OGR--Linux speed vs. NT speed

[MaxCrunch]

I have two PCs with identical hardware, one running RH6.0 (with a 2.2.16 kernel), the other running WinNT Workstation 4 (SP6a). All the machines do 24x7 is crunch OGR data. It seems that the Linux box is only crunching about 4.67 Mnodes/s while the NT box is crunching about 6.3 Mnodes/sec. Is the Linux client really that much slower, or should I be looking for other problems with the Linux box? There is nothing else running on that box, and the Dnet client is getting 99.9% of the processor time.

TIA!
MaxC

 

[Joe O]

MaxCrunch, What proxies are you running on the two machines?
We have found that 463B is 20 to 30 % faster than 462 or 460 when running under the same Windows.

 

[bd]

They operating system shouldn't matter. You prolly having something running on the linux box in the background that is taking some time? What kind of install did you do for the RH6 box ?

 

[MaxCrunch]

Joe O -- I am using the latest client on both machines.

bd -- top shows nothing else big running, and the Dnet client is getting all of the CPU cycles on the Linux box. I did a minimal Linux install, with basically only the networking packages and the programs needed for compiling new kernels on the machine. No X, no Netscape, no server software running, nothing. I don't even have the sendmail daemon running!

MaxC

 

[Joe O]

MaxCrunch, I wonder if this is the same issue as with GammaFlux. The DCypher.Net *nix clients are slower than the Windows clients because the *nix C compilers do not optimise as well as Microsoft's C compiler.

 

[TwoFace]

Hey MaxCrunch...

I've got the same problem also with both running the latest client... They aren't identical tho' but after calculating using only the output from the clients and the MHZ I found the linux client to be about 22% slower/MHZ This is probably not a very accurate way to do things but it's what I found anyway...

Hope Dnet releases a new client soon that addresses (spelling?!?) this...

 

[Lord Demios]

Do this as root on your box. (warning only do this if it IS YOUR BOX)

nice -10 ./dnetc -? -? -?


the nice -10 will put this process as having higher priority then most non-critical system processes. This is what I run mpg123 at.

LD

 

[MaxCrunch]

Lord D--I'm not sure how changing the nice level would affect things, as my Dnet client appears to already be getting all of my CPU's cycles! I have a feeling that, as Joe O said, the Linux client simply isn't optimized to the extent of the NT client. I'd hate to have to convert the box to NT simply to crank OGR stubs faster...

By the way, I appreciate all of the "Linuxy" things you post Lord D--good stuff!

MaxC

 

[Verdoux]

I also observed that the NT OGR client was 25-30% faster than the linux one.

In fact running the NT client under linux with wine, i got the same speed than under NT.
That shows that this is a compiler issue.
It's a real problem because with wine you can't cleanly shutdown the NT client: you lose the current stub

 

[Qythyx]

I did a little benchmarking when I upgraded my NT boxes to the latest client and for NT the 463b client is faster. For Linux, however, the 462 client is faster and the the 463 client slows down signifcantly, so, downgrade those linux boxes.

P.S. I dunno why, maybe because they compile static libraries now and maybe they're not optimized. ???

 

[Verdoux]

You're right. The 462 is faster than the 463.

The changelog says that the new client is compiled against the old libc5 lib. That also may be the cause of the slowness.

Hope they will change that for the next releases.

 

[TwoFace]

Where can I get the 462 for linux then? At DNet there is only the newest available

 

[Joe O]

Try this FTP page ftp://ftp.distributed.net/pub/dcti/v2.8010/ or this one ftp://ftp.distributed.net/pub/dcti/v2.8009/

 

[TwoFace]

Thanks Joe O but the first one has only got the 463 and the last one only 460 Nogo I'm afraid...

 

[Joe O]

For OGR, 460 and 462 benchmarked the same nodes/sec under Windows.
It's worth a try.

 

[TwoFace]

Oki... I'll try the 460 then

 

[Verdoux]

Twoface> Did you carefully read the listing ?
The 462 linux client is still there.

 

[TwoFace]

OMG! You're right Verdoux... Must've hit a blindspot that! Cool then I'll just get the 462 then...

 

[TwoFace]

I have the 462 installed now... will see how it does and report back later ... hope it's faster so we can trash the / !

 

[MaxCrunch]

Thanks for the input, everybody! I'll give 462 a try when I get home this evening.

MaxC

 

[TwoFace]

Well I've been running the 462 client on my linux box for approx 42 hours now and here's what I've found:

It's about 96,88% as effective as the win32 client... this is a huge improvement over 463 which was about 22% slower ... I'm happy now but I really hope DNet will send out a new client for Linux since I don't really like running anything but the newest... Anyway happy cracking and let's trash the /