| OCR Text |
Show 4 Simon is intended for the simulation of medium to large parallel systems. An inefficient processor model could seriously limit the usefulness of Simon for simulating even moderately large systems. Consider a system of 64 processors (certainly a modest system compared to current product announcements.) A "typical" register transfer simulation model with a simulation overhead of 1000:1 per target processor could require 18 hours of host CPU time to simulate a single second of the system, exclusive of additional simulation overhead for message passing and synchronization. 1.4 Processor Simulation Models: An Overview A wide variety of different techniques have been used to simulate computer processors. These techniques may be classified into five broad categories : 1. Probabilistic models. 2. Instruction mix models and synthetic benchmarks. 3. Trace driven models. 4. Register transfer models. 5. Hardware and firmware emulators. The ordering in the list is roughly from highly simplified models to highly detailed and accurate models. There is also an ordering from (1) to (4) of reduced relative efficiency. Probabilistic models are characterized by simulation overheads that are often better than 1:1, while register transfer models suffer from overheads that are two to four orders of magnitude larger. Emulators are often much more efficient than register transfer models. This is not surprising, since emulation can be viewed as a register transfer model where significant portions of software have migrated to firmware and/or hardware. In discussing the models certain characteristics are presented in either a positive or negative light. It should be noted that these judgements are relative to the goals, requirements, and intended uses of the Simon System, and should not be taken as absolute measures of "good" or "bad." It is possible to |
