Autonomic Concurrency Regulation in Transactional Memory

Participant: Diego Rughetti

Home Institution: 
Sapienza Università di Roma 

Home Country: Italy

Host: Prof. Paolo Romano

Host Institution: INESC-ID
 
Host Country: Portugal

Start Date:
 2013-11-25

End Date: 2013-12-20 

Description: 
Transactional Memory (TM) has emerged as a powerful programming paradigm for concurrent applications. TM allows encapsulating the access to data shared across concurrent threads within transactions, thus avoiding the need for synchronization mechanisms to be explicitly coded by the programmer. On the other hand, synchronization transparency must not come at the expense of performance. Hence, TM-based systems must be enriched with mechanisms providing optimized run-time efficiency. Among the issues to be tackled, a crucial one is related to determining the optimal level of concurrency (number of threads) to be employed for running the application. When using too low levels of concurrency, the intrinsic parallelism of applications may not be fully untapped. On the other hand, an excessively high concurrency level may lead to thrashing phenomena caused by excessive data contention and consequent transaction aborts.
In this relation we present the results of a study aimed to evaluate to what extent existing techniques for self-tuning the concurrency level of Software Transactional Memory can be effectively used in the context of Hardware Transactional Memory (Intel TSX). Our study highlights that the techniques developed for STM can exhibit significant drawbacks when employed in a HTM-based system. This finding has motivated the investigation of two novel approaches for optimal concurrency level prediction explicitly tailored for Hardware Transaction Memory. By means of an experimental study based on the STAMP benchmark suite, we show that the proposed techniques are not only lightweight, but also that they can achieve a significantly higher accuracy than pre-existing optimization techniques proposed for STM.

Report: here