Center for Petascale Computing
A unique capability in Charm++/AMPI, and one that makes a compelling case for its adoption,
is measurement-based dynamic load balancing, relying on our empirical observation called the
principle of persistence: once an application has been expressed in terms of its natural objects, the
object loads and communication patterns persist over time, even in a dynamic application whose
load evolves over time. The RTS mediates communication and scheduling of objects. Therefore,
it can easily keep track of the CPU time consumed by each object, and the number of bytes and
messages exchanged between objects. Since the principle of persistence implies that the recent
past is a good predictor of the near future, the RTS can then rebalance the computation and
communication by reassigning objects to processors. The RTS includes a suite of load balancing
strategies that are useful in varying circumstances. However, some
open research issues still remain in the face of large petascale machines. One major challenge is the scalability and efficiency of the load balancer itself on petascale
machines. We have proposed and developed a prototype hierarchical scheme that combines the
advantages of those centralized and distributed schemes that potentially can scale to petascale parallel
machines. We plan to explore topology-aware load balancing strategies in the scalable
hierarchical load balancer which improves the work-to-processor mapping by taking the machine
topology into account in a distributed fashion.
Investigator: L.V. Kale