O(n^2) or worse ? cat /proc/net/tcp

O(n^2) or worse ? cat /proc/net/tcp

[Matti Aarnio]

I have about 750 TCP socket on one multi-threaded server application, and 
about 250 on three single-threaded clients.

I did observe odd slowness at first with  "lsof -p .."  program, then at
"netstat -na".  Same slowdown happens with:  cat /proc/net/tcp
(that 'cat' took 25 seconds...)

Sure this is not earth shatterinly dangerous thing, rather a nuisance
when a monitoring application is not as fast as one would expect.


Application works fine, though.

  /Matti Aarnio

Re: O(n^2) or worse ? cat /proc/net/tcp

[David Miller]

From: Matti Aarnio <matti.aarnio@zmailer.org>
Date: Mon, 24 Mar 2008 00:43:07 +0200

> I have about 750 TCP socket on one multi-threaded server application, and 
> about 250 on three single-threaded clients.
> 
> I did observe odd slowness at first with  "lsof -p .."  program, then at
> "netstat -na".  Same slowdown happens with:  cat /proc/net/tcp
> (that 'cat' took 25 seconds...)
> 
> Sure this is not earth shatterinly dangerous thing, rather a nuisance
> when a monitoring application is not as fast as one would expect.

There have been some improvements in this area over the
past few releases.

If you tell us what version you're running we can try to
dig out the changes that matter.

Re: O(n^2) or worse ? cat /proc/net/tcp

[Matti Aarnio]

On Sun, Mar 23, 2008 at 04:20:21PM -0700, David Miller wrote:
> From: Matti Aarnio <matti.aarnio@zmailer.org>
> Date: Mon, 24 Mar 2008 00:43:07 +0200
> 
> > I have about 750 TCP socket on one multi-threaded server application, and 
> > about 250 on three single-threaded clients.
> > 
> > I did observe odd slowness at first with  "lsof -p .."  program, then at
> > "netstat -na".  Same slowdown happens with:  cat /proc/net/tcp
> > (that 'cat' took 25 seconds...)
> > 
> > Sure this is not earth shatterinly dangerous thing, rather a nuisance
> > when a monitoring application is not as fast as one would expect.
> 
> There have been some improvements in this area over the
> past few releases.
> 
> If you tell us what version you're running we can try to
> dig out the changes that matter.

Bleeding edgeish Fedora:

  2.6.25-0.90.rc3.git5.fc9 #1 SMP Tue Mar 4 20:19:33 EST 2008 x86_64 x86_64 x86_64 GNU/Linux

/Matti Aarnio

Re: O(n^2) or worse ? cat /proc/net/tcp

[Andi Kleen]

Matti Aarnio <matti.aarnio@zmailer.org> writes:
> 
> I did observe odd slowness at first with  "lsof -p .."  program, 

You could oprofile it.

> then at
> "netstat -na".  Same slowdown happens with:  cat /proc/net/tcp
> (that 'cat' took 25 seconds...)

/proc/net/tcp should be only O(n^2) if the application does a lot of 
open()/seek()/read()/close() for very small chunks, but at least netstat
doesn't do that. Not sure about lsof though.

-Andi

Re: O(n^2) or worse ? cat /proc/net/tcp

[Matti Aarnio]

On Mon, Mar 24, 2008 at 05:06:28AM +0100, Andi Kleen wrote:
> Matti Aarnio <matti.aarnio@zmailer.org> writes:
> > 
> > I did observe odd slowness at first with  "lsof -p .."  program, 
> 
> You could oprofile it.

Ah, I forgot to start learning of oprofile..

> > then at
> > "netstat -na".  Same slowdown happens with:  cat /proc/net/tcp
> > (that 'cat' took 25 seconds...)
> 
> /proc/net/tcp should be only O(n^2) if the application does a lot of 
> open()/seek()/read()/close() for very small chunks, but at least netstat
> doesn't do that. Not sure about lsof though.

Or  cat  -- which definitely does not seek around..

