Event model for tracking dependencies of MPI partitioned point-to-point communication

• Ereignismodell zum Nachvollziehen von Abhängigkeiten in MPI Partitioned Point-to-Point Communication

Thärigen, Isabel; Müller, Matthias S. (Thesis advisor); Wolf, Felix (Thesis advisor); Hermanns, Marc-André (Consultant); Geimer, Markus (Consultant)

Aachen : RWTH Aachen University (2023)
Master Thesis

Masterarbeit, RWTH Aachen University, 2023

Abstract

The new MPI 4.0 standard introduces the concept of partitioned point-to-point (P2P) communication. While this new subtype of persistent point-to-point communication brings many advantages, especially for multi-threaded MPI programs, it also introduces challenges for measurement and analysis tools which have to integrate the new concept in their execution. The tools work on profiles or traces which follow a specific event model that specifies which events are recorded for which operations as well as the information they provide. To enable tools to integrate partitioned P2P communication into their analysis, the event model has to provide all information necessary to understand the communication that happened in the original application. This work discusses possibilities to extend the well-known OTF2 event model to support partitioned P2P communication and introduces seven new events to do so. The functionality of this extension is demonstrated by explaining how the communication visualization in Vampir and the wait-state analysis in Scalasca can be extended to integrate partitioned P2P communication. The latter is also implemented in form of a prototype and calculates the desired results for a set of small test cases with partitioned P2P communication.

Institutions

• IT Center [022000]
• Department of Computer Science [120000]
• Chair of Computer Science 12 (High Performance Computing) [123010]