Pipe issues:

• Have the overhead of starting a separate process. (you may or may not have this overhead when doing a TCP socket connection depending on if the server is forking or threading)

• Most OSes have highly optimized TCP/IP stacks, generally much more optmized than the OSes core pipe processing functions. For today's applications network performance is typically much more important than pipe performance.
• localhost TCP/IP connections (where client and server are on the same host) do not go out "on the wire", they are typically highly optimized with direct memory-to-memory exchange of the data. Localhost network connections have practially 0 network overhead.

i stand absolutely corrected. i guess i am showing the fact that some of the things i learned are no longer quite as accurate as they were when i picked them up :) so, i wrote a bit of code to test this out:

#!/usr/bin/perl -w

use strict;
use IO::Socket;
use Benchmark;

my $to_mail_addy     = "junk\@????????.net";
my $from_mail_addy   = "junk\@????????.net";
my $mailhost         = "localhost";

sub with_pipe {
   open(SENDMAIL, "|/usr/sbin/sendmail") or die "cannot open sendmail"
+;
   print SENDMAIL "To: $to_mail_addy\n";
   print SENDMAIL "From: $from_mail_addy\n";
   print SENDMAIL "Subject: A speed test...\n";
   print SENDMAIL "Content-type: text/plain \n\n";
   print SENDMAIL "Blah blah blah... this is my content...";
   close(SENDMAIL);
}

sub with_socket {
   my $sock = new IO::Socket::INET(PeerAddr => 'localhost',
                                    PeerPort => 25,
                                    Proto    => 'tcp');

   die "unable to create $sock" unless defined $sock;

   print $sock "Hello arino.net\n";
   print $sock "MAIL FROM: $from_mail_addy\n";
   print $sock "RCPT TO: $to_mail_addy\n";
   print $sock "DATA\n";
   print $sock "Subject: A speed test...\n";
   print $sock "Blah blah blah... this is my content...";
   print $sock ".\n";
   print $sock "quit\n";
   close($sock);
}

timethese (1000, {"WITH PIPES" => \&with_pipe, 
          "WITH SOCKET" => \&with_socket });


[ed@darkness ed]$ perl speedtest.pl 
Benchmark: timing 1000 iterations of WITH PIPES, WITH SOCKET...
WITH PIPES: 22 wallclock secs ( 0.26 usr  0.68 sys +  6.95 cusr  4.93 
+csys = 12.82 CPU) @ 1063.83/s (n=1000)
WITH SOCKET: 17 wallclock secs ( 1.76 usr +  1.13 sys =  2.89 CPU) @ 3
+46.02/s (n=1000)
[download]

so we see that sockets ARE much faster than pipes, which to me was counter intuitive... I guess they really have optimized the bejeezus out of the TCP stack... time to have a little chat with my networking professor :)


UPDATE: out of curiosity, people have asked what the system specs were like on the machine i ran the tests on:

Processor	Pentium II (Deschutes) 348.491034 Mhz
RAM	Memtotal - 193060 kB MemFree - 20864 kB
OS	Linux 2.2.13 i686 (redhat)

Eduardo, as we mentioned in the Chatterbox, I think the problem with this benchmark lies with process creation, not TCP vs. PIPEs. With your PIPE code, you are spawning a new process of sendmail each time, however with your TCP code, sendmail is already running.

Additionally, I have a test case on my computer at home (which I cannot get to easily at the moment), with a GO client/server program. Before I separated the two programs, the run time was about 25 seconds, full CPU usage. Once I moved to TCP my run time jumped to almost 2 minutes with CPU never topping over 30-40%. Changing my code once more to use UNIX sockets, I regained full CPU usage and an average run time of 29 seconds. I did not implement a PIPE solution, but I believe UNIX sockets are implemented as a PIPE so the number should between 25-29 seconds if I did, no? :)

Anyway I just wanted to post officially something before everyone ran off making all there PIPE'd programs TCP! :)

Ciao,
Gryn

(p.s. By GO client/server I mean the board game GO, the server sends a 19x19 text array to the client and the client responds back with a move in this case two random numbers. in the test cases mentioned above this was repeated for about 1,000 moves)