|
>Hi
>
>To digress slightly,
>
>How about this for a scenario,
Hi,
The balancing strategy you explain sounds a lot like Adaptive Separable
Policy described in [WS88] and studied for example in [BZ92].
This policy, like NQ, is derived from SED. For every server i it uses a
cost function AS(i)=(ni+1)/(ui*pi), where ni current job count on server
i, ui is the fixed service rate ("weight") for server i, and pi is the
current utilization of server i. Note that like in SED, current job
count is incremented by 1. pi is the fraction of time that the server i
has been idle weighted exponentially to the latest load levels. Every
server measures its own load levels, and sends the value periodically to
the load balancers.
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. A lot of traffic is needed
for the load level information to be current at the load balancer. If
that information isn't current, it has no practical use.
Regards,
Marko
[WS88] A. Weinrib and S. Shenker, Greed is not enough: Adaptive load
sharing in large heterogeneous systems. In Proceedings IEEE INFOCOM'88,
pages 986-994, 1988.
[BZ92] S. A. Banawan and N. M. Zeidat, A comparative study of load
sharing in heterogeneous multicomputer systems. 25th Annual Symposium on
Simulation, pages 22-31, may 1992.
|
