国外操作系统相关论文《Efficient SystemEnforced Deterministic Parallelism》.pdfVIP

Efficient System-Enforced Deterministic Parallelism Amittai Aviram, Shu-Chun Weng, Sen Hu, Bryan Ford Yale University Abstract manifestation depends subtly on inputs and code path lengths instead of on “real” time. Third, the user-level Deterministic execution offers many benefits for debug- instrumentation required to isolate and schedule threads’ ging, fault tolerance, and security. Current methods memory accesses can incur considerable overhead, even of executing parallel programs deterministically, how- on coarse-grained code that synchronizes rarely. ever, often incur high costs, allow misbehaved software to defeat repeatability, and transform time-dependent To meet the software development, debugging, and races into input- or path-dependent races without elim- security challenges that ubiquitous parallelism presents, inating them. We introduce a new parallel program- it may be insufficient to shoehorn the standard nonde- ming model addressing these issues, and use Determina- terministic programming model into a synthetic execu- tor, a proof-of-concept OS, to demonstrate the model’s tion schedule. Instead we propose to rethink the basic practicality. Determinator’s microkernel API provides model itself. We would like a parallel environment that: only “shared-nothing” address spaces and determinis- (a) is “deterministic by default” [12, 40], except when tic interprocess communication primitives to make ex- we inject nondeterminism explicitly via external inputs; ecution of all unprivileged code—well-behaved or not— (b) introduces no data races, either at the memory ac- prec