Next-Generation Message Passing for InfiniBand Clusters

Colloq: Speaker: 
Matthew Koop
Colloq: Speaker Institution: 
Ohio State University
Colloq: Date and Time: 
Fri, 2009-01-30 10:00
Colloq: Location: 
5700, L202
Colloq: Host: 
Jeffrey Vetter
Colloq: Host Email:
Colloq: Abstract: 
The need for computational cycles continues to exceed availability, driving commodity clusters to increasing scale. With upcoming clusters containing hundreds-of-thousands of cores and more, InfiniBand is a popular interconnect for these clusters. Over 25% of the Top 500 supercomputers and 50% of the top 100 use the InfiniBand interconnect. As these systems scale, increasing attention must be given to achieving scalable performance and resource usage within the message passing library. In this talk, I will first focus on trends in large-scale systems as well as the scalability concerns that accompany these trends. Next, I will present new message passing designs that leverage the different network hardware features of InfiniBand to increase performance and scalability for upcoming large-scale systems. Finally, this talk will address the challenges in designing support for upcoming parallel architectures.
Colloq: Speaker Bio: