Bug #496

DRb.start_service(nil) is very slow

Added by Hongli Lai over 3 years ago. Updated 9 months ago.

[ruby-core:18381]
Status:Closed Start date:08/25/2008
Priority:Normal Due date:
Assignee:Masatoshi Seki % Done:

0%
Category:lib
Target version:1.9.2
ruby -v:-

Description

On some systems - such as mine - DRB.start_service(nil) is very slow. This is caused by the fact that DRb.open_server calls TCPServer.open(nil, 0).addr[1]. On my system, this takes about 3 seconds, and during those 3 seconds there is 0% CPU usage.

I suspect it has got something to do with the fact that retrieving the port of a server socket that's bound to 0.0.0.0 is for some unknown reason very slow.

The problem can be fixed by replacing the following line in drb.rb

  uri = 'druby://:0' unless uri

with:

  uri = 'druby://localhost:0' unless uri

This does not violate DRb.start_service's contract, because a nil indicates that it will bind to the default local host name.

The following monkeypatch works around the issue:

  module DRb
    class << self
      alias orig_start_service start_service
    end

    def self.start_service(uri = nil, front = nil, config = nil)
      if uri.nil?
        orig_start_service("druby://localhost:0", front, config)
      else
        orig_start_service(uri, front, config)
      end
    end
  end

Related issues

related to ruby-trunk - Feature #1811: Default BasicSocket.do_not_reverse_lookup to true Closed 07/23/2009

History

#1
Updated by Brian Candler over 3 years ago

I suspect a DNS/hostname resolution problem on your system.

Looking at 1.8.6p114, notice that DRbTCPSocket.getservername first calls Socket.gethostname to get the hostname, then calls Socket.gethostbyname which does both forward and reverse name resolution, presumably to map to a "canonical" name. Then DRbTCPSocket.open_server_inaddr_any also calls Socket.getaddrinfo, although it discards the results apart from using them to decide whether to bind to IPv4, IPv6, or both.

Try the following in irb - how long does each step take?

{{{
require 'socket'
h1 = Socket.gethostname
h2 = Socket.gethostbyname(h1)[0]
Socket.getaddrinfo(h2, nil, Socket::AF_UNSPEC, Socket::SOCK_STREAM, 0, Socket::AI_PASSIVE)
}}}

If one of these steps is slow, you may be able to solve the problem by adding an entry in /etc/resolv.conf to map your hostname to your primary interface IP.

#2
Updated by Hongli Lai over 3 years ago

All of those steps finish instantaneously.

I'm on Ubuntu 8.04.

#3
Updated by Brian Candler over 3 years ago

OK, how about this:

  require 'socket'
  t = TCPServer.new("", nil)
  t.addr
  t.addr

  Socket.do_not_reverse_lookup=true
  t.addr

On my machine, the first 't.addr' invocations take about 5 seconds. The last one is instant. I believe this is because of the delay looking up 0.0.0.0 in the DNS.

#4
Updated by Hongli Lai over 3 years ago

The last t.addr is instant for me as well.

#5
Updated by Hongli Lai about 3 years ago

Any update on this? Is my proposed fix acceptable?

#6
Updated by Roger Pack about 3 years ago

Sounds like the problem is that 
  require 'socket'
  t = TCPServer.new("", nil)
  t.addr

is slow, is that the root of the problem?
-=r

#7
Updated by Shyouhei Urabe about 3 years ago

  • Assignee set to Masatoshi Seki

#8
Updated by Marc-Andre Lafortune over 2 years ago

  • Category set to lib
  • ruby -v set to -

#9
Updated by Yusuke Endoh almost 2 years ago

  • Status changed from Open to Feedback
  • Target version set to 1.9.2
Hi,

I cannot reproduce the issue with trunk on Linux.
Does this still reproduce?
Could you elaborate your platform and network configuration?

-- 
Yusuke Endoh <mame@tsg.ne.jp>

#10
Updated by Roger Pack almost 2 years ago

This might be fixed now, since Socket.do_not_reverse_lookup defaults to true now in trunk (maybe somebody should note that in the NEWS?)
Thanks.
-rp

#11
Updated by Yusuke Endoh almost 2 years ago

  • Status changed from Feedback to Closed

Also available in: Atom PDF