Leveraging Binary Translation for Heterogeneous Profiling
Department of Computer Science, University of Virginia
4th Workshop on Architectural and Microarchitectural Support for Binary Translation, Held in conjunction with the 38th Int’l Symposium on Computer Architecture (ISCA-38)
@article{upton2011leveraging,
title={Leveraging Binary Translation for Heterogeneous Profiling},
author={Upton, D. and Hazelwood, K.},
year={2011}
}
Heterogeneous systems, such as those including a graphics processor for general computation, are becoming increasingly common. While this increases the potential computing power that can be leveraged, it also increases the complexity of the system. This in turn increases the complexity of understanding behavior of the system, which is important when developing new software as well as when designing a new heterogeneous platform. Most current approaches profile individual computational resources in isolation, making it difficult to understand the characteristics of the full system. In this paper, we present an approach based on binary instrumentation that simplifies the task of profiling the behavior of heterogeneous systems by enabling a user to easily combine profiling data from multiple resources. We accomplish this by using communication and computation functions as markers to align the profiles. By using a lightweight method, we can achieve an average overhead of 35% overall and only 1% overhead on the compute kernels.
October 30, 2011 by hgpu
