A visual performance analysis framework for task-based parallel applications running on hybrid clusters

摘要

Programming paradigms in High-Performance Computing have been shifting toward task-basedrnmodels that are capable of adapting readily to heterogeneous and scalable supercomputers. Thernperformance of task-based application heavily depends on the runtime scheduling heuristicsrnand on its ability to exploit computing and communication resources. Unfortunately, the traditionalrnperformance analysis strategies are unfit to fully understand task-based runtime systemsrnand applications: they expect a regular behavior with communication and computation phases,rnwhile task-based applications demonstrate no clear phases. Moreover, the finer granularity ofrntask-based applications typically induces a stochastic behavior that leads to irregular structuresrnthat are difficult to analyze. Furthermore, the combination of application structure, scheduler,rnand hardware information is generally essential to understand performance issues. This paperrnpresents a flexible framework that enables one to combine several sources of information andrnto create custom visualization panels allowing to understand and pinpoint performance problemsrnincurred by bad scheduling decisions in task-based applications. Three case-studies usingrnStarPU-MPI, a task-based multi-node runtime system, are detailed to show how our frameworkrncan be used to study the performance of the well-known Cholesky factorization. Performancernimprovements include a better task partitioningamongthemulti-(GPU, core) toget closer to theoreticalrnlower bounds, improved MPI pipelining inmulti-(node, core,GPU) to reduce the slow start,rnand changes in the runtime system to increaseMPI bandwidth, with gains of up to13%in the totalrnmakespan.