LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: [lvs-users] [LVS - NAT] InActConn in Ipvsadm with UDP stop increment

To: "LinuxVirtualServer.org users mailing list." <lvs-users@xxxxxxxxxxxxxxxxxxxxxx>
Subject: Re: [lvs-users] [LVS - NAT] InActConn in Ipvsadm with UDP stop incrementing during load test, and message are not correctly routed
From: Joseph Mack NA3T <jmack@xxxxxxxx>
Date: Wed, 26 Mar 2008 06:49:31 -0700 (PDT)
On Wed, 26 Mar 2008, pierrick grasland wrote:

> Hello,
>
> After correcting my hardware problem, I rerun my load test on LVS with a
> simple test for UDP.
>
> Actually, my setup is just the following :
>
> Client (CIP : 192.168.1.209) ==> Director  (VIP : 192.168.1.231, DIP:
> 192.168.2.231) ==> realserver (Only one, RIP : 192.168.2.3), with the
> following configuration :
>
> Client : XP, using SIPp in order to generate a lot of SIP messages in
> parallel (500 differents sockets open at max). These message are a single
> UDP packet, with only 425 Bytes each
> But, when I'm trying the same test, with UDP, I actually have very bad
> result (between 500 and 3000 successful call for test with ten of thousands
> calls). Results are UDP_test1.txt, UDP_test2.txt and UDP_test3.txt

There are a couple of problems with trying to figure out 
what's going on.

o LVS UDP has problems (see the HOWTO). AFAIK there isn't a 
test for high throughput single port UDP that we can use to 
check that UDP load balancing is working. ntp was used as a 
test in the early days, but the packet throughput is 
designed to be low. I don't know if your UDP test qualifies 
(it might, but you don't say so). Is your UDP test a single 
port send-reply system? If not should be easy enough to 
write a pair of programs in perl that are a udp ping-echo 
pair.

o no-one has SIP working on LVS, although people like 
yourself are putting some effort into it.

So at the moment it's hard to test LVS for its ability to 
load balance SIP when we don't know that it's load balancing 
UDP.

> It's seems that LVS doesn't increase InActConn anymore, 
> and, in SIPp, only the first calls are correctly routed, 
> all others finished as timeout.

When there are problems with UDP, won't there be a whole lot 
of ports tied up on the UDP server with packets that haven't 
found anywhere to go?

Joe
>

-- 
Joseph Mack NA3T EME(B,D), FM05lw North Carolina
jmack (at) wm7d (dot) net - azimuthal equidistant map
generator at http://www.wm7d.net/azproj.shtml
Homepage http://www.austintek.com/ It's GNU/Linux!


<Prev in Thread] Current Thread [Next in Thread>