Binary analysis for measurement and attribution of program performance

Abstract

Modern programs frequently employ sophisticated modular designs. As a result, performance problems cannot be identified from costs attributed to routines in isolation; understanding code performance requires information about a routine's calling context. Existing performance tools fall short in this respect. Prior strategies for attributing context-sensitive performance at the source level either compromise measurement accuracy, remain too close to the binary, or require custom compilers. To understand the performance of fully optimized modular code, we developed two novel binary analysis techniques: 1) on-the-fly analysis of optimized machine code to enable minimally intrusive and accurate attribution of costs to dynamic calling contexts; and 2) post-mortem analysis of optimized machine code and its debugging sections to recover its program structure and reconstruct a mapping back to its source code. By combining the recovered static program structure with dynamic calling context information, we can accurately attribute performance metrics to calling contexts, procedures, loops, and inlined instances of procedures. We demonstrate that the fusion of this information provides unique insight into the performance of complex modular codes. This work is implemented in the HPC-TOOLKIT 1 performance tools.

Identifying Program Loop Nesting Structures during Execution of Machine Code

“…Several tools that analyze serial code and provide useful information for performance tuning and parallelization have been introduced in the recent past [6]- [8], [12]. However, their loop detection mechanisms are neither precise nor inter-procedural.…”

“…However, there might be situations in which loop structures in source code do not correspond exactly to these in binary code due to aggressive loop transformations and function inlining. Even in those cases, we can map the structures without inconsistency of semantics as discussed in [8]. Also, we note that the loop structures in the machine code is one of the important factors in determining the data access locality and performance.…”

“…The HPCToolkit presented by Tallent et al [8] performs binary analysis that recovers its static structure of loop nests using Havlak's algorithm and maps it into the debug information. While theirs aim at tuning of programs performed by post-mortem, two-pass analysis, the proposed pMarker method is based on one-pass analysis and can obtain dynamic information on-the-fly.…”

“…There are three significant advantages of our loop nest detection system over the previous methods [6]- [8]. The first is that it enables precise detection of inter-procedural dynamic loop nests.…”

Visualizing complex dynamics in many-core accelerator architectures

“…Instead of relying on compiler-generated debug information to map an instruction to its associated source code line (what AerialVision currently does in source code view), HPCToolkit analyzes the application binaries directly to recover the program structure [40]. It then correlates the statistically sampled performance metrics with the source code structure and presents the metrics with the associated source code.…”