Messages in this thread | | | Date | Wed, 28 Jan 2009 16:32:04 +1100 | From | Bron Gondwana <> | Subject | Re: [PATCH 1/3] epoll: increase default max_user_instances to 1024 |
| |
On Tue, Jan 27, 2009 at 08:00:30PM -0800, Davide Libenzi wrote: > On Wed, 28 Jan 2009, Bron Gondwana wrote: > > > Both Postfix and Apache use an epoll instance per child, which > > leads to significant scalability issues with max_user_instances > > set so low. Bump the default to 1024 so medium sized sites are > > not impacted. > > NACK. Epoll allocates globally about 100 to 160 bytes (32/64 bit) for each > file added to the interface: > > for i 1..1024 > for j 1..1024 > if i!=j > add j -> i > > That's (N^2 * {100, 160}) = 100MB to 160MB of pinned kernel memory,
Woah - that's serious.
This:
instances_uid 0 (root) num_instances 142 max_instances 4096 watches_uid 107 (postfix) num_watches 1097 max_watches 266555
isn't serious. It's pretty sane. 142 processes with an epoll watcher, and fewer than 10 fds per epoll. Unfortunately, it wouldn't work on an unpatched and un-specially-configured stock kernel. That's steady-state too, not a peak. I just grabbed it off a running MX:
[brong@mx1 ~]$ free total used free shared buffers cached Mem: 4151652 3113128 1038524 0 130808 2014152 -/+ buffers/cache: 968168 3183484 Swap: 2047992 50364 1997628 [brong@mx1 ~]$ uptime 00:31:05 up 2 days, 18:03, 2 users, load average: 0.86, 1.23, 1.08
Hardly looking stressed right now.
If I'm reading it right, your concern is the massively recursive case, where every single epoll gets added to every other epoll as a chained file descriptor?
That's clearly not happening here - so it seems that maybe our "happy medium" is actually in closer inspection of what's going on rather than a blanket low N to keep N^2 down.
Bron.
| |