bind 9.3.0 performance problem on x86 Linux

[yan, tommy (CSG-MKG-BJ)]

Hi list,

In my testing enviroment x86 Redhat Linux is having
trouble in 4 CPUs config, I'd summary below, hope you
could help.

I use queryperf, which is included in bind, to test
bind 9.3.0 performance on Linux(rhel3 u3), both HP
Proliant(dl580 G2) and Itanium(rx4640). This is just a
simple test, only 5 records in a master domain, so all
data are in cache. named is compiled and configured to
enable multi-threads. I use 4 queryperf clients, add
presure to a simgle server (4 IPs over a single NIC)
separately.

For Integrity Linux(2.4), 2 cpu got ~50000 qps, and 4
cpu got ~75000. usr% is about 70%, sys% is below 30%.
I suppose this is ok.
For Proliant Linux(2.4), 2 cpu got ~50000 qps, but 4
cpu got almost also 50000. sys% is 40%.

Network is not a problem, since rx4640(4cpu) took
15MB/s, which dl580(4cpu) only took 11MB/s.

I suspect either the redhat kernel has a scalability
problem, or SMP architecture has trouble handling mass
dns requests.

Do you have comments?

Thx.

=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
/etc/init.d/named  only added a -n 4 option


/etc/named.conf   only added a minimal-responses option

Tommy Yan
=20
Linux & Opensource Lab China
IDD=A3=BA86-10-656345365
Mobile=A3=BA86-13910290301
Mail=A3=BAtommy.yan at hp.com