A Sampling Method Focusing on Practicality

Daniel Gracia-Perez (1) , Hugues Berry (1) , Olivier Temam (1)

1 ALCHEMY - Architectures, Languages and Compilers to Harness the End of Moore Years

In the past few years, several research works have demonstrated that sampling can drastically speed up architecture simulation, and several of these sampling techniques are already largely used. However, for a sampling technique to be both easily and properly used, i.e., plugged and reliably used into many simulators with little or no effort or knowledge from the user, it must fulfill a number of conditions: it should require no hardware-dependent modification of the functional or timing simulator, it should simultaneously consider warm-up and sampling, while still delivering high speed and accuracy.\\ \indent The motivation for this article is that, with the advent of generic and modular simulation frameworks like ASIM, SystemC, LSE, MicroLib or UniSim, there is a need for sampling techniques with the aforementioned properties, i.e., which are almost entirely \emph{transparent} to the user and simulator agnostic. In this article, we propose a sampling technique focused more on transparency than on speed and accuracy, though the technique delivers almost state-of-the-art performance. Our sampling technique is a hardware-independent and integrated approach to warm-up and sampling; it requires no modification of the functional simulator and solely relies on the performance simulator for warm-up. We make the following contributions: (1) a technique for splitting the execution trace into a potentially very large number of variable-size regions to capture program dynamic control flow, (2) a clustering method capable of efficiently coping with such a large number of regions, (3) a \emph{budget}-based method for jointly considering warm-up and sampling costs, presenting them as a single parameter to the user, and for distributing the number of simulated instructions between warm-up and sampling based on the region partitioning and clustering information.\newline \indent Overall, the method achieves an accuracy/time tradeoff that is close to the best reported results using clustering-based sampling (though usually with perfect or hardware-dependent warm-up), with an average CPI error of 1.68\% and an average number of simulated instructions of 288 million instructions over the Spec benchmarks. The technique/tool can be readily applied to a wide range of benchmarks, architectures and simulators, and will be used as a sampling option of the UniSim modular simulation framework.

Domaines: Architectures Matérielles [cs.AR]

Daniel Gracia-Perez, Hugues Berry, Olivier Temam. A Sampling Method Focusing on Practicality. IEEE Micro, 2006. ⟨inria-00158808v2⟩