|
Hi,
we're happily using LVS (DR+wrr) for balancing load (http/http/soap as well
as an inhouse-designed session service) of currently four clusters with a
total of around 45 realservers.
As a central syslog server is not yet overloaded but wouldn't be useless
to prepare some clustering, I thought about adding a balancer for syslog
services using the sourcehash scheduler (so that all logs of a distinct
host is only sent to exactly one realserver, easing file aggregation and
distribution a lot).
Problem:
-syslog is udp
-sh, rr and lc seem to balance tcp well, but udp performs somewhat
unbalanced:
IP Virtual Server version 1.0.7 (size=4096)
Prot LocalAddress:Port Scheduler Flags
-> RemoteAddress:Port Forward Weight ActiveConn InActConn
UDP 172.21.0.2:514 sh
-> 172.21.0.49:514 Route 1 0 6
-> 172.21.0.48:514 Route 1 0 3541
TCP 172.21.0.2:101 sh
-> 172.21.0.49:101 Route 1 0 7
-> 172.21.0.48:101 Route 1 0 4
101/tcp is a test service, 514/udp the "real" syslog; test service was tried
from a few single machines, syslog is live-syslog from a few hundred
different hosts. Not only sh but also rr and lc were tested - with the
same results.
Ok, according to ipvsadm-howto par 5.17 this is somehow expected:
"On a short time scale (ca. timeout), there will be no load balancing
of UDP services (e.g. as was found for ntp). All requests will go to
the same realserver. On a long time scale (>>timeout) loadbalancing will
occur."
In fact, on a long time scale (over the night) very few requests were sent
to the second realserver; but as I reduced the timeout to a few seconds, the
load was distributed fine along the realservers:
IP Virtual Server version 1.0.7 (size=4096)
Prot LocalAddress:Port Scheduler Flags
-> RemoteAddress:Port Forward Weight ActiveConn InActConn
UDP 172.21.0.2:514 sh
-> 172.21.0.49:514 Route 1 0 1174
-> 172.21.0.48:514 Route 1 0 1191
From a
# while /bin/sleep 1 ; do ipvsadm -Ln -c --sort | grep ${srchost}:514 ; done
packets are balanced to the same realserver for $srchost (for a couple
of hosts).
As the first test gave somehow strange results (balancing 99% to a single
host), I'd like to know if I can expect the balancer to work as sh is
expected to do (really always balance by source-ip always to the same
realserver) or if somehow things might break up like they do when
having the udp-timeout at 300 seconds.
Regards,
Anders
|
