Thomas M. Stricker
Graduate Student, Computer Science
Degrees held:
1985 Propaedeutical Deg. in Comp.Sci. & Comp. Eng., Swiss Federal Inst. of Technology (ETHZ), Zurich, Switzerland.
1988 Graduate Deg. in Comp.Sci. & Comp. Eng., Swiss Federal Inst. of Technology (ETHZ), Zurich, Switzerland.
1991 M.Sc., Carnegie Mellon University, Pittsburgh, USA
Entered program:
Fall 1989
Thesis Advisor
Prof. Thomas Gross
Research keywords:
Software and hardware architecture of communication systems in massively parallel computers. Applied parallel algorithms. Run-time support and OS for massively parallel computers.
Description of current research:
My major interest focuses on communication systems for supercomputers, in particular those for massively parallel machines. Powerful micro-processors have made it easy to concentrate tremendous computation power in a very small space, but at the same time they made it much more difficult to bring enough data as operands to the respective arithmetic units. Since a computation is distributed among hundreds of processors in such a supercomputer, extremely powerful networks are needed for transfers fast enough to keep up with the computation. Several studies pointed out that simply increasing the throughput and decreasing the latency of the network hardware is not sufficient to solve the problem of data transfers in supercomputers. Many types of overhead remain in handling the data appropriately before and after transfers through the network. In fast networks it becomes increasingly difficult to incorporate the transferred data seamlessly into a running computation. This overhead creates a critical performance bottleneck in most scientific application programs running on massively parallel computers. My research goal is to investigate different communication styles and determine their suitability for different parallel programming methods and their performance limits. The programming methods considered include a wide spectrum of paradigms, from manually coded systolic algorithms to automatically generated parallel FORTRAN programs, with an an emphasis on recent programming tools that automatically map programs to parallel systems. The specific communication method I am currently looking into is "message passing," where every node can transfer a message to any other node at any given time. This communication style has many advantages with respect to flexibility, but it is particularly difficult to implement when high communication performance is required. Based on my observations synchronization and data transfer semantics of message passing are best handled separately. For building research prototypes I am partially relying on the Carnegie Mellon's iWarp machines, but I am also working on porting my ideas to several other supercomputing platforms. I expect to conclude this investigation in 1995 and publish my results as a Ph.D. thesis. (last updated Jan 28, 1994)