Exploring multiple levels of performance modeling for heterogeneous systems.

摘要

The current trend in High-Performance Computing (HPC) is to extract concurrency from clusters that include heterogeneous resources such as General Purpose Graphical Processing Units (GPGPUs) and Field Programmable Gate Array (FPGAs). Although these heterogeneous systems can provide substantial performance for massively parallel applications, much of the available computing resources are often under-utilized due to inefficient application mapping, load balancing, and tuning. While several performance prediction models exist to efficiently tune applications, they often require significant computing architecture knowledge for reliable prediction. In addition, they do not address multiple levels of design space abstraction and it is often difficult to choose a reliable prediction model for a given design. In this research, we develop a multi-level suite of performance prediction models for heterogeneous systems that primarily targets Synchronous Iterative Algorithms (SIAs). The modeling suite aims to produce accurate and straightforward application runtime prediction prior to the actual large-scale implementation. This suite addresses two levels of system abstraction: 1) low-level where partial knowledge of the application implementation is present along with the system specifications and 2) high-level where the implementation details are minimum and only high-level computing system specifications are given. The performance prediction modeling suite is developed using our proposed Synchronous Iterative GPGPU Execution (SIGE) model for GPGPU clusters, motivated by the RC Amenability Test for Scalable Systems (RATSS) model for FPGA clusters. The low-level abstraction for GPGPU clusters consists of a regression-based performance prediction framework that statistically abstracts system architecture characteristics, enabling performance prediction without detailed architecture knowledge. In this framework, the overall execution time of an application is predicted using regression models developed for host-device computations and network-level communications performed in the algorithm. We have used a family of Spiking Neural Network (SNN) models and an Anisotropic Diffusion Filter (ADF) algorithm as SIA case studies for verification of the regression-based framework and achieved over 90% prediction accuracy compared to the actual implementations for several GPGPU cluster configurations tested. The results establish the adequacy of the low-level abstraction model for advanced, fine-grained performance prediction and design space exploration (DSE). The high-level abstraction consists of the following two primary modeling approaches: qualitative modeling that uses existing subjective-analytical models for computation and communication; and quantitative modeling that predicts computation and communication performance by measuring hardware events associated with objective-analytical models using micro-benchmarks. The performance prediction provided by the high-level abstraction approaches, albeit coarse-grained, delivers useful insight into application performance on the chosen heterogeneous system. A blend of the two high-level modeling approaches, labeled as hybrid modeling, is explored for insightful preliminary performance prediction. The performance prediction models in the multi-level suite are verified and compared for their accuracy and ease-of-use, allowing developers to choose a model that best satisfies their design space abstraction. We also construct a roadmap that guides user from optimal Application-to-Accelerator (A2A) mapping to fine-grained performance prediction, thereby providing a hierarchical approach to optimal application porting on the target heterogeneous system. The end goal of this dissertation research is to offer the HPC community a thorough, non-architecture specific, performance prediction framework in the form of a hierarchical modeling suite that enables them to optimally utilize the heterogeneous resources.