Optimistic simulation of parallel architectures using program executables

  • Authors:

  • Sashikanth Chandrasekaran;Mark D. Hill

  • Affiliations:

  • Computer Sciences Department, University of Wisconsin-Madison, 1210 West Dayton Street, Madison, WI;Computer Sciences Department, University of Wisconsin-Madison, 1210 West Dayton Street, Madison, WI

  • Venue:
  • PADS '96 Proceedings of the tenth workshop on Parallel and distributed simulation
  • Year:
  • 1996

Quantified Score

Hi-index 0.00

Visualization

Abstract

A key tool of computer architects is computer simulation at the level of detail that can execute program executables. The time and memory requirements of such simulations can be enormous, especially when the machine under design-the target-is a parallel machine. Thus, it is attractive to use parallel simulation, as successfully demonstrated by the Wisconsin Wind Tunnel (WWT). WWT uses a conservative simulation algorithm and eschews network simulation to make lookahead adequate. Nevertheless, we find most of WWT's slowdown to be due to the synchronization overhead in the conservative simulation algorithm. This paper examines the use of optimistic algorithms to perform parallel simulations of parallel machines. We first show that we can make optimistic algorithms work correctly even with WWT's direct execution of program executables. We checkpoint processor registers (integer, floating-point, and condition codes) and use executable editing to log the value of memory words just before they are overwritten by stores. Second, we consider the performance of two optimistic algorithms. The first executes programs optimistically, but performs protocol events (e.g., sending messages) conservatively. The second executes everything optimistically and is similar to Time Warp with lazy message cancellation. Unfortunately, both approaches make parallel simulation performance worse for the default WWT assumptions. We conclude by speculating on the performance of optimistic simulation when simulating (1) target network details, and (2) on hosts with high message latencies and no synchronization hardware.