| Designing and deploying application-specific integrated circuits (ASICs) is becoming a prohibitively expensive and risky endeavor. While there are many different possible silicon implementation alternatives that may replace ASICs, application-specific instruction processors (ASIPs) have the most promise because they have the potential to provide a flexible solution with high performance and shorter design times. Unfortunately, programming an ASIP today is already a time consuming error-prone task that will not scale to larger more complex architectures. Ideally, ASIPs should be programmed using a popular application environment---a productive design entry language that appeals to the application community. But the asymmetry between natural-to-use application environments and ASIP architectures prevents a direct mapping.; The goal of this research is to develop a methodology to bridge this gap by providing a more productive design flow while sacrificing little system performance. Our approach is to define a programming model that combines an application development environment and abstractions of salient features of the target architecture. We demonstrate the utility of a programming model in the networking domain, where the past five years have witnessed over 30 attempts at programmable solutions for packet processing. We create a programming model (called NP-Click) that aids mapping Click, a popular networking application environment, to the Intel IXP1200, a common network processor. Across three common networking applications, we compare the performance, design process and resource usage of using NP-Click to the current state-of-the-art programming approach. We show that the programming model approach is mostly able to remain within 10% of the performance of the current programming approach but reduces design time significantly. Moreover, NP-Click allows the designer to more effectively trade-off performance and design time. Lastly, this research describes how the programming model concept can be applied to additional network processors and ASIP architectures in other application domains. |
