Στο πλαίσιο της διοργάνωσης των σεμιναρίων του τμήματος θα πραγματοποιηθεί διαδικτυακά μέσω της εφαρμογής Ms Teams την Παρασκευή 17/09/2021 και ώρα 12:00, ομιλία με τίτλο «Non-blocking operations in GraphBLAS: tradeoffs and performance». Ομιλητής θα είναι ο  κ. Αριστείδης Μάστορας, Postdoctoral Researcher, Huawei Zurich Research Center. Μπορείτε να παρακολουθήσετε την ομιλία μέσω του παρακάτω συνδέσμου: Link MsTeams

ΠΕΡΙΛΗΨΗ

GraphBLAS is a recent standard that allows expression of graph algorithms in the language of linear algebra and enables easy code parallelization and optimization. Programs executed with our GraphBLAS implementation are automatically vectorized and parallelized by choosing, at compile-time, one of the available backends for shared and/or distributed memory. In this talk, we present the design of a new backend for shared memory systems that aims at improving data locality by executing GraphBLAS operations in a non-blocking fashion. The GraphBLAS API remains the same, and the non-blocking backend achieves faster execution of programs as a result of better data locality. The non-blocking backend exploits lazy evaluation and dynamic parallelism to group together the execution of operations that reuse data in cache. A preliminary evaluation of the non-blocking backend shows promising results for various datasets of two algorithms written in GraphBLAS.