Evaluating locality-aware extensions for task migration in distributed memory

Marvin Porsil, Jannis Klinkenberg, Matthias S. Müller, Karl Fuerlinger
2021
Ich versichere hiermit, dass ich die vorliegende Arbeit selbständig und ohne Benutzung anderer als der angegebenen Hilfsmittel angefertigt habe. Alle Stellen, die wörtlich oder sinngemäß aus veröffentlichten und nicht veröffentlichten Schriften entnommen sind, sind als solche kenntlich gemacht. Die Arbeit ist in gleicher oder ähnlicher Form noch nicht als Prüfungsarbeit eingereicht worden. Abstract In HPC applications one approach for dynamic load balancing is the usage of taskbased
more » ... ion frameworks, which are suited for irregular workloads. The OpenMP tasking construct however is restricted to the boundaries of a single process. Chameleon, a library for reactive load balancing for hybrid MPI+OpenMP task-parallel applications, provides a way to balance the load in distributed memory systems, across process boundaries. This thesis presents the following three approaches to optimize Chameleon. A task affinity extension considering data locations, which may be accessed by tasks, when selecting a task to execute. The affinity extension showed performance improvements of up to 35% when compared to the baseline runtime in multiple scenarios. Considering the topological distances between computing nodes or processes running on these nodes in a cluster network when choosing a migration victim for load balancing proved to result in only small improvements, depending on the size of the tasks. The utilization of the communication thread of Chameleon for the execution of tasks, as Chameleon currently uses one thread per rank exclusively for communication to handle the load balancing. Utilizing the communication thread for task execution additional to the load balancing resulted in runtime improvements when the degree of contribution allows for enough communication, which is dependent on the degree of imbalance.
doi:10.18154/rwth-2021-09161 fatcat:sov5xbys2jcfnioyy7asci2ybm