TCP IP Retransmits every two minutes on WTS 2003 with Citrix 4.5, in the Performance Tuning / Server Sizing forum on BrianMadden.com

Reply

Stephan Ortman Posted: Tue, Jun 16 2009 12:52 PM

rated by 0 users

Hello - any insight is appreciated on the following issue:

We use a large Citrix XenApp server farm to host about 3000 concurrant users.  We publish an entire desktop to our user base.  95% of our users are on the local LAN and this problem happens for LAN users. 

In general, performance of our infrastructure is very good.  However, we do experience "typing stalls" and "screen stalls" periodically.  I followed some of the guidance provided on this website and ran Perfmon and worked with our network shop to conduct more in-depth troubleshooting. 

The "smoking gun" we have found for the typing stalls is that every 2 minutes (exactly), each server will experience a spike in TCP IP packet retransmits.  This spike lasts maybe two or three seconds, during which time typing will stall and/or any screen updates will stall, depending on what you are doing.

Our network shop did an in depth analysis of packet traffic from the server and says we should turn off TCP Keep Alives.  I've done some reading on Keep Alives and see there are ICA and Windows TCP K.A.s.  We have them set to various levels on our servers... I am not sure if I can turn them off completely - and if yes, what side effects that may have.

My questions:  Does the TCP Keep Alive settings we have something to do with the 2 minute spikes?  If yes, what settings should we adjust them to.  If no, are there other high-probability causes for 2 minute retransmit spikes we should investigate?

Thanks!  Steve

Reply

Alan Osborne replied on Tue, Jun 16 2009 4:09 PM

rated by 0 users

Hi,

A client of mine had a very similar issue, only the keepalives weren't getting ACKed which caused a TCP RESET to be issued. This was causing user sessions to be disconnected. The issue was traced back (using WireShark) to an F5 hardware load balancer with an incorrectly configured session timeout. We found a KB article from F5 that provided the recommended config changes on the Big IP for handling ICA traffic which corrected the problem.

If you are using a hardware load balancer, you might want to review its configuration.

Reply

Stephan Ortman replied on Fri, Jun 19 2009 4:23 PM

rated by 0 users

Thanks Alan - we use F5s heavily in our environment, so it's something I'll be sure to check.  Thanks again!