access_time 04 de dezembro de 2019 às 15:00 até 04 de dezembro de 2019 às 17:00
place Sala V0.15 (Piso 0, Pavilhão de Civil), IST, Alameda
With the advent of Chip-Multiprocessors (CMPs), Transactional Memory (TM) emerged as a powerful paradigm to simplify parallel programming. Unfortunately, as more and more cores become available in both affordable and high-end parallel machines, the scalability limits of a wide class of TM applications become more evident. Many TM applications reach a performance plateau after a certain number of threads, and most of such applications even observe performance drops after such plateau. Therefore, online parallelism tuning techniques were proposed to find the optimal number of threads that a TM application needs. However, state-of-the-art solutions are exclusively tailored to single-process scenarios with rather static workloads and they perform poorly when running in dynamic multi-process environments. This work proposes novel methods for parallelism tuning and space-sharing for co-located transactional processes. Our proposals let the parallel applications set their parallelism level, so that they can, fairly and efficiently, space-share the system and the system does not become oversubscribed. Our evaluation with different workloads and scenarios shows that using our methods substantially boosts the system’s overall performance with respect to the state-of-the-art parallelism tuning techniques and the system converges to a fair and efficient state, unlike the other techniques in multi-process environments.
local_offer Prova de Doutoramento
person Candidato: Amin Mohtasham
supervisor_account Orientador: Prof. João Pedro Faria Mendonça Barreto