ABSTRACT:
A very fast network is an essential part of any supercomputer. A quick look to the latest top 500 list shows that around half the clusters have other interconnect rather than Ethernet. IBM has its own proprietary network and an interface to access it. This library is called Low-level Application Programming Interface (LAPI) and consists in a set of message-passing functions based on a one-sided communication model. Additionally, LAPI provides an active message infrastructure that can accelerate communication between processes by only involving one of them in any transmission. The upcoming Blue Waters supercomputer will offer LAPI as one of the libraries to interact with the network.

In this talk, we will describe the challenges in designing and constructing a LAPI network layer for MPICH2. At the beginning, we will see how two-sided communication operations are implemented into MPICH2 library and how 4.5 us latency is achieved (compared to bare LAPI 4 us latency). Then, we will cover the opportunities LAPI opens up for improving one-sided communication operations and non-contiguous data transfers.

SHORT BIOGRAPHY:
Esteban Meneses received a B. Eng. degree in Computing Engineering (2001) and a M.Sc. degree in Computer Science (2007) from the Costa Rica Institute of Technology. Currently, he is a third year PhD student working with Prof. Laxmikant V. Kale at the Parallel Programming Laboratory in the University of Illinois at Urbana- Champaign. From 2007 to 2009, Esteban held a Fulbright-LASPAU scholarship for the first two years of his PhD program. His research interests span the areas of scalable fault tolerance and load balancing for supercomputing applications.