Author: Roger Kowalewski
Title: Nasty-MPI: A Systematic Stress Testing Approach for Correctness Debugging in MPI-3 RMA
Abstract:
The Message Passing Interface (MPI) as the de-facto standard for
developing High Performance Computing (HPC) applications provides two
communication models. The traditional one is two-sided communication,
where MPI processes share data with point-to-point messages or
collective operations and are implicitly synchronized by message
matching. In contrast, the one-sided model defines an interface for
Remote Memory Access (RMA), allowing one process to specify all
communication parameters for both the sending and receiving side.
One-sided communication semantically decouples data movement and
synchronization by providing support for asynchronous reads and
updates of distributed shared memory. While such an interface can be
extremely efficient, it poses several challenges. Improper
synchronization of possibly conflicting asynchronous memory accesses
to shared data may result in errors, such as data races or ordering
violations. Due to the ever increasing scale of applications and
machines, such synchronization errors are hard to debug and cannot
always be easily reproduced. This master thesis presents Nasty-MPI, a
new tool that supports developers in verifying the correctness of
their MPI-3 RMA programs. The tool can be attached to any MPI program
without code modification. It dynamically intercepts and transforms
RMA operations into different executions which are still valid in
terms of the MPI standard. Given an application with a latent
synchronization error, Nasty-MPI will force a manifestation of the
error that can easily be detected with the help of program invariants
(e.g., assert statements). An experimental evaluation shows that the
tool can detect synchronization errors in even simple configurations
which would otherwise only shown in much more complex scenarios.