raindrops RubyGem user+dev discussion/patches/pulls/bugs/help
 help / color / mirror / code / Atom feed
From: Ryan King <ryan@twitter.com>
To: raindrops@librelist.com
Subject: Re: queued is always 0
Date: Thu, 1 Jul 2010 11:37:15 -0700	[thread overview]
Message-ID: <AANLkTinP5fBfcRa6Y3-MrMzDqjRRrgEOIunvn_h7q2WS@mail.gmail.com> (raw)
In-Reply-To: 20100701101809.GA23938@dcvr.yhbt.net

On Thu, Jul 1, 2010 at 3:18 AM, Eric Wong <normalperson@yhbt.net> wrote:
> Ryan King <ryan@twitter.com> wrote:
> Sorry for the late response, I was afk the entire day.
>> On Tue, Jun 29, 2010 at 7:45 PM, Eric Wong <normalperson@yhbt.net> wrote:
>> > Ryan King <ryan@twitter.com> wrote:
>> >> On Mon, Jun 28, 2010 at 5:50 PM, Eric Wong <normalperson@yhbt.net> wrote:
> <snip>
>> > What happens when you run fewer Unicorn workers or increase concurrency
>> > with ab?
>> I bumped the concurrency up to 2000 (with 16 workers) and still no
>> queuing (according to raindrops).
>> Interestingly, when I ran this benchmark ab died because of timeouts.
>> I lowered the concurrency back to 16 and ran it again, which quickly
>> caused timeouts. When I looked at the unicorn processes it was obvious
>> that they were still processing a backlog processes from the first ab
>> run. I'm not sure if this is how it is supposed to work, but it was
>> surprising to me.
> They were buffered in the kernel, and Unicorn doesn't have a good way to
> know if a client disconnected until Unicorn tries writing the response
> (often not even the first packets).
>> > How large are your responses?
>> About 43 KB.
> That should be small enough to fit inside the default socket buffers
> without ever blocking.
>> > If you have small responses, Unicorn could've also finished writing the
>> > response to the socket buffers and accepted another connection by the
>> > time ab gets around to reading the socket.
>> >
>> > Under Linux, Unicorn (and Mongrel) also uses TCP_DEFER_ACCEPT with a
>> > value of 1 (second)[1], so it won't even register in the queue if
>> > ab opened the socket and didn't have a chance to write to it, yet...
>> >
>> > You could try commenting out the setsockopt call to set
>> > TCP_DEFER_ACCEPT in lib/unicorn/socket_helper.rb for testing
>> > and see if the queue jumps up, too..
>> We're not setting the tcp_nodelay option, so that code shouldn't be
>> getting run, right?
> It does run and we always set TCP_DEFER_ACCEPT (since the Mongrel days).
> It is safe to disable for Unicorn if you're behind a reverse proxy,
> but it shouldn't be a noticeable performance hit either way.
> I'll definitely make this configurable in the future (since higher
> values are desirable against some attacks), but there's no reason to
> disable it for anything other than benchmarking/testing you're doing
> now..
>> > I noticed that the usage string for linux-tcp-listener-stats.rb was
>> > wrong yesterday (and updated it in rainbows.git), but you can try
>> > running it in a tight delay loop with "-d <seconds>":
>> >
>> >  ruby linux-tcp-listener-stats.rb -d 0.01 | awk '$3 != 0 { print $0}'
>> >
>> > The above will only print lines if there are queued connections.  You
>> > can replace "-d 0.01" with a smaller number, but Raindrops (and the
>> > underlying tcpdiag kernel module) can only give a snapshot of the
>> > current queue size).
>> I had done basically the same thing in previous test runs. Never any
>> queued requests.
>> I'm really hoping that I just misunderstand something here, but it
>> seems that our unicorn processes are accepting connections long before
>> they can service them.
> Now that I think about it, it could be possible you're making a small
> mistake in your linux-tcp-listener-stats invocation.
> Unicorn listens to all interfaces on a specified port by default (e.g.
>  So you would need to point linux-tcp-listener-stats.rb
> to "" and not "" (which may provide you
> with a list of _client_ sockets, not server ones).

God, you're right. This is the problem. When I switch the invocation
it works as expected.

> It looks like raindrops may double count localhost connections
> because the client and server address are on the same host, and
> I seem to recall tcpdiag not providing a way to filter those...
> (It is very late at night for me, so I'll look at this again
> sometime tomorrow.)

In this test, its counting the ab connections on, but the
server connections on


  reply	other threads:[~2010-07-01 18:37 UTC|newest]

Thread overview: 9+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2010-06-29  0:26 queued is always 0 Ryan King
2010-06-29  0:50 ` Eric Wong
2010-06-30  1:12   ` Ryan King
2010-06-30  2:45     ` Eric Wong
2010-06-30 18:51       ` Ryan King
2010-06-30 23:18         ` Ryan King
2010-07-01 10:18         ` Eric Wong
2010-07-01 18:37           ` Ryan King [this message]
2010-07-02  2:51             ` Eric Wong

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:

  List information: http://yhbt.net/raindrops/

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=AANLkTinP5fBfcRa6Y3-MrMzDqjRRrgEOIunvn_h7q2WS@mail.gmail.com \
    --to=ryan@twitter.com \
    --cc=raindrops@librelist.com \


* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
Code repositories for project(s) associated with this public inbox


This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).