bind9 performance cannot ramp up cpus

[Rick Jones]

"yan, tommy (CSG-MKG-BJ)" <tommy.yan at hp.com> wrote:
> I'm using a 6 cpu hp Itanium hpux server (as bind server)for testing
> bind performance, but can only ramp up to 4 cpus,and thus does not
> generate fairly good performance.

which system are you using?

> I use bind9.2.0, hpux11.23, and queryperf, which is the perf test

I think that isc.org is up to a 9.3.something.

> tool from isc. queryperf input file arround 60000 lines, and named
> data only has 10 records(so this is a simple test).

> named.conf, nsswitch.conf, resolv.conf are also very simple, I'm
> sure they point to right dns and resolve using dns first.

> It is interesting that each queryperf client(with 50 or more
> concurrent thread) can only take up 1MB network throughput, I add
> more clinet, and get better performance.

Well, if there are a fixed number of threads and those threads will
only have so many queries outstanding at a time...  (I don't know if
queryperf is synchronous or async wrt its generation of queries)

> Clients are servers with
> dedicated LAN connecting to bind server. The test result as below:

> num. of client   query per second
> 1                   7000
> 3                   15000
> 5                   25000

> monitering with glance, named wait event is stream and lan, system
> call mainly on messege send/recieve, while system tables are not
> full at all.  I've checked that dns service use udp socket for
> communication, and guest maybe the bottomneck resides on network. So
> I tunned some socket cache/connection parameters, but does not
> improve much.

What do you mean by socket cache?  Anyhow, what sort of NIC are you
using in the 6-CPU server?  Please be specific about model - take the
data from ioscan -fk | grep lan - don't just say 'Gigabit' :)

If you go to the 'a' screen of glance and look at per-CPU utilization
what do you see? Consider both user, kernel and interrupt time.

If you go to the process system calls page of glance "L" (IIRC) what
other system calls do you see besides the send/recv? (go ahead and
cut/paste the screen).

What happens if you add IP addresses to the server and spread the
queries across those IPs?  Just adding IPs, not adding NICs.

rick jones
ftp://ftp.cup.hp.com/dist/networking/briefs/
-- 
oxymoron n, commuter in a gas-guzzling luxury SUV with an American flag
these opinions are mine, all mine; HP might not want them anyway... :)
feel free to post, OR email to raj in cup.hp.com  but NOT BOTH...