raffaele.ronchi@xxxxxxxxx wrote:
> I have installed Fedora Core2 (Kernel 2.6.5-1.358smp) on a system
> with Xeon MP processors. I ran the same program (multithreaded) on
> this system and compared the results Vs. what obtained on the same
> hardware running RH Advanced Server 3.0 (Kernel 2.4.21-9.ELsmp).
Four points:
1. As other people have said, it's very difficult to provide meaningful
answers when all we know is that the program is multi-threaded. Does it
use NPTL threads, for example?
2. It looks as though your program spends a lot of time in kernel mode.
Kernel mode switches are relatively slow on Fedora 2 SMP, because it has
the 4G/4G kernel mode enabled by default. This forces the MMU to throw
away its TLB on each kernel / user context switch, and your program looks
like it has a lot of them.
To get around this, you would have to recompile your kernel. Under
"Processor type and features" and "High Memory Support", choose the
*lowest* amount of memory you need.
3. It is known that in some areas 2.6 will perform worse than 2.4.
Usually this is aimed at providing a more consistent environment where
performance won't become spectacularly bad as you approach the worst-
case scenarios. If you can identify particular areas of your program
that have slowed down, and can provide a test case to the Linux Kernel
Mailing List, mentioning the word "regression", I'm sure they'd be very
interested.
4. When posting to a mailing list, it is normally considered polite to
keep your lines under 80 characters (under 72 is often considered good
practice).
Hope this helps,
James.
--
E-mail address: james@ | "Never trust a species that grins all the time.
westexe.demon.co.uk | It's up to something."
| (Terry Pratchett on dolphins)
