OS/Networking Seminar

This talk gives a brief overview of MPI and compares and contrasts MPI to other message passing systems. Specific features of LAM/MPI will also be discussed, including LAM's dynamic process control capabilities, and its utilization of an underlying daemon-based run-time environment. Finally, a potential collaboration to integrate LAM into Condor's run-time environment will be introduced. Topics include the possibility of dynamic MPI processes under Condor as well as checkpointing, migrating, and resuming Condor LAM/MPI jobs.

For over three decades, computers have made a continuous, exponential improvement in performance, as predicted by Moore's Law. Feature size in each new generation of circuits is smaller than the last; yielding an ever- increasing number of components on a single die and continually rising clock frequencies. In recent years, the architectures exploiting these resources have begun to change, requiring more sophistication in the compilers and tools that map applications onto hardware. In particular, designs increas- ingly rely on software to expose parallelism and keep processor utilization high. During this talk, I will discuss the state of parallel software tools, industry efforts to aid the creation of parallel software, and oppor- tunities for future research and development in this area.

Paul M. Petersen is the Principal Engineer tasked with the research and development of parallel tools at KAI Software, an Intel company. Dr. Petersen graduated from the University of Illinois (UIUC) in 1992 with a PhD in Computer Science specializing in the analysis and generation of parallel programs. After graduation and a short postdoc, Dr. Petersen accepted a position at Kuck & Associates, Inc. At KAI Dr. Petersen was responsible for the early technical work which lead to OpenMP, and is currently closely involved with the technical committee's drafting the future revisions of the OpenMP specification. Dr. Petersen is also the architect of the Assure product line, whose goal is to detect errors in parallel programs. KAI cur- rently fields two versions of Assure; these are Assure for OpenMP, which detects errors on OpenMP programs, and Assure for Threads, which detects deadlocks and data-races in arbitrary POSIX Threads programs.

In April 2000, Intel acquired Kuck & Associates, Inc. as a wholly owned subsidiary. The company name was changed to KAI Software, an Intel company and the Champaign, IL site is currently undergoing expansion to serve as a parallelism center for Intel.

Due to the diverse storage needs of these two projects, no singular storage appliance would be ideal for either. NeST therefore is about the search for configurable, flexible storage. To insure portability and enable easy deployment, NeST is user level software that runs almost entirely without super-user privilege.

This talk will focus on both the design and implementation of the NeST software and will be motivated by a discussion of the various storage requirements that necessitate configurable, flexible storage solutions. Individual NeST software modules will be examined to show how this flexibility is achieved. Particular attention will be shown to the concurrency architecture, to the protocol layer and to the disk management.

Jim's talk covers some of the research he did last summer while at Lucent Bell Labs. The talk explores the metrics associated with administering fault-tolerant remote storage. Recovery precedence and replication width are described along with storage Quality of Service levels. The research identified new metrics that could be used by both vendors and customers to monitor and administer remote storage contracts.

1. They do not overlap I/O and computation, leading to wasted allocations.
2. They have no facility for scheduling traffic.
3. They employ end-to-end flow control, thus limiting throughput to that of the slowest link at any given moment.
4. They are very brittle with respect to common network conditions such as broken links and crashed servers.