UC San Diego

The IBM Cell processor is a heterogeneous multi-core architecture designed to demonstrate exceptional levels of performance improvement for compute-intensive applications. The streamlined design of its Synergistic Processing Units (i.e. small local store, no cache, limited branch prediction, no dynamic instruction reordering) presents a new set of challenges for application developers as they are now required to explicitly control the flow of data amongst the processors. Therefore, in this thesis we introduce a lightweight, flexible framework library called Cell Architecture Framework and Extensions (CAFE) to assist developers in taking advantage of this computational power without forcing a single programming model onto their applications. CAFE takes a more minimalistic approach than other frameworks by presenting low-level abstractions and utilities designed to help partition and transfer data between the cores. As a result, programmers can develop applications using a reasonably high level interface, yet still retain explicit control over the flow of data. We believe that these characteristics are very important to developing high-performance applications in the Cell environment. To motivate our design decisions, we provide an in-depth examination of where these aforementioned implementation challenges appear and discuss a range of countermeasures that may be successfully employed. We also present an assortment of example applications that each utilize CAFE in a different manner to help show its versatility