Date: Monday, May 6, 2024

Time: 2:00 pm - 4:00 pm ET

Location: Klaus 3402 and Virtual (https://gatech.zoom.us/j/3092779704?pwd=SDVZaTlkTFpsZDRoYUU3Y0pzdmNvdz09)

Lechen Yu

Ph.D. Student

School of Computer Science

Georgia Institute of Technology

Committee:

Dr. Vivek Sarkar (Advisor) – School of Computer Science, Georgia Institute of Technology

Dr. Santosh Pande – School of Computer Science, Georgia Institute of Technology

Dr. Qirun Zhang – School of Computer Science, Georgia Institute of Technology

Dr. Hyesoon Kim – School of Computer Science, Georgia Institute of Technology

Dr. Tiago Cogumbreiro – Department of Computer Science, University of Massachusetts Boston

Abstract:

OpenMP is a popular intra-node parallel programming model that supports several

problem decomposition approaches, including task parallelism, data parallelism, and heterogeneous 

parallelism. When OpenMP introduces new parallel paradigms or features, it must

ensure that these additions align with the existing constructs to maintain consistency and

avoid unspecified behaviors. New features can result in revisions to the behavior of existing

constructs, which may in turn require programmers to re-evaluate their understanding of

existing constructs and adapt their code accordingly. Consequently, writing correct OpenMP

programs can be challenging even for experienced programmers.

To alleviate the intricacy of writing correct OpenMP programs, this proposal outlines

several dynamic analysis techniques that help programmers identify programming errors in

OpenMP programs. First, we describe various studies on device offloading, a recent OpenMP

feature still in its developmental phase. Our studies reveal discrepancies between the LLVM

implementation and the intended runtime behavior of device offloading. Additionally,

erroneous usage of device offloading constructs can lead to an assortment of memory

anomalies. Since these memory anomalies can generate disparities between host variables

and their corresponding counterparts on accelerator devices, we classify such bugs as data

inconsistencies. By establishing permissible memory accesses on the host and accelerator,

this proposal introduces a dynamic approach to detect data inconsistencies automatically.

The dynamic approach leverages a per-variable state transition model, which can be used to

establish the validity of the memory location before executing any memory accesses. Beyond

data inconsistencies, this proposal also delves into novel dynamic approaches for identifying

data races in OpenMP programs. By extending the SPD3 race detection algorithm, originally

designed for async-finish task parallelism, our dynamic race detection approach can handle a

large subset of parallel constructs in OpenMP, thereby checking more precise happens-before

relations among tasks relative to existing per-thread vector-clock-based approaches.