Performance Forecasting of Distributed Transactional Memory Systems

Participant: Pierangelo Di Sanzo

Home Institution: 
Sapienza Università di Roma 

Home Country: Italy

Host: Prof. Paolo Romano

Host Institution: INESC-ID
Host Country: Portugal

Start Date:

End Date: 2013-06-04 

This report describes the work I carried out and the results that have been achieved in the context of my Short Term Scientific Mission (STSM) at INESC-ID, where I had the opportunity of closely collaborating with Prof. Paolo Romano and the researchers in his group. The STSM has been focused on the analysis and the development of performance forecasting tools for Distributed Transactional Memories (DTM) systems, which is one of main topic I deal with in my research work. Indeed, one of the main issues in Transactional Memory (TM) systems (both centralized and distributed) is the performance degradation that can be experienced when a sub- optimal system configuration is set. I remark that, unlike non-transactional systems, in TM- based applications performance is affected, in addition to contention on hardware resources (as CPU, system bus), also by contention on shared data. This increases the complexity of predicting the system's performance, because the various affecting factors are tightly intertwined and show complex inter-dependencies. The complexity is particularly exacerbated in the case of DTM, due to the presence of factors such as data distribution/replication and shared network links. As a consequence, developing performance analysis tools for a DTM system entails addressing a number of different aspects. 

Report: here