Introduction to high-performance computing.

                         Thomas Ludwig


ABSTRACT

The supercomputer parallel calculations find an ever increasing number
of applications to solving numerous topical problems in the modern
science and technology. Their use is governed by emergence of a new
class of superlarge problems. The course of lections will include the
history of development of supercomputing, review of the most powerful
supercomputers, description of the classes of problems requiring
parallel computations, and the main trends in development of
technologies. Architectures of the supercomputers with shared and
distributed memories will be described as well as distinctions between
supercomputers and parallel clusters. Specific features of data
storage during parallel computations will be considered as well as the
technologies of parallel programming.


Lecture 1 - Architecture of High Performance Computers

To start with, we will discuss the architectural principles of high
performance computers and, in particular, of compute clusters. We will
have a closer look to processors, interconnect technology, storage,
and in particular, to the memory architecture. The latter defines the
classes of shared and distributed memory computers. The lecture will
also present some data from the current TOP500 list of the strongest
computers in the world. Finally, an overview over operating system
aspects will be presented.


Lecture 2 - Parallel Programming Principles

We will now learn how parallel programs are characterized and how in
principle we design and implement such programs. A good knowledge of
compiler and hardware details is often necessary in order to get
optimal performance of the program. The parallelization paradigm of
data partitioning and message passing will be introduced. Two measures
will be presented to evaluate the performance of the parallel program.


Lecture 3 - Message Passing with MPI

The first step into parallel programming will be done based on the
Message Passing Interface (MPI). We will write a small program that
distributes data to different compute nodes, calculates some data, and
finally collects the results. A few basic library calls for message
passing will be introduced, which are already sufficient to write a
first parallel program. Problematic issues like debugging and
performance analysis will be covered.


Lecture 4 - Advanced Issues with Message Passing

MPI offers a huge number of library calls, most of which do just
combine several basic calls and thus realize complicated activities in
a single call. We will have a look at collective calls and
sophisticated communication patterns. As bioinformatics is
particularly data intensive, a first introduction to parallel
input/output via MPI will be given. We will present an outlook onto
advances features in the MPI-2 standard and what they are used for.


KEYWORDS

Parallel Architectures, Parallel Programming, Message-Passing, MPI