Netbench results on a linux 4-way system
habanero at us.ibm.com
Thu Aug 1 11:31:02 GMT 2002
David Collier-Brown wrote:
> Andrew Theurer wrote:
> > Hyperthreading. With 2 physical processors, I can get 25% better results!!!
> > with 4 physical processors, I only get 2% better. I may be running into
> > other bottlenecks on the 4 physical/8 logical CPU case, so I hope there is
> > room for improvement. However there are probably a lot more 2-way P4 systems
> > out there than 4-way, so I bet this could really benefit a lot of people out
> > there.
> Hmmn: can you tell if you're hitting a network limitation
> or if your disks are saturated? It would be cool to know if
> it's an external or a kernel bottleneck.
> I'll opine that it's not Samba itself, as a colleague has
> run many many sambae on a largish Sun multiprocessor (;-))
Probably not samba, unless the spinlocks for the db are getting to be
too much (the reason for futexes). I _really_ doubt it is disk IO, as
most of the data is cached, and the majority of the disk IO is journal
activity. It could be network, sort of, but doubtful. The cards (4
gigabits) are capable to handle much more, but we have seen that network
throughput is very sensitive to FSB speed. I have not taken a profile
yet, and hopefully when I do, we will understand the problem more.
One other thing, P4 by default sends all ints to CPU0. This _was_ a
problem for 4 physical/8 logical processors. A single physical
processor is enough to handle all the ints from these network cards at
the throughput we are seeing. However, when you split the physical CPUS
into 2 logical CPUS, (hyperthreading) "1/2" of a CPU is not enough to
support processing all the ints. I used a new irq balancing feature to
spread the ints to multiple processors to avoid this.
More information about the samba-technical