Subject: Re: 2.0.2 and NAT xmit performance
To: Steve Paul <xeglon@earthlink.net>
From: Steven M. Bellovin <smb@cs.columbia.edu>
List: port-i386
Date: 06/02/2005 15:08:32
In message <429F56EC.9040708@earthlink.net>, Steve Paul writes:
> >> 3) Now go to client 192.168.1.20 to send the 200mb file directly to 
>site
> >> XYZ on the internet with FTP, it stalls on the first 1K for about 5-10
> >> seconds, then continues sending at ~1-2k/s tops.
>
> >Can you test transfering a big file from the client to the server?
> >This might be a local network problem (duplex mismatch or what have you)
> >Martin
>
>That was step 1 on the original email.  Sending a big, 200mb file from 
>the client 192.168.1.20 to server/gateway 192.168.1.1 is fast.. Full 
>network speed, 100base xfer rate.  Step 2 was to ftp that file from the 
>server to an external ftp site on the internet, which is also fast.  
>It's just when NAT comes into play it initially stalls for several 
>seconds, then continues very, very slow.
>
>For a test, I took my old 1.5Z server on the LAN as 192.168.1.70 and 
>tried ftping a file from LAN -> internet and it doesn't suffer the same 
>problem, so this is definately a Windows client issue.  We have four(4) 
>Windows XP PC's on the LAN, all with XP Pro + SP2 and all of them are 
>doing this.  They all have Yukon/Marvell Gigabit PCI network adapters so 
>it's definately a Windoze -> BSD issue instigated sometime between 1.6 
>and 2.0.2 as the 1.5Z server doesn't have this problem if I place it at 
>.1 and re-enable NAT.
>
>I'm going to keep plugging away and hope if anyone has any recollection 
>of things either in the SIP driver, NAT or ftp proxy that has changed 
>that might instigate this behavior.  tcpdump's haven't been too helpful 
>but I'm still analyzing them.. may attach one if it would be helpful.
>

Might it be an MTU issue with PPPoE?  Your original note mentioned DSL.
I see you have mssclamp specified; is that actually taking effect in 
the packets sent and received?  What happens if you lower that value to 
1420?


		--Steven M. Bellovin, http://www.cs.columbia.edu/~smb