Performance Visualization of MPI-IO Operations in Parallel Programs

                              Thomas Ludwig


ABSTRACT

With scientific computations, e.g. in the field of bioinformatics, we
observe a dramatic increase in the amount of data that is
computed. Even with modern high performance computers the storage
capabilities often form the bottleneck in computing more detailed
results. We focus our research on parallel file systems that can be
found in cluster environments.

The critical issue here is the overall performance that you will get
from your system. Scalability is a problem because of sequential parts
in the code of the parallel file system when it comes to metadata
operations. What is needed are tools that give insight into the
internal behavior of parallel file systems and relate this information
to the user level. By doing so we can see what activity in the user
program triggers which low level read/write operations.

The talk will present an enhanced tool environment that is based on
PVFS2 and MPICH2. We add tracing facilities to the parallel file
system and thus can investigate its behavior and relate it to the
parallel user program.

In particular we will show, how different concepts of MPI-IO
operations relate to I/O performance. We will compare non-collective
and collective calls and requests for contiguous and non-contiguous
data regions. Conclusions wil be drawn for application programers.


KEYWORDS

Performance Visualization, Performance Evaluation, Parallel I/O, PVFS2