Predictive analysis of message passing applications

Update Item Information
Title Predictive analysis of message passing applications
Publication Type dissertation
School or College College of Engineering
Department Computing
Author Sharma, Subodh
Date 2013-08
Description Message passing (MP) has gained a widespread adoption over the years, so much so, that even heterogeneous embedded multicore systems are running programs that are developed using message passing libraries. Such a phenomenon is a shift in computing practices, since, traditionally MP programs have been developed specifically for high performance computing. With growing importance and the complexity of MP programs in today's times, it becomes absolutely imperative to have formal tools and sound methodologies that can help reason about the correctness of the program. It has been demonstrated by many researchers in the area of concurrent program verification that a suitable strategy to verify programs which rely heavily on nondeterminism, is dynamic verification. Dynamic verification integrates the best features of testing and model checking. In the area of MP program verification, however, there have been only a handful of dynamic verifiers. These dynamic verifiers, despite their strengths, suffer from the explosion in execution scenarios. All existing dynamic verifiers, to our knowledge, exhaustively explore the nondeterministic choices in an MP program. It is apparent that an MP program with many nondeterministic constructs will quickly inundate such tools. This dissertation focuses on the problem of containing the exponential space of execution scenarios (or interleavings) while providing a soundness and completeness guarantee over safety properties of MP programs (specifically deadlocks). We present a predictive verification methodology and an associated framework, called MAAPED(Messaging Application Analysis with Predictive Error Discovery), that operates in polynomial time over MP programs to detect deadlocks among other safety property violations. In brief, we collect a single execution trace of an MP program and without re-running other execution schedules, reliably construct the artifacts necessary to predict any mishappening in an unexplored execution schedule with the aforementioned formal guarantee. The main contributions of the thesis are the following: The Functionally Irrelevant Barrier Algorithm to increase program productivity and ease in verification complexity. A sound pragmatic strategy to reduce the interleaving space of existing dynamic verifiers which is complete only for a certain class of MPI programs. A generalized matches-before ordering for MP programs. A predictive polynomial time verification framework as an alternate solution in the dynamic MP verification landscape. A soundness and completeness proof for the predictive framework's deadlock detection strategy for many formally characterized classes of MP programs. In the process of developing solutions that are mentioned above, we also collected important experiences relating to the development of dynamic verification schedulers. We present those experiences as a minor contribution of this thesis.
Type Text
Publisher University of Utah
Subject Dynamic verification; MPI; Predictive analysis
Dissertation Institution University of Utah
Dissertation Name Doctor of Philosophy
Language eng
Rights Management Copyright © Subodh Sharma 2013
Format application/pdf
Format Medium application/pdf
Format Extent 809,589 bytes
Identifier etd3/id/2593
ARK ark:/87278/s6mk9n24
Setname ir_etd
ID 196169
Reference URL https://collections.lib.utah.edu/ark:/87278/s6mk9n24
Back to Search Results