The Department of Computer Science at the University of Cyprus cordially invites you to the PhD Defense entitled:

The TFlux Parallel Processing Platform

Speaker: Mr. Kyriakos Stavrou
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: Friday, January 24th, 2009
Time: 11:00 - 12:00 EET
Host: Pedro Trancoso (pedro AT

This work presents the TFlux (Thread Flux) Parallel Processing Platform, a complete system that offers an efficient dataflow-like thread-based model of execution, the Data-Driven Multithreading (DDM), to its users using commodity components, i.e. unmodified Operating System, unmodified compiler and unmodified ISA hardware making it applicable to off-the-shelf systems. TFlux provides a complete solution from the programming toolchain to the hardware implementation. The abstraction layer TFlux exports to its users hides all the details of the underlying machine allowing different hardware configurations to support its model of execution transparently to the programmer. The user of TFlux can develop applications using a set of simple but powerful compiler directives. Then the TFlux-C-Preprocessor coverts this code to an ANSI C program that includes the Runtime Support for TFlux and all calls to the system’s scheduler. This code can be compiled with a commodity C compiler resulting in a binary that is executable by any commodity Operating Systems on any commodity CPU processors. The layered design of TFlux has been tested on different Unix-based multiprocessor systems. Moreover, this design enabled the porting of TFlux to different machines with minimum effort. In this work, two TFlux implementations are presented: TFluxHard and TFluxSoft. For TFluxHard the thread Scheduler is implemented as a hardware unit whereas for TFluxSoft, the Scheduler’s functionality is provided at the software level. As such, TFluxHard is applicable to systems that offer the ability to augment the machine with a small hardware module while TFluxSoft is directly applicable to any existing off-the-shelf systems. To evaluate the TFlux implementations, a benchmark suite based on existing and synthetic applications was developed. The applications in this suite were carefully chosen in order to have different characteristics both in terms of their dynamic behavior and complexity of their dataflow graph. Although for most applications the performance of the two implementations is close, TFluxHard shows an advantage over TFluxSoft arising from offloading the Scheduler’s functionality to the hardware module. In addition, the experimental results also show that TFlux is able to exploit better the parallelism for applications with complex dependency graphs, compared with traditional parallel programming model approaches. Overall, TFlux proves to be a platform that allows applications to efficiently exploit the available parallelism. This is achieved by supporting dataflow-like thread scheduling in commodity multicore/multiprocessor systems.

Short Bio:
Kyriakos Stavrou is a PhD Candidate at the Department of Computer Science, University of Cyprus.

  Other Presentations Web:
  Colloquia Web: