Bind 9.1.3 stop resolving but is still running.

[Brad Knowles]

At 5:23 PM -0400 9/5/01, Jason Larke wrote:

>  I ran the queryperf tool from BIND 9 against BIND 8.2.5RC1 and
>  9.2.0RC1. 8.2.5RC1 answered about four times as many queries per
>  second as 9.2.0RC1. It's possible that my test load- which was
>  gained from the querylog of a production server- had some oddity
>  in it that made 9 underperform badly, or that the default options
>  in the BIND 8 makefile did a better job than configure did.

	Well, I wouldn't consider 9.2.0 to be a version suitable for 
doing final benchmarking, not just yet.  Try it against 9.1.3-REL, 
and make sure you try both threaded and non-threaded versions.  Also 
make sure you've got enough memory, so that you don't run into 
thrashing issues.


	Out of curiosity, is your server authoritative for the zones that 
you're querying, or are you testing the caching performance of BIND 9 
versus BIND 8?

	If you're testing the caching performance, I'd encourage you to 
take a look at the archives of this mailing list and look for the 
name "Matt Simerson", as he has also done some interesting benchmarks 
of the relative caching performance of BIND 8 and BIND 9 (using some 
older tools), and you may gain some additional insight by reading his 
posts on the subject.

>  This was on a two-processor machine, too.

	Just because it's a two-processor machine doesn't necessarily 
mean anything -- if the version of BIND is running non-threaded, then 
this won't help at all, and may even hurt.

	Of course, it may very well make a big difference what OS you're 
doing the benchmarking with, too.

>  I'm not a C programmer and I didn't have time to do a lot of work
>  analyzing the issue, but it looked to me something odd had to be
>  going on with BIND 9.

	The thing that Rick Jones teaches us is that we really need to 
get pretty deep into the BIND code to fully understand why it's 
bottlenecking on some operations, and then work to try to optimize 
those specific operations.

	If you're running on a platform where both gcc and "professional" 
compilers are available directly from the vendor, I would encourage 
you to try the benchmarking with both sets of compilers, and with 
various optimizations turned on.  There's going to be a limit to how 
far you can go with gcc, but Rick Jones shows us that there may be 
quite a bit more you can do with vendor compilers.


	See <ftp://ftp.cup.hp.com/dist/networking/briefs/> if you want to 
read the various BIND benchmarking papers that Rick has written.

-- 
Brad Knowles, <brad.knowles at skynet.be>

H4sICIFgXzsCA2RtYS1zaWcAPVHLbsMwDDvXX0H0kkvbfxiwVw8FCmzAzqqj1F4dy7CdBfn7
Kc6wmyGRFEnvvxiWQoCvqI7RSWTcfGXQNqCUAnfIU+AT8OZ/GCNjRVlH0bKpguJkxiITZqes
MxwpSucyDJzXxQEUe/ihgXqJXUXwD9ajB6NHonLmNrUSK9nacHQnH097szO74xFXqtlbT3il
wMsBz5cnfCR5cEmci0Rj9u/jqBbPeES1I4PeFBXPUIT1XDSOuutFXylzrQvGyboWstCoQZyP
dxX4dLx0eauFe1x9puhoi0Ao1omEJo+BZ6XLVNaVpWiKekxN0VK2VMpmAy+Bk7ZV4SO+p1L/
uErNRS/qH2iFU+iNOtbcmVt9N16lfF7tLv9FXNj8AiyNcOi1AQAA