Compiler Support for Fine-Grain Software-Only Checkpointing

Abstract

We explore three application areas for our fine-grain checkpointing support. First, we utilize our efficient software-only checkpointing to support overlapping execution with delinquent loads through ii the prototyping and evaluation of both control-speculation and data-speculation transformations. We further propose a theoretical timing model and confirm its effectiveness with a real-machine workload.Second, we investigate its application to debugging, in particular by providing the ability for a debugger to rewind to an arbitrarily-placed point within the execution of a buggy program. A study using BugBench applications shows that our compiler-based fine-grain checkpointing has more than a factor of 100 less overhead than full-process, unoptimized checkpointing. Finally, we demonstrate that compilerbased checkpointing support can be leveraged to free the programmer from manually implementing and maintaining detailed software rollback mechanisms when coding a backtracking algorithm for a realistic CAD application, with runtime overhead of only 15% compared to the manual implementation.iii

Speculative Memory Checkpointing

“…User-level techniques can be easier to deploy, but incur significant run-time overhead because memory management at the application-level is more costly than from within the kernel [9]. Other user-level approaches, rely on compiler-based program instrumentation [10,14,39,40,64,65,76], which require source-code and recompilation of the target programs and all used libraries. Using dynamic instrumentation at the binary level [53,70] can provide checkpointing for unmodified binaries but incurs even higher performance overheads.…”

“…In debugging applications, it enables users to efficiently navigate through several program states observed during the execution, while empowering advanced debugging techniques such as reverse/replay debugging [27,34,60,61]. Memory checkpointing also serves as a key enabling technology for important first-class programming abstractions like software transactional memory [39], application-level backtracking [11,76], and periodic memory rejuvenation [68].…”

“…Traditional memory checkpointing techniques rely on commodity hardware-a strategy that provides superior deployability compared to instrumentation-based strategies [10,14,39,40,53,64,65,70,76]-to incrementally copy memory pages that were modified by the running program [11,20,21,34,37,51,54,56,58,60,63]. While incremental memory checkpointing is regarded as an efficient alternative to disk-based or full memory checkpointing [52], it still incurs nontrivial memory tracing costs for every taken checkpoint, resulting in relatively infrequent checkpoints used in practice.…”

“…Traditional incremental checkpointing mechanisms are usually page-granular, that is, a memory page is the smallest data block copied (although more fine-grained techniques exist [17,39,53,70,76]). Below we discuss the core mechanisms and techniques employed by these approaches.…”