strange for me, because I have these measurement results on several installations:
1. Workstation always under Linux 24x7: 700M
2. Workstation at home with primarly under Windows (when gaming), and Linux: 1400M vs 700M
3. New Windows worksation with zero result returned so far: yep, again, 1400M.
Like I said before, that is
)
Like I said before, that is because the boinc client itself does the benchmark. The standard boinc client for linux is not optimized, so it gives a bad benchmark. Switch to an optimized client for Linux and it will give a correct measurement.
The measurement itself really doesn't matter though, do all 3 computers complete the work units in about the same amount of time?
such things just should not be writ so please destroy this if you wish to live 'tis better in ignorance to dwell than to go screaming into the abyss worse than hell
In comparison to Windows,
)
In comparison to Windows, Linux does not benchmark as well on otherwise identical machines, using the recommended Boinc clients. Since claimed credit is a funcrion of the sum of the benchmarks multiplied by processing time, the lower benching Linux machines will yield lower credit claim than their Windows counterparts for equal work.
What I've read from a few crunchers who've recently tried both OSes on the same machine is that the latest Linux app processes work slightly faster than the Windows app (depending on which flavor of Linux) by something like 5%.
Some of the largest and most productive "farms" (for example, those of Steffen Grunewald and Lunarcom) are exclusively Linux rigs, and if you look through the results of some of their computers, there are very few problems with invalidation arising from crunching alongside Windows (or Mac) machines.
Michael
microcraft
"The arc of history is long, but it bends toward justice" - MLK
RE: The measurement itself
)
I'd never compare the rate, due only one host is 24x7 running.
My experience indicates that
)
My experience indicates that Linux seems to process a "little" bit faster. I have always attributed this to Linux not needing to display graphics. Can anybody suggest a test to determine if this is so? Also, What Linux Distributions are best, and how can one ever prove it?
RE: My experience indicates
)
With Albert it became harder to make such comparisons. Some time ago
I found a WU, gone by now, which was crunched by two identical hosts,
one Linux, one Windows. The numbers can be found here.
Someone else stated that the assumption, Linux faster than Windows, only holds
on Intel CPU's. Couldn't find an example though. Maybe someone else
finds more examples.
Michael
Team Linux Users Everywhere
Well, rather than drone on
)
Well, rather than drone on and on about this OS vs. that OS...
I Just set up 3 IDENTICAL computers up in the lab. 1 (Moe) is XP, 2 (Larry) is Ubuntu Linux (hoary hedgehog), and 3 was running 2000 Pro when I got to it. I will give them all a few days to fight it out (and bump my stats) and then compare thier numbers.
OS freaks and fans are welcome to give me thier input as to how to tweak the various units as I am a Linux and BOINC newbie. When all is said and done, I expect that Linux will be a winner. However, Like a good game you never really know who will win and I figured this to be the best way to cut through the hype.
"inquiring minds want to know"
RE: Well, rather than drone
)
Alan,
Thanks in advance for the sacrifice. I say "sacrifice" only because I think you'll have to go to Seti or something to find WUs close enough to equal to use for "calibration". You might be able to get away with downloading a datafile of large albert WUs to only one machine and cloning it into the others, but the schedulers on the others may not recognize it.
Interesting, though. I'd like to learn the results.
Michael
microcraft
"The arc of history is long, but it bends toward justice" - MLK
The other alternative is to
)
The other alternative is to do a lot of work, then use the averages produced. In my case I could differentiate between AMD and Intel and even P4, Xeon ... but, not windows/linux (OS-X is sufficiently different with the hand coding that there is no point).
If I get a couple votes for interest I will write the SQL to do the query against my BOINC View logs (yes I log all work and usually put it into a database for questions like these that can be answered this way).