better performance with 32 bit ! why?
sven at whgl.uni-frankfurt.de
Wed Jun 29 20:00:39 UTC 2011
Thanks for that insight. I already considered something like the 'single
core per udp socket' problem.
One thing that just popped up my mind:
Does it increase performance, when you, let's say, bind multiple IPs to
the same NIC and make bind listen to all of those IPs, while of course
taking care to fix the corresponding NS RRs to contact all of thos IPs.
(I'm thinking of something like:
in NS ns1
in NS ns2
ns1 in a ip1
ns1 in a ip2
ns2 in a ipa
ns2 in a ipb
While ip1 and ip2 in the example are on the same NIC
If I understood you right, this should probably scale better on machines
with a high number of cores and improve the number of queries that can be
processed. More available UDP sockets should lead to more cores being used
for the initial IO stuff.
I am just asking out of curiosity though.
On Wed, June 29, 2011 18:49, Michael Graff wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> On 6/29/11 9:08 AM, Sven Eschenberg wrote:
>> Maybe some bind developer can shed a light on this:
>> Does bind use epoll()?
>> AIO (as in Posix RT extensions)
> BIND 9 uses epoll() I believe, but AFAIK does not touch AIO. I've not
> touched that code recently though, so this is possibly out of date
> information. Also, which version of BIND 9.x introduced the epoll() I
> am not certain of.
> That said, the single biggest problem in BIND 9 currently in terms of
> using all CPUs on a high-core machine is that all initial I/O is
> processed on a single core per UDP socket.
> Thus, you had 32 real cores in a server today and would use this beast
> for DNS, you'd not be happy unless you somehow spread this across
> multiple UDP listening sockets.
> We're working on this though. As with all things, we work on the things
> that are most important to our customers and users, so making your voice
> heard that this is important is helpful to us. Running code is always
> good too, as is a monetary donation for a development effort you want to
> see moving forward faster.
> - --Michael
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.4.11 (Darwin)
> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
> -----END PGP SIGNATURE-----
> Please visit https://lists.isc.org/mailman/listinfo/bind-users to
> unsubscribe from this list
> bind-users mailing list
> bind-users at lists.isc.org
More information about the bind-users