Multi-valued Expression Analysis for Collective Checking

Abstract

Determining if a parallel program behaves as expected on any execution is challenging due to non-deterministic executions. Static analyses help to detect all execution paths that can be executed concurrently by identifying multi-valued expressions, i.e. expressions evaluated differently among processes. This can be used to find collective errors in parallel programs. In this paper, we propose a new method that combines a control-flow analysis with a multi-valued expressions detection to find such errors. We implemented our method in the PARCOACH framework and successfully analyzed parallel applications using MPI, OpenMP, UPC and CUDA.

Rethinking Data Race Detection in MPI-RMA Programs

“…Our software stack is built with LLVM-15 and we used an Eviden OpenMPI implementation of MPI, built in its 4.1.5.2 version. For the experiments, we use RMA-Analyzer integrated in PARCOACH [7], a tool dedicated to the detection of deadlocks caused by an improper use of collective calls in parallel applications. The performance analysis is done with MUST-RMA v1.9.0, available on github 1 .…”

“…In this example, the node corresponding to the MPI_Put, ([2...12], 𝑅𝑀𝐴_𝑅𝑒𝑎𝑑) is inserted in the left subtree of the node ([4], 𝐿𝑜𝑐𝑎𝑙_𝑅𝑒𝑎𝑑). Then, when searching for the intersecting accesses with new access ([7], 𝐿𝑜𝑐𝑎𝑙_𝑊 𝑟𝑖𝑡𝑒), a new access is inserted in the BST. The algorithm works as follow: when a new access 𝑛𝑒𝑤_𝑎𝑐𝑐 is inserted in the BST, if this 𝑛𝑒𝑤_𝑎𝑐𝑐 is intersecting with another access 𝐵𝑆𝑇 _𝑎𝑐𝑐 already present in the BST, three new accesses representing the parts of the intersection of the two…”

PARCOACH Extension for Static MPI Nonblocking and Persistent Communication Validation

“…This trade-off allows the best of both approches. PARCOACH [2], [3] uses this method to find collectives mismatch. It raises warnings for potential errors with debugging information like the conditionals responsible for them.…”

“…After the matching is done, a basic data-flow analysis is performed to find statements that can lead to race conditions inside each overlapping window. Collective mismatch detection is performed by PARCOACH using the algorithm presented in [1] and extended in [3].…”

“…PARCOACH [1]- [3] is a framework built on top of LLVM to detect misuse of collectives in MPI programs. A first analysis studies the program at compile-time and issues warnings when a potential error is detected.…”