The Department of Computer Science at the University of Cyprus cordially invites you to the PhD Defense entitled:
Data-Driven Multithreading Virtual Machine
Speaker: Samer Arandi
Affiliation: University of Cyprus, Cyprus
Category: PhD Defense
Location: Room 148, Faculty of Pure and Applied Sciences (FST-01), 1 University Avenue, 2109 Nicosia, Cyprus (directions)
Date: Thursday, November 17, 2011
Time: 10:00-11:00 EET
Host: Paraskevas Evripidou (skevos AT cs.ucy.ac.cy)
Since the advent of digital computers, chip designers built faster computers by relying on improvements in fabrication technologies and architectural/organizational optimizations. However, the inability of the sequential model to tolerate long latencies (manifested mainly in the Memory Wall) combined with the Power and Instruction Level Parallelism (ILP) Walls problems eventually rendered this approach ineffective. The envisaged solution was to switch to multi-core architectures. This switch was an engineering effort that did not address the long memory latencies and the complexity of the designs. In addition, this switch elevated concurrency as a major challenge as it became evident that new concurrent models/paradigms are needed to efficiently utilize the resources of multi-core chips. The Data-flow model is a formal model that can handle concurrency and tolerate memory and synchronization latencies. Data-Flow systems can also be simpler and so more power efficient than conventional systems. In this thesis, we propose re-visiting the Data-flow model and adopting it as the basis for an execution model that efficiently exploits the resources of multi-core architectures. We design and implement a virtual machine supporting the Data-Driven Multithreading (DDM) model of execution (which combines Dynamic Data-Flow concurrency with efficient sequential execution) on homogeneous and heterogeneous multi-core systems with a host/accelerator organization and a software-managed memory hierarchy. A special prefetching software cache -based on data-driven caching policies- is developed for the automatic management of such memory hierarchies. The developed virtual machine also supports distributed DDM execution on a cluster of multi-core nodes connected over an off-chip network. We evaluate the homogeneous and heterogeneous implementations of the VM for single multi-core nodes and clusters of nodes using a suite of benchmarks. The evaluation demonstrates that the VM scales well and tolerates latencies and synchronization overheads efficiently achieving very good performance and outperforming other state-of-the-art systems. The results indicate that the virtual machine efficiently exploits data-flow concurrency on conventional/commercial multi-core systems, which strengthens the case that hybrid models that combine Data-Flow concurrency with efficient control-flow execution are candidates for adoption as the basis of a new execution model for Multi core systems.
Samer Arandi is a PhD. Candidate at the Department of Computer Science, University of Cyprus. He received his BSc. in Computer Engineering from An-Najah University in Palestine in 2004 and his MSc. in Advance Computing with distinction from the Imperial College - UK in 2006. His research interests include Data-flow and Parallel Architectures and Multi-core Systems.
|Other Presentations Web: https://www.cs.ucy.ac.cy/colloquium/presentations.php|
|Colloquia Web: https://www.cs.ucy.ac.cy/colloquium/|