Netbench results on a linux 4-way system

[Andrew Theurer]

I will put up the throughput vs client data this week on a web page.  I'll 
send out a URL in a day or two..  

-Andrew

On Thursday 01 August 2002 2:58 am, Goetz Rieger wrote:
> Hey Andrew,
>
> On Wed, 31 Jul 2002 12:56:04 -0500
> Andrew Theurer <habanero at us.ibm.com> wrote:
>
> great stuff. As you can think, I have never got the chance to run "real"
> netbench-runs. So I am not very familiar with interpreting the results....
>
> Could you publish more data, like the plot throughput against clients?
>
> Any performance data on Samba/Linux is relevant. ;-)
>
> Regards,
> Goetz
>
> > I though I'd share some NetBench results on one of our servers.
> >
> > Server:
> > 4 x 1.5 GHz P4, 256K L3, 32MB L4, 2 GB memory
> > 4 x 1Gbps acenic ethernet
> > 14 SCSI disks in hardware RAID1 with 128 MB writeback NVRAM
> > SuSE 8.0, 2.4.18 kernel
> > Samba 2.2.3a
> > Ext3 fs
> >
> > Clients:
> > 48 x 866 MHz PIII running Windows 2000
> >
> > Results:
> >
> > Baseline				576 Mbps
> > ext3 data=writeback		623 Mbps
> > samba smblog=1			673 Mbps
> > sendfile/zerocopy		801 Mbps
> > O(1) scheduler			809 Mbps
> > Evenly affined IRQs		800 Mbps *needed to get process affinity correct
> > Process affinity			848 Mbps
> > /proc/sys/net/hll=764		853 Mbps
> > case sens enforced		895 Mbps
> > samba spinlocks			912 Mbps
> > dcache read copy update 	923 Mbps (also had 5% idle time)
> >
> > I have also achieved 1002 Mbps with ext2.
> >
> > Some other things I think may be worth investigating:
> >
> > gettimeofday().  Samba calls this a lot, one for every reply I think, to
> > check for connection timeout.  This means we go into kernel mode every
> > single time we call this, something I'd like to avoid.  And I also don't
> > think we need the resolution of gettimeofday for this.  How about some
> > sort of timer in samba with a 1 second granularity?  I admit I have not
> > thought about how to do this, but there's gotta be a way.
> >
> > locking for the samba db.  Spin locks got us a little better than
> > flocks, but again I'd rather not go into kernel mode every time.  Has
> > anyone considered using Rusty Russell's futexes for this?
> >
> > 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.
> >
> > Anyway, I have a lot of analysis data (kernel profiles mostly) that goes
> > along with this stuff if anyone is interested in looking at it  My
> > intention was to make some sort of article out of this, so I wanted to
> > get some feedback from this list.  Is this stuff interesting or relevant
> > to anyone out there?  What else would you like to see in terms of samba
> > on linux  performance?
> >
> > Thanks,
> >
> > Andrew Theurer