Linux vs Windows benchmarking
22 Feb 2005 7:23:52 UTC
Topic 187846
(moderation:
I've got 3 computers running this thing, 1 is a windows XP machine with an Athlon XP 2400+ (http://einsteinathome.org/host/33405) and the other is a Linux machine running a AthlonXP 3000+ (here: http://einsteinathome.org/host/20488).
Now, if you look at the benchmarking you'll notice the windows 2400+ got 1836.29 FLOPS and 4432.19 MIPS, while the linux 3000+ got 1157.19 FLOPS and 2646.83 MIPS. Why the massive difference? Any why is the 2400 scoring lower than the 3000? :p.
Linux vs Windows benchmarking
)
http://einsteinathome.org/node/187761
Applies for both Linux and Mac.
---
"Chance is irrelevant. We will succeed."
- Seven of Nine
>
)
> http://einsteinathome.org/node/187761
>
> Applies for both Linux and Mac.
>
> ---
>
I went looking for stuff like that before posting :p, aw well I'll try out the developers versions when I've finished my current WUs then. Thanks for replying :).
OK, this is really screwed. I
)
OK, this is really screwed. I just ran the BOINC client on my linux box via WINE and got the same high scores for the benchmark :p. After that actually running the e@h downloaded binary slowed wine right down but the lack of linux support or correctness or whatever you want to call it is really dissapointing :(.
If you're interested, here's the page for the wine'd client:
http://einsteinathome.org/host/34782
BTW, WU's under the linux native client are taking ~10hrs, with the wine'd client they were expected to take 5hrs 30m.
EDIT: I'll leave both the linux native client and the windows command line client wine'd overnight with the same nice level and report back with results in the morning...
I have recently tried E@H on
)
I have recently tried E@H on two IBM laptops, an old A21 with
Windows, a newer A31 with Linux. I don't have any faith in
benchmarks, so ran a simple double precision N-body integrator
on both; as expected the A31 was about 3x as fast.
g77 A21(0.6GHz) A31(1.6GHz)
(using cygwin)
--- 53.45s 5.38s
-O1 10.49s 3.79s
-O2 10.10s 3.48s
And yet E@H takes very similar times on the two machines to
complete a WU. Does this all come down to -- as I saw mentioned
in another thread -- really bad optimization under Linux?
> And yet E@H takes very
)
> And yet E@H takes very similar times on the two machines to
> complete a WU. Does this all come down to -- as I saw mentioned
> in another thread -- really bad optimization under Linux?
Looks that way. The same box which takes 5 1/2 hrs to complete under Windows takes a whopping 9 1/2 with Linux.
Although not a Linux end-user
)
Although not a Linux end-user yet, having a Linux-optimized E@H client would be a good incentive for me make the switch on at least one box; I wanna run E@H on every box I can. ;)
"Chance is irrelevant. We will succeed."
- Seven of Nine
I wonder if an Admin can move
)
I wonder if an Admin can move this to the crunching forum? Anyway...
I discovered this site: http://www.pperry.f2s.com/boinc-compile.htm and followed the instructions and ended up finally achieving a high whetstone benchmark in Linux ;). I went from 1269.41 million ops/sec to 2036Mo/s :D. This is faster than the 1994.68 I got with the Wine'ed Windows client. The dhrystone performance is still less (4756.62 from Wine'ed client, 3413 for the optimized Linux one), but I don't believe that E@H uses it so it don't really matter ;).
So devs, here are your flags:
export CFLAGS="-march=athlon-xp -O3 -fomit-frame-pointer -funroll-loops -fforce-addr -ffast-math -ftracer"
Can we have an optimized client pretty please? :D.
> I've got 3 computers
)
> I've got 3 computers running this thing, 1 is a windows XP machine with an
> Athlon XP 2400+
> (http://einsteinathome.org/host/33405) and the other
> is a Linux machine running a AthlonXP 3000+ (here:
> http://einsteinathome.org/host/20488).
>
> Now, if you look at the benchmarking you'll notice the windows 2400+ got
> 1836.29 FLOPS and 4432.19 MIPS, while the linux 3000+ got 1157.19 FLOPS and
> 2646.83 MIPS. Why the massive difference? Any why is the 2400 scoring lower
> than the 3000? :p.
did you compile and optimize the linux client? If not, that's why...
I run an xp box and a linux box. Anyway, before I compiled the linux software and just installed it without doing anything, my benchmarks were something like:
~900 whetstone
~1900 drystone
after I compiled and optimized it I get...
1600 whetstone
2600 drystone
so yeah, it makes a big diff. If your running linux you should ALWAYS compile your own stuff and that's one reason why.
Go here to compile linux boinc
http://www.pperry.f2s.com/boinc-compile.htm
PS. they don't need to optimize it, you do... if you've never done it, the instructions are pretty good on that site on how to do it... ;-)
> Although not a Linux
)
> Although not a Linux end-user yet, having a Linux-optimized E@H client would
> be a good incentive for me make the switch on at least one box; I wanna run
> E@H on every box I can. ;)
http://www.pperry.f2s.com/downloads.htm
And if you're lazy like me, the ready-made Pentium3-optimized client at the page above roughly doubled the CPU stats of my Linux boxes.
Linux is a great way to breathe life into an old box. I resurrected my Pentium-133 last year with Slackware. It performs great! Of course, I don't run X on it (no GUI) - it's command line only. But you'd never know it's a P-133 by the way it performs as a web server/mail server. Try running IIS on an old P-133 with 48MB RAM! Yeah, I know, apples to oranges, but still.
Have fun exploring the different Linux distributions, it can be a blast if you have the time.
Benchmarks don't tell the
)
Benchmarks don't tell the real story, completion times do.
Look at the benchmarks using the same box -
Linux
Measured floating point speed 2363.73 million ops/sec
Measured integer speed 4647.4 million ops/sec
Win
Measured floating point speed 2163.99 million ops/sec
Measured integer speed 5243.92 million ops/sec
Would you of guessed the Linux client completed in 9.5 hrs vs just 5.5 for Windows by relying on the benchmarks? FYI the optimized linux client at the pperry site was used.