Also after the first Mindcraft's test, it is written in their
report that among NT tuning procedure they have installed
some experimental patch which assigns each network card
directly to each CPU, so that CPU's do not compete for
network cards and in fact do not go through the single
(as in Linux, I suppose) TCP stack.
Anatolii
----Original Message Follows----
From: [EMAIL PROTECTED]
To: [EMAIL PROTECTED]
Date: Tue, 6 Jul 1999 12:24:19 -0500
Another interpretation of the Mindcraft and c't measurement results is that
for
static, in-cache page-serving, that the
limiting factor is the bandwidth of the network interface. If the network
interface is the bottleneck, then adding processors
is not a very interesting test of scalability. Also, it might be the
case
that in order for the 4-way NT system to be bottlenecked on
something other than the network card (once again for static, in-cache
page-serving), you have to have 4 network interfaces.
If this is true, then the reason that Linux doesn't do as well in the MP
cases
merely reflects that Linux MP doesn't scale well.
I've heard (but haven't personally checked) that Linux networking code in
2.2.x
is single-threaded. If this is correct, then
that just exacerbates the scaling problem.
Best Regards,
Ray Bryant
[EMAIL PROTECTED]
512-838-8538
Best Regards,
Ray Bryant
[EMAIL PROTECTED]
512-838-8538
-
Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/mentre/smp-faq/
To Unsubscribe: send "unsubscribe linux-smp" to [EMAIL PROTECTED]
_______________________________________________________________
Get Free Email and Do More On The Web. Visit http://www.msn.com
-
Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/mentre/smp-faq/
To Unsubscribe: send "unsubscribe linux-smp" to [EMAIL PROTECTED]
