Re: Postfix benchmark: bug or performance regression ?

2019-03-29 Thread Wietse Venema
Ralf Hildebrandt: > * Viktor Dukhovni : > > > On Mar 28, 2019, at 12:03 PM, Wietse Venema wrote: > > > > > > And thank you for your thorough investigation that helped to narrow > > > down the root cause: under high traffic conditions, LMTP connections > > > are cached but never reused, therefore

Re: Postfix benchmark: bug or performance regression ?

2019-03-29 Thread Ralf Hildebrandt
* Viktor Dukhovni : > > On Mar 28, 2019, at 12:03 PM, Wietse Venema wrote: > > > > And thank you for your thorough investigation that helped to narrow > > down the root cause: under high traffic conditions, LMTP connections > > are cached but never reused, therefore those idle cached connections

Re: Postfix benchmark: bug or performance regression ?

2019-03-28 Thread Viktor Dukhovni
> On Mar 28, 2019, at 12:03 PM, Wietse Venema wrote: > > And thank you for your thorough investigation that helped to narrow > down the root cause: under high traffic conditions, LMTP connections > are cached but never reused, therefore those idle cached connections > are exhausting server

Re: Postfix benchmark: bug or performance regression ?

2019-03-28 Thread Wietse Venema
Juliana Rodrigueiro: > > To get rid of the 2s delays: > > > > /etc/postfix/main.cf: > > lmtp_connection_cache_on_demand = no > > > > Please let us know if that helps. Meanwhile we can develop a proper fix. > > And yes, it worked, that helped a lot. Although not as fast as before, but now > I

Re: Postfix benchmark: bug or performance regression ?

2019-03-28 Thread Juliana Rodrigueiro
On Wednesday, 27 March 2019 20:01:49 CET Viktor Dukhovni wrote: > On Wed, Mar 27, 2019 at 03:36:28PM +0100, Juliana Rodrigueiro wrote: > > However, during a benchmark, we realized 3.3.2 was 5 times slower than the > > version before. > > This is misleading. Postfix is not 5 times slower, your

Re: Postfix benchmark: bug or performance regression ?

2019-03-28 Thread Wietse Venema
Wietse Venema: > Juliana Rodrigueiro: > > Excerpt of maillog version > 2.11.1: > > Mar 27 14:46:50 localdomain postfix/lmtp[24750]: 6CEFF61: > > to=, orig_to=, > > relay=localdomain.com[/var/ > > imap/socket/lmtp], delay=0.02, delays=0.01/0/0.01/0, dsn=2.1.5, status=sent > > (250 2.1.5 Ok

Re: Postfix benchmark: bug or performance regression ?

2019-03-27 Thread Viktor Dukhovni
> On Mar 27, 2019, at 3:01 PM, Viktor Dukhovni > wrote: > > There's likely a bug. We should either simulate a synthetic nexthop > ($myhostname?) for unix-domain destinations, and then do nexthop > reuse (and perhaps do no caching by endpoint address for unix-domain > destinations), or go back

Re: Postfix benchmark: bug or performance regression ?

2019-03-27 Thread Wietse Venema
Juliana Rodrigueiro: > Excerpt of maillog version > 2.11.1: > Mar 27 14:46:50 localdomain postfix/lmtp[24750]: 6CEFF61: > to=, orig_to=, > relay=localdomain.com[/var/ > imap/socket/lmtp], delay=0.02, delays=0.01/0/0.01/0, dsn=2.1.5, status=sent > (250 2.1.5 Ok SESSIONID=) > Mar 27 14:46:50

Re: Postfix benchmark: bug or performance regression ?

2019-03-27 Thread Viktor Dukhovni
On Wed, Mar 27, 2019 at 03:36:28PM +0100, Juliana Rodrigueiro wrote: > However, during a benchmark, we realized 3.3.2 was 5 times slower than the > version before. This is misleading. Postfix is not 5 times slower, your benchmark appears to be measuring the LMTP delivery rate to a single sink