NYCCS/Computer Science Seminar
Tulin Kaman, SBU Friday, June 5, 2009 9:00AM-1:00PM
9:00AM Breakfast
9:30-10:00AM: Tuning and Analysis Utilities: TAU
The TAU toolkit for the performance analysis and tuning of distributed
and multithreaded programs can be used by New York Blue users who want
to identify and eliminate the bottlenecks. TAU performance evaluation
tool is available for codes which run on NYBlue L and P platforms.
10:00-1030AM: Debugging on the Blue Gene/P System
Debugging parallel applications on Blue Gene can be achieved using the
GNU debugger. The GNU debugger gdb that comes with the Blue Gene systems
has the ability to do remote debugging via a GDB server(gdbserver) on the
remote machine. Starting gdbserver through the mpirun command allows the
programmer to get the debug server running first. The GDB server runs on
I/O nodes, and each GDB client debug one Compute Node. GDB connections
can be established via TCP/IP.
10:30-11:00AM Break
11:00-11:30AM: Intel Multi-core Architecture
Multi-core architectures that can hold dozens or even hundreds of
processors on a compute node have become a worldwide effort for the future
of computing technology. The goal is to improve performance, increase
energy-efficiency. Intel's microarchitecture, code named Nehalem, represents
the latest in processor innovation. The new generation of multi-core processor
architectures provide new levels of energy efficiency and performance
scalability.
11:30AM-12:00PM: Hybrid-Multi/Core Programming
Computers of today and the future will have many cores per processor.
Efficient parallelism will require a mixed used of threads across
cores on one processor and message passing between distinct nodes. Here
we address a novel strategy, which appears to be practical for at least a
large class of scientific problems.
Hybrid (threads + messages) programming should allow improved strong
scaling performance, and the messages are in play only for larger
segments of the memory. It is not reasonable to expect memory per thread
to keep pace with the increasing number of threads, so that failure to use
threads (using only MPI capabilities) may lead to degradation of performance
on new architectures.
12:00-1:00PM Lunch