Distributed Clustering and Scheduling of Object-Oriented Virtual Machines

This report presents an overview of several approaches to provide a Single System Image view of a cluster, particularly concerning the view of a single address space. The main focus of our work is to understand the current approaches for clustering a regular multithreaded and non-cluster-aware Java application, as well as the current techniques and metrics for scheduling threads in a heterogeneous cluster. We will start by addressing the Distributed Shared Memory (DSM) approach, together with the consistency models studied in the academic world to improve performance and reduce the communication. Software Transactional Memory (STM) will also be addressed as a possible alternative to lock-based approach for providing data synchronization between nodes. In section 3.2 we will take a closer look to current Java approaches for clustering and their main features (e.g. global GC, support for JIT, native code and load-balancing mechanisms). Algorithms for scheduling threads and thread migration will also be considered in 3.3, complemented by a small reference to live migration of virtual machines. Finally, we propose a new extension to provide Terracotta with a global scheduler in an attempt to make it a viable platform for general-purpose multithreaded applications without having the programmer to worry about