Design and resource management of reconfigurable multiprocessors for data-parallel applications

FPGA (Field-Programmable Gate Array)-based custom reconfigurable computing machines have established themselves as low-cost and low-risk alternatives to ASIC (Application-Specific Integrated Circuit) implementations and general-purpose microprocessors in accelerating a wide range of computation-intensive applications. Most often they are Application Specific Programmable Circuiits (ASPCs), which are developer programmable instead of user programmable. The major disadvantages of ASPCs are minimal programmability, and significant time and energy overheads caused by required hardware reconfiguration when the problem size outnumbers the available reconfigurable resources; these problems are expected to become more serious with increases in the FPGA chip size. On the other hand, dominant high-performance computing systems, such as PC clusters and SMPs (Symmetric Multiprocessors), suffer from high communication latencies and/or scalability problems. This research introduces low-cost, user-programmable and reconfigurable MultiProcessor-on-a-Programmable-Chip (MPoPC) systems for high-performance, low-cost computing. It also proposes a relevant resource management framework that deals with performance, power consumption and energy issues. These semi-customized systems reduce significantly runtime device reconfiguration by employing userprogrammable processing elements that are reusable for different tasks in large, complex applications. For the sake of illustration, two different types of MPoPCs with hardware FPUs (floating-point units) are designed and implemented for credible performance evaluation and modeling: the coarse-grain MIMD (Multiple-Instruction, Multiple-Data) CG-MPoPC machine based on a processor IP (Intellectual Property) core and the mixed-mode (MIMD, SIMD or M-SIMD) variant-grain HERA (HEterogeneous Reconfigurable Architecture) machine. In addition to alleviating the above difficulties, MPoPCs can offer several performance and energy advantages to our data-parallel applications when compared to ASPCs; they are simpler and more scalable, and have less verification time and cost. Various common computation-intensive benchmark algorithms, such as matrix-matrix multiplication (MMM) and LU factorization, are studied and their parallel solutions are shown for the two MPoPCs. The performance is evaluated with large sparse real-world matrices primarily from power engineering. We expect even further performance gains on MPoPCs in the near future by employing ever improving FPGAs. The innovative nature of this work has the potential to guide research in this arising field of high-performance, low-cost reconfigurable computing. The largest advantage of reconfigurable logic lies in its large degree of hardware customization and reconfiguration which allows reusing the resources to match the computation and communication needs of applications. Therefore, a major effort in the presented design methodology for mixed-mode MPoPCs, like HERA, is devoted to effective resource management. A two-phase approach is applied. A mixed-mode weighted Task Flow Graph (w-TFG) is first constructed for any given application, where tasks are classified according to their most appropriate computing mode (e.g., SIMD or MIMD). At compile time, an architecture is customized and synthesized for the TFG using an Integer Linear Programming (ILP) formulation and a parameterized hardware component library. Various run-time scheduling schemes with different performanceenergy objectives are proposed. A system-level energy model for HERA, which is based on low-level implementation data and run-time statistics, is proposed to guide performance-energy trade-off decisions. A parallel power flow analysis technique based on Newton\u27s method is proposed and employed to verify the methodology