[bind10-dev] recent performance improvements

[JINMEI Tatuya / 神明達哉]

At Tue, 20 Mar 2012 08:13:14 -0500 (CDT),
"Jeremy C. Reed" <jreed at isc.org> wrote:

> Okay. I am now specifically setting it to -n 1 by default. The updated 
> performance results (bind9 without threads is slower then bind9 built 
> with threads and -n 1):
> 
> builtin--.nxdomain 3202.635569 qps 3% slower
> builtin--.soa 3408.178849 qps 5% slower
> builtin--.success 15266.703820 qps 6% slower
> root-memory-.nxdomain 38362.727878 qps 38% slower
> root-memory-.soa 25547.769731 qps 56% slower
> root-memory-.success 16004.849085 qps 68% slower
> smallzone-memory-.nxdomain 40865.231371 qps 34% slower
> smallzone-memory-.soa 35941.285741 qps 42% slower
> smallzone-memory-.success 37586.834986 qps 39% slower
> 
> (By the way, this is different than I had found a couple years ago, 
> where the possible overhead of one thread caused worse performance than 
> no threads).

'named -n 1' w/ threads could be faster than named w/o threads since
9.4, especially on a multicore machine.  In the authoritative-only
(and no frequent dynamic changes) scenario, the thread overhead is
minimal because it works almost lock-free, and the advantage of the
(single) worker thread not having to watch network/timer events can
outweigh other overhead of the use of the thread.

At least I observed that when I did performance measurement for
multi-thread enhancements before they were merged to 9.4.

> > In the context of comparing BIND 9 and BIND 10, it just doesn't make
> > sense to compare "enabling threads without -n" BIND 9 with a
> > single-process b10-auth if the machine has multiple cores.

> Single b10-auth versus named built with threads using -n 1:

> I need to redo my research of BIND 9 comparing no -n and -n 1 through 
> -n32. Last time I did it on same system with older BIND 9, named 
> detected cpus was 84% the speed of the best -n number. Since you 
> reminded me about -n 1 again that may explain why -n 4 was slower in 
> some cases versus no -n. So maybe a newer BIND 9 improved this 
> autodetection or it better used a different amount of worker threads.  
> Then I will use the best -n to compare against best b10-auth amount.

Without -n, named just tries to use all available CPU (core)s.  It may
or may not be the best choice depending on the number itself or the
system's thread model.

---
JINMEI, Tatuya