Alexander Hück

In the context of MPI (Message Passing Interface), our talk presents a long-standing collaboration between RWTH Aachen and TU Darmstadt to further extend the MPI correctness checker tool MUST with memory allocation checking capabilities in the context of MPI (communication) calls for C/C++ target programs. To that end, we developed the LLVM compiler plugin TypeART. Both tools are open source and available under the BSD 3-clause license, see https://itc.rwth-aachen.de/must/ and https://github.com/tudasc/TypeART.

In general, MUST’s correctness checking includes errors that already manifest - segmentation faults or incorrect results - and many errors that are invisible to the developer or only manifest with certain HPC systems and MPI libraries. MUST works by intercepting MPI calls of a target application at runtime, allowing for bookkeeping of the current program state. Thus, MUST can cope with the complex MPI semantics of, e.g., (a) collectives, (b) wildcards, or (c) datatypes.

As a consequence of relying on intercepting MPI calls, though, MUST is unaware of the effective type of the allocated void* buffers used for the low-level MPI API. To that end, TypeART was developed to track memory (de-) allocation relevant to MPI communication. TypeART instruments heap, stack and global variable allocations with a callback to our runtime. The callback consists of (a) the memory address, (b) the type-layout information of the allocation (built-ins, user-defined structs etc.) and (c) number of elements. Thus, with TypeART, MUST can check for type compatibility between the type- less MPI communication buffer and the declared MPI datatype. Our tools also handle derived datatypes with complex underlying C/C++ data structures.

This talk gives an overview of the various open source HPC-related research projects at the Institute for Scientific Computing at TU Darmstadt. For each of the projects, we present an overview of the tool and its application from a user perspective. In addition, we cover more technical details, e.g., main programming language, build system or the general development stack. We have consolidated our efforts into a single GitHub space at https://github.com/tudasc.

MACH MACH (MPI Assertion Checking) is a Clang/LLVM-based static analysis tool to detect if one of the recently proposed MPI-assertions can be applied to an MPI communicator. A programmer may specify these assertions, such that the MPI Library can use more optimized operations, as some of the standard's constraints are relaxed in that case, e.g., strict message ordering. MACH statically checks whether the relaxation of some of the MPI standard's guarantees will cause errors in an application. The current version supports C and C++ applications, with the scope of the analysis being limited to one translation unit. The project is available under the Apache-2.0 License.

PIRA The Performance Instrumentation Refinement Automation framework is a profiler built on top of Clang/LLVM and Score-P. It performs automatic instrumentation refinement for performance measurements of sequential and MPI-parallel applications. To reduce runtime overhead it constructs, fully automatically, filter lists for Score-P based on static source-code features and profile information. The project is available under BSD 3-clause license.

SimAnMo The Simulated Annealing Modeler is a software tool to model the runtime behavior of programs depending on the size of one input parameter x. Its modeling procedure is based on runtime measurements for small values of x which result in acceptable runtimes and are so-called training points. SimAnMo supports modeling for programs with polynomial behavior, i.e., the program's runtime t(x) only grows in polynomial terms with the size of the input x. In addition, SimAnMo recognizes and models and exponential growth in the runtime. SimAnMo is highly configurable and parameterizable and automatically produces PDF reports of runtime behavior. The project is available under the MIT License.

TypeART TypeART is a memory type sanitizer based on the LLVM compiler framework. In particular, TypeART ist used for type correctness checks of MPI communication buffers and checkpointing libraries. In both cases, data is passed as a type-less void buffer, and the user is required to specify its type manually. This is error prone. To avoid a mismatch of types during runtime, TypeART instruments memory allocations (heap, stack, global) and extracts their type information to enable runtime comparisons. Detected mismatches are reported at runtime. The project is available under BSD 3-clause license.

Other people: Jan-Patrick Lehr Tim Jammer Michael Burger