[RndTbl] NIC performance with iperf

john at johnlange.ca john at johnlange.ca
Sat Apr 17 22:28:22 CDT 2010

Isn't Windows still running a BSD decendent IP stack?

Just thought I'd troll a little ;)


Sent from my BlackBerry device on the Rogers Wireless Network

-----Original Message-----
From: "Adam Thompson" <athompso at athompso.net>
Date: Sat, 17 Apr 2010 14:34:39 
To: 'Continuation of Round Table discussion'<roundtable at muug.mb.ca>
Subject: Re: [RndTbl] NIC performance with iperf

> I installed a cheap GigE switch for a customer using 2006-era P4 Dell
> desktops,
> running XP with cheap Broadcom GigE NICs. Constant 600-700mbit for
> file
> transfers.
> CPU I think was ~30%, but that may have included the AV scanning the
> file.
> I assumed Iperf is developed/optimized natively under *NIX and the
> Windows
> version is compiled under Cygwin, do you have any indication that it's
> optimized?

I'm certain iperf is NOT optimized for Windows.  However, empirically I 
can determine using application-level tests (i.e. FTP, SAMBA, etc.) that 
the numbers I'm seeing are consistent.

Worth mentioning is that Microsoft shipped "ttcp.exe" on most versions of 
Windows XP install media, under VALUEADD\NET\something, IIRC.  Presumably 
that binary is reasonably-well optimized.  I recall using it to test 
between Windows XP and Solaris, and seeing half-decent speeds but nothing 
wonderful (~400Mbps).

Equally, I know that I too, have seen lesser Windows clients talking to 
Windows servers at higher speeds.  It's also worth noting that SMB traffic 
to my Linux box is noticeably slower than it was to the Sun Ultra 
Enterprise 450 it replaced.  While there's obviously something wrong with 
the networking stack on my Windows 7 PC (100% CPU utilization just to do 
TCP?!) it's also possible that there's some specific interaction between 
the Linux and Windows stacks.

As I demonstrated, Linux-to-linux works great.  Kevin demonstrated that 
MacOSX-Linux works great.  Many of us have anecdotal evidence that 
Windows-Windows works well, and several people have recently demonstrated 
that Windows-Linux does NOT work well.   And I've personally demonstrated 
(although I haven't posted any numbers here) that Windows-Solaris (*not* 
OpenSolaris, if it makes any difference) works better than Windows-Linux.

(OK, I engaged in a bit of Windows-bashing when I blamed the Windows 
networking stack *exclusively*.)

I don't have any desire to set up a test lab in my living room just to try 
out the various combinations... how much effect can the GigE switch have 
on performance?  I'm using dirt-cheap D-Link 5-port switches if that makes 
any difference.

-Adam Thompson
 <athompso at athompso.net>
 (204) 291-7950

Roundtable mailing list
Roundtable at muug.mb.ca

More information about the Roundtable mailing list