Re: [lvs-users] ip_vs_send_async error

To: " users mailing list." <lvs-users@xxxxxxxxxxxxxxxxxxxxxx>
Subject: Re: [lvs-users] ip_vs_send_async error
From: Siim Põder <siim@xxxxxxxxxxxxxxx>
Date: Mon, 14 Dec 2009 13:10:06 +0200

Aleksey Chudov wrote:
> All test was on the same machine. There is no problems with only few
> connections while testing. Now we have above 8000 conn/ps. But it is only
> half of planned rate. I can see connections from lvs master on lvs backup so
> connection sync is work. May be not all connections is synchronized, but I
> don't know how to test this.

The kernel only wakes connsync thread up every second to send sync
messages in one batch. When the UDP send buffer cant fit all that data
any more, you will get those errors on sync. To get around it, make
your buffers bigger.

net.core.{w,r}mem_{max,default} are the right sysctls IIRC. Also be
sure to increase the receiver side buffer, otherways you will drop the
packets on the other end and that is more difficult to spot (Udp
InErrors in netstat -s can hint about it, for example).

Alternatively, you may want to build your kernel with the following
patch (sorry about the broken wrapping, starts the sync operation 10
times as often):

--- a/net/ipv4/ipvs/ip_vs_sync.c        2008-02-11 05:51:11.000000000
+++ b/net/ipv4/ipvs/ip_vs_sync.c        2008-10-06 12:17:07.883882381
@@ -679,7 +686,7 @@
                if (stop_master_sync)

-               msleep_interruptible(1000);
+               msleep_interruptible(100);

Please read the documentation before posting - it's available at: mailing list - lvs-users@xxxxxxxxxxxxxxxxxxxxxx
Send requests to lvs-users-request@xxxxxxxxxxxxxxxxxxxxxx
or go to

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