> -Andi

  /Matti Aarnio

Re: O(n^2) or worse ? cat /proc/net/tcp

[Matti Aarnio]

On Sun, Mar 23, 2008 at 04:20:21PM -0700, David Miller wrote:
> From: Matti Aarnio <matti.aarnio@zmailer.org>
> Date: Mon, 24 Mar 2008 00:43:07 +0200
> 
> > I have about 750 TCP socket on one multi-threaded server application, and 
> > about 250 on three single-threaded clients.
> > 
> > I did observe odd slowness at first with  "lsof -p .."  program, then at
> > "netstat -na".  Same slowdown happens with:  cat /proc/net/tcp
> > (that 'cat' took 25 seconds...)
> > 
> > Sure this is not earth shatterinly dangerous thing, rather a nuisance
> > when a monitoring application is not as fast as one would expect.
> 
> There have been some improvements in this area over the
> past few releases.
> 
> If you tell us what version you're running we can try to
> dig out the changes that matter.


I cooked up a small perl "client-server application" (at the end) that
opens up several hundred TCP sockets in between its two halfs.
It doesn't do anything else, that is the machine is _IDLE_ after socket
creations:

$ time cat /proc/net/tcp|wc;time cat /proc/net/tcp|wc;time cat /proc/net/tcp|wc
   2215   35650  332250

real    0m35.731s
user    0m0.005s
sys     0m35.721s
   1815   29445  272250

real    0m25.328s
user    0m0.004s
sys     0m25.309s
   1815   30850  272250

real    0m29.521s
user    0m0.004s
sys     0m29.508s


Kernel:
 2.6.25-0.90.rc3.git5.fc9 #1 SMP Tue Mar 4 20:19:33 EST 2008 x86_64 x86_64 x86_64

Machine:

vendor_id       : AuthenticAMD
cpu family      : 15
model           : 35
model name      : AMD Athlon(tm) 64 X2 Dual Core Processor 4400+
stepping        : 2
cpu MHz         : 2211.331

bogomips        : 4425.68



Odd..  Another machine with slightly older kernel is definitely swifter
although it clocks less than half of the bogomips..

$ time cat /proc/net/tcp | wc
   1822   30969  273300

real    0m6.674s
user    0m0.008s
sys     0m6.669s


 2.6.25-0.40.rc1.git2.fc9 #1 SMP Wed Feb 13 17:17:48 EST 2008 x86_64 x86_64 x86_64 GNU/Linux


vendor_id       : AuthenticAMD
cpu family      : 15
model           : 75
model name      : AMD Athlon(tm) 64 X2 Dual Core Processor 4200+
stepping        : 2
cpu MHz         : 1000.000

bogomips        : 2006.08





#!/usr/bin/perl                                                                                                                     
#  Simple test load generator for Linux kernel related                                                                              
#  "cat /proc/net/tcp" performance issue..                                                                                          

use IO::Socket::INET;
my @rd = ();
my @wr = ();

my $rc = fork();
if ($rc == 0) { ## Child                                                                                                            
    $srv = IO::Socket::INET->new( Type => SOCK_STREAM, Listen => 900, LocalPort => 9988,
                                  ReuseAddr => 1, Blocking => 1 );
    while (1) {
        my $s = $srv->accept();
        push @rd, $s;
    }
    exit;
}

sleep 1; # make sure the "server" starts before our client-side does connects..

foreach my $i (1..900) {
    my $s = IO::Socket::INET->new( Proto => TCP, PeerPort => 9988, PeerAddr => 'localhost' );
    push @wr, $s;

};

while (1) {
    sleep 100;
}

exit 0;

Re: O(n^2) or worse ? cat /proc/net/tcp

[David Miller]

From: Matti Aarnio <matti.aarnio@zmailer.org>
Date: Thu, 27 Mar 2008 13:49:11 +0200

> Odd..  Another machine with slightly older kernel is definitely swifter
> although it clocks less than half of the bogomips..

The machine with the faster cpu likely has much more memory,
thus the hash tables allocated are much larger, thus the
table scan is significantly more expensive.