Multiple CPUs not working

Steven Job list3 at wwwcrazy.com
Mon Jan 3 14:22:28 UTC 2005


A good (single CPU) comparison ca be found at:
http://www.shub-internet.org/brad/papers/dnscomparison/

I have done a test (using Bind 9) between two systems (multi and single CPU). 
But it was really for a comparison between Gentoo (using the Gentoo Kernel with
all of their patches) and RedHat Enterprise (using a custom compiled kernel). 
So a total of 4 systems were used (2 for each OS):

System #1
P4 800Mhz - 3.0 Ghz
2048 MB - DDR 400 (ECC and unbuffered)
Gentoo OS: 15.2k queries per second (on average of many runs)
RedHat: 14.8k queries per second (on average of many runs)


System #2
Dual Xeon 533Mhz - 2.8 Ghz
2048 MB - DDR 266 (ECC and registered / buffered)
Gentoo OS: 13.7k queries per second (on average of many runs)
RedHat: 12.9k queries per second (on average of many runs)

Each machine had 85k zones loaded.

Now the single cpu easily outperformed the multi-cpu but that makes sense since
the memory was faster, FSB was faster, and mem was unbuffered (actually not
sure if the unbuffered memory would make a difference).  I'm sure the single
CPU would even be faster if the memory was non-ecc.

I would really like to have a test of the same memory compenents and FSB.  Does
anyone have this?

So we decided that it would be better to do IP anycast with more single CPU
machines (even load balance them at each location) then to use the multi-cpu
option.  But then again the memory components in the Dual Xeon were a little
slower.
The strangest part though is that if we yanked out a CPU from the Dual Xeon
there was a "zero" noticeable difference. (And yes we did compile Bind
correctly and set the run time options correctly.)

Regards,
-Steve




Quoting James Bourne <jamesb at al.com.au>:
> Actually on that note anyone got any performance stats (eg. BIND v TDNS v
> MSDNS, uni/multi-proc etc.)?
>



More information about the bind-users mailing list