A Virtual Platform for Performance Estimation of Many-core Implementations

This paper presents a prototype for a virtual platform to estimate performance of OpenMP parallelized programs in shared-memory many-core platforms at early stages of the design flow. This is a challenging problem because, at these stages, the particular details of the final platform are unknown, but early performance estimations are needed to choose between different parallel implementations. The tool presented enables fast modelling of the SW and HW components in a complete platform model and also has the advantage of enabling configurable models of any many-core platform. This can be achieved because of two novel ideas that are explained in this paper: a native simulation framework that enables the modelling of concurrent threads described in OpenMP and a novel use of "shared" and "private" clauses that models the data transfers. The advantages of using the proposed tool are explained with a specific example.