Automatic Parameter Tuning in Distributed Load Balancer

What is it about?

In distributed computation, it is important that all the computers used remain busy as much as possible. As it is sometimes not possible to split the work evenly, dynamic load balancers relocate part of the load from busy nodes to idle nodes during the computation. Here, we focus on the so-called "lifeline-based global load balancer." In this scheme the main procedure consists in performing a certain amount of computation (the granularity) and checking if load balancing is needed. But just how much this computation amount should be is not obvious ...

Featured Image

Photo by Piret Ilver on Unsplash

Photo by Piret Ilver on Unsplash

Why is it important?

If the granularity is too small, the load balancing procedures are performed often and redundantly, diminishing the performance of the program. On the contrary, if the granularity is too large, the load balancing procedures are seldom performed, causing the performance of the program to decrease. A just balance needs to be found. To further complicate matters, the value to use depends on the computation given to the load balancer. To resolve this issue, we introduced a tuning mechanism which automatically adjusts the granularity during execution.