Probabilistic loop scheduling considering communication overhead

Data dependent loop scheduling based on genetic algorithms for distributed and shared memory systems

“…The performance of our approach is evaluated by comparing the resulting schedule length. For the set of experiments we test the percentage of individuals of the final population that represent the optimal scheduling, the execution time to obtain it, the schedule length of the best individuals of our model (we compare that with previous results [4,12,18]) or the number of iterations to obtain optimal results. In our experiments, we assume a fully connected homogeneous distributed-memory system composed of 8 processors ðK ¼ 8Þ and each operation (task) takes 1 time unit.…”

“…Refs. [12,14] present a technique, which is applied to a communication sensitive data flow graph that represents a nested loop. This technique takes into account the data transmission time, the loop carried dependencies, and the target architecture to schedule the set of nodes of the graph.…”

“…Each of the algorithms assumes that we work on a shared-memory system. Many other works have studied the parallelism across iterations to consider the loop carried dependencies [4,8,[12][13][14]16,17]. Refs.…”

“…Previous approaches have attempted to achieve the minimum completion time for the parallel loop scheduling problem only by distributing the workload as evenly as possible while minimizing the number of synchronization operations required and the communication overhead caused by access to non local data on shared-memory systems [2,3,5,6,11,18]. Other authors have studied the parallelism across iterations to consider loop carried dependencies, proposing different techniques to improve this parallelism [4,8,[12][13][14]16,17]: cyclo-compaction scheduling, loop pipelining, etc.…”

Pitfalls in Parallel Job Scheduling Evaluation

“…If a parameter space exploration is not feasible, model the communication and detail all the assumptions made [26,60], preferably basing the model on real workloads or measurements. Confidence intervals can also be used to qualify the results [69]. Research.…”