A new load balancing algorithm for discussion and consideration

Jeremy Kerr jeremy at redfishsoftware.com.au
Fri May 9 00:58:02 BST 2003


> The problem with this policy is that it generates quite a bit of network
> traffic between load balancer and the servers and therefore doesn't
> perform well on highly loaded large systems.

With the simulaton cluster I was trying feedbackd on (4 servers, 1 director), 
a health report once every 2 seconds was sufficient to prevent the 'herd 
effect'. With eight or less services being monitored on each server, the 
control packets are 52 bytes each, and require an equal sized ack - so will 
fit into an ethernet frame easily (52 byte packet + 20 byte IP header + 20 
byte TCP header < 1500 byte MTU).  Assuming that the first NECP ack will 
include the TCP ack, we're transmitting three distinct frames over the 
physical layer every two seconds (two necp packets, plus the tcp ack for the 
second).

Given Joseph Mack's ~8000 frames per second figure for 100Mbps ethernet, each 
NECP session will use 1.5/8000 of the medium - 0.02% of available bandwidth 
per server.

Granted, 4 servers isn't a huge cluster, but compared to the network load of 
the service provision, the load from monitor processes isn't large at all. I 
think the bigger concern is the resource load on the director from the master 
process handling load updates.


Jeremy


More information about the lvs-users mailing list