Predictive, reactive and replication-based load balancing of tasks in Chameleon and sam(oa) ²

Abstract

Increasingly complex hardware architectures as well as numerical algorithms make balancing load in parallel numerical software for adaptive mesh refinement an inherently difficult task, especially if variability of system components and unpredictability of execution time comes into play. Yet, traditional predictive load balancing strategies are largely based on cost models that aim to predict the execution time of computational tasks. To address this fundamental weakness, we present a novel reactive load balancing approach in distributed memory for MPI+OpenMP parallel applications that is based on keeping tasks speculatively replicated on multiple MPI processes. Replicated tasks are scheduled fully reactively without the need of a predictive cost model. Task cancellation mechanisms help to keep the overhead of replication minimal by avoiding redundant computation of replicated tasks. We implemented our approach in the Chameleon library for reactive load balancing building upon previous work on reactive task migration. Our experiments in the parallel dynamic adaptive mesh refinement software sam(oa) 2 demonstrate performance improvements in the presence of wrong cost models and artificially introduced noise to simulate imbalances coming from hardware variability. CCS CONCEPTS• Computing methodologies → Parallel algorithms.

From reactive to proactive load balancing for task‐based parallel applications in distributed memory machines

“…Because of reactive task migration, we do call "offloading" tasks instead of "stealing," and reactive action is taken from the overloaded/slow processes. 16 As shown in Figure 4B, the decision time of offloading tasks (t k ) is earlier than work stealing in Figure 4A. A detailed example of how tasks are offloaded reactively can be found in Appendix A.…”

“…11,15 The following idea is task replication that aims at tackling unexpected performance variability. 16 However, this is difficult to know how many tasks should be offloaded at once and which processes are truly underloaded/fast in a short period. Without prior load knowledge, replication strategies need to fix the target process for replicas, such as left/right neighbor ranks.…”

“…Instead of waiting for a process queue empty, the reactive approach relies on monitoring the queue status to offload tasks from an overloaded process to underloaded targets § 11,15 . The following idea is task replication that aims at tackling unexpected performance variability 16 . However, this is difficult to know how many tasks should be offloaded at once and which processes are truly underloaded/fast in a short period.…”

“…When the imbalance condition is met, tasks can be offloaded in advance. This is why we can reduce the impact of migration overhead better than work stealing 11,16 . However, reactive operations might be settled wrong in the cases of high imbalance because the most current status only reflects as well as conjectures an unbalanced situation for a short period.…”

“…This is why we can reduce the impact of migration overhead better than work stealing. 11,16 However, reactive operations might be settled wrong in the cases of high imbalance because the most current status only reflects as well as conjectures an unbalanced situation for a short period. We still lack information about load and which process is a potential victim to offload tasks.…”