[dns-operations] DNS server benchmarking sanity check
dns at fl1ger.de
Mon Aug 15 05:19:34 UTC 2016
On 15 Aug 2016, at 5:29, Mark Delany wrote:
> My total guess is that the kernels are doing some sort of hash onto a
> lock table based on the destIP, destPort tuple so a unique destIP
> would be as useful a tie-breaker as destPort appears to be.
I think that guess is correct as there are some ethtools settings to it
for a lot of network cards. I use the following setting for most of my
testing and this gets better results when using only one IP.
/sbin/ethtool -N eth0 rx-flow-hash udp4 sdfn
> Having said all that, many of these issues seem obvious. Is it really
> the case that UDP kernel performance has not been extensively analyzed
> in the context of DNS servers?
I'm not sure. Most of the people who have done some work on this thread
seem to be in agreement. More ports/IP and SO_REUSE_PORT give you better
performance if the kernel and application support it. Scaling above 8
cores doesn't give you the same increase of performance on pure UPD
answering as up to 8 cores. That matches my experiences also. And I am
sure Anand will talk about his results in an upcoming RIPE conference.
More information about the dns-operations