Load Balancing

[Stuart Card]

At 01:21 PM 2000-04-06 +1000, richard@iguana.co.nz wrote:

>Ok, first up, love the load balancing thus far, good stuff. My question is
>basically looking at how this can be further advanced, particular with
>then end in mind of getting the fastest response possible from the
>balanced service.
>
>Currently, to my knowledge, we utilise a simple count of who has the least
>connections to work out which of the balanced servers gets the next
>connection. I was wondering if anyone had experimented or implemented more
>advanced algorithms...

We are working on a packet-by-packet approach to load balancing.
It doesn't address the issue of server load, only network load; but
rather than balancing a count of connections, or even an estimated
traffic load based upon connections weighted with their expected loads,
it forwards each packet down the path that _at this particular moment_
seems appropriate (based upon path bandwidth, latency, drop rate,
error rate, queue length, buffer memory size, etc.).  It is probably
not scalable to very large numbers of paths nor to high-bandwidth
paths, but we are dealing with small numbers (1-4) of low-speed
(<=64kbps) wireless links, each of which is the first hop of a
different end-to-end path.

------------------------------------------------------------------------
Stuart W. Card, Chief Scientist & Vice-Pres., Critical Technologies Inc.
Suite 400 Technology Center, 4th Floor 1001 Broad Street, Utica NY 13501
315-793-0248   FAX -9710    <stu@critical.com>   http://www.critical.com