Man vs. Machine : Comparing Handwritten and Compiler-generated Application-Level Checkpointing
Year: 2004
Type: report
Abstract: The contributions of this paper are the following. We describe the implementation of the $C^3$ system for semi-automatic application-level checkpointing of C programs. The system has (i) a pre-compiler that instruments C programs so that they can save their states at program execution points specified by the user, and (ii) a novel memory allocator that manages the heap as a collection of pools. We describe two static analyses for reducing the overhead of saving and restoring the application state. The first one optimizes stack variables, while the second one optimizes heap data structures. To benchmark our system, we compare the overheads introduced by our semi-automatic approach with the overhead of handwritten application-level checkpointing in an n-body code written by Joshua Barnes. Except for very small problem sizes, these overheads are comparable. We highlight various algorithmic challenges in the optimization of application-level checkpointing that should provide grist for the mills of the PLDI community.
Source: eCommons (Cornell University)
Institution: Cornell University
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: National Science Foundation
Awards: Synthesis of Block-recursive Codes for Deep Memory Hierarchies, MATLAB Extensions and Compiler Techniques for High-Performance Computing, ITR: Adaptive Software for Field-driven Simulations, Next Generation Software: A Framework For Developing Complex Applications On High-End Petaflop-Class Machines, CISE Research Infrastructure: A Two-tier Computation and Visualization Facility for Multiscale Problems
OpenAlex ID: W43490847