Optimization Of The Phase Change Memory As Main Memory Of Stream Processor

The development of single-Chip Multiprocessor and new application's need for more accuracy and better perforamance require main memories of computer systems increase its capacity to store more worksets. DRAM has been the primary technology to build main memories of computer systems. However with feature size scaling down DRAM is in jeopardy as charge storge and sensing mechanisms become less reliable. The capacity increament of DRAM main memory result in a significant portion of the total system power and the total system cost is spent in the memory system. As integrated circuit technology to and beyond the end of CMOS dimensional scaling, CPU needs new memory technology to replace SRAM. Therefore, technology researchers have been studying new memory technologies that can provide more memory capacity than DRAM while still being competitive in terms of performance, cost, and power. PCRAM and STT-RAM are those new memory technologies. PCRAM has advantages in scalability and dencity compared with DRAM. However, PCRAM have longer read and write latency than DRAM, PCRAM's read and write operations need consume more energy compared with DRAM, as a non-valitile memory, PCRAM has limit write endurance also. Those drawbacks of PCRAM will be less important with feature size scaling. At current technology node we need to improve PCRAM's properties to make it compacatiable with DRAM.This dissertation focuses on the optimation of PCRAM main memory for stream processor. The major research contents and achievements cover the following aspects. (1)Bases on the study of PCRAM storage mechanisms, we model PCRAM in a fundamental way from Architecture point, use this model we build main memory for stream processor. (2) In order to enhance limited PCRAM write endurance, we use redundant bit-writes removal technology extended the endurance of PCRAM to an average 3.2x longer, and reduced write energy to 42% of original PCRAM. We study the way to use nonvolatility to improve memory access performance and make applications go faster with PCRAM main memory. The use of nonvolatility to avoid not necessary row buffer write back can improve applications'performance and consume less energy. (3) Base on the analysis result of stream memory system and PCRAM technology, we test several memory access scheduling algorithms on applications. Experiment result show that the row/open memory access scheduling algorithm is more suitable for PCRAM than other algorithms.