TITLE: The Changing Face of Storage I/O in High Performance Computing
ABSTRACT: In broad terms, HPC systems consist of 3 major components:
processing, networking and storage. Significant attention is given to
the processing and networking components, yet storage is often
overlooked, but is
frequently the bottleneck for many applications. This problem has been
exacerbated in recent years by the advent of Linux clusters with
commodity
hardware and software which has made high performance computing (HPC)
affordable for a much wider market segment then in the past. But many of
the new players in this field do not have experience with these systems
and
adopt simple, yet inefficient storage I/O models. Beginning with a
historical overview, this presentation examines the range of
applications
and system architectures used in HPC today with a special focus on the
challenges of storage I/O, including both disk and tape. It examines a
taxonomy of disk storage solutions commonly used in today's HPC
environment
and gives examples of various storage system architectures that the
author
has designed for his customers.
BIO: Dr. Ray Paden is currently an HPC Technical Architect with world
wide scope
in IBM's Deep Computing organization, a position he has held since June,
2000. His particular areas of focus include HPC storage systems,
performance optimization and cluster design. Before joining IBM, Dr.
Paden
worked as software engineer doing systems programming and performance
optimization for 6 years in the oil industry. He also served in the
Computer Science Department at Andrews University for 13 years,
including 4
years as department chair. He has a Ph.D. from the Illinois Institute
of
Technology in Computer Science. He has done research and published
papers
in the areas of parallel algorithms and combinatorial optimization,
performance tuning, file systems, and computer education. He has served
in
various capacities on the planning committee for the Supercomputing
conference since 2000.
|
