ASIC-系统软件与软件安全研究组.ppt

* * * * * OECD = Organization of Economic and Cooperation and Development. In 1998, US only connect small number of countries for research collaborations, such as UK, Canada, France, Germany, and the rest of countries are insignificant or in the long tail. The data in right hand side shows the collaboration pattern with many countries in the world. The two figures show the change from power law to stretched exponential. * * * * * * * * * * * * * * * * * * Since Spark is designed around RDD, Can we simply offload RDD operations to GPUs to get better performance? The answer is no due to the mismatches between Spark’s designs and GPU’s unique characteristics * * * A case for the highly efficient specialization Crypto currency mining is high computing intensive task in bitcoin system Since 2008, the mining has been evolved on CPU, GPU, FPGA, and ASIC The efficiency factors are CPU=1, multicore = 10, GPU=100, FPGA=1,000 ASIC is the most efficient in throughput and power: 10,000 to 1,000,000 * High Performance Computing is Scale-up based Scaling up Maximize parallelism as the number of computing nodes and workload scale, and minimize latency High scalability by advanced hardware Achieving high parallelism Exploiting locality to reduce latency High performance is gained by HPