Thursday, 5th of October 2017, 12:00 – 1:00

Scientific Workflows on Manycore Computers

Venue: 
SR 1, ICT Building,
Technikerstraße 21a, 6020 Innsbruck

Lecturer:

Matthias Janetschek
Research Assistant at DPS group, University of Innsbruck

Abstract:

Modern shared-memory heterogeneous manycore systems have becomeincreasingly complex to program, and traditional programming methods arestruggling to fully exploit the performance of today’s heterogeneoussystems.  One of the most successful paradigms for implementinghigh-performance computing applications in Grid- and Cloud-computing arescientific workflows, originally designed to easily create large andcomplex applications by reusing existing (often legacy and monolithic)software components and assembling them through well-defined controlflow and data-flow dependencies. Because of the similarity in terms ofscale and heterogeneity, workflow systems represent today a promisingalternative for development and execution of scientific applications onshared-memory heterogeneous manycore architectures. To test this thesiswe implemented a new Manycore Workflow Runtime Environment (MWRE) whichwe specifically designed for the unique requirements of shared-memoryheterogeneus manycore systems. MWRE is compiler-based and translatesworkflows from a XML-based workflow specification language into anequivalent C++-based program. Experimental results show that MWRE canachieve performance similar to OpenMP while better supportingheterogeneity and concepts like full-ahead scheduling.