Section: New Results

Load-balancing with TreeMatch

In the context of the Joint Laboratory for Petascale Computing (JLPC) included Inria and the University of Illinois at Urbana, we developed two load balancers for Charm++.

The two load-balancers we wrote take into account both the computing power and the hierarchical topology depending on the fact that the application is compute-bound or communication-bound. This work is based on our TreeMatch library that computes process placement in order to reduce an application communication costs based on the hardware topology. Compared to some other solutions based on weighted topologies (latency, bandwidth, ...), ours is fully dynamic because we use only a qualitative approach for our representation of the hardware architecture.

The first load balancer is designed for compute-bound applications as it favours the leveling of CPU loads. The second load balancer focuses on communication-bound applications as it first reduces the congestion on the upper links in the topology tree.

These two load balancers gave us improvements for some applications up to 10% of the execution time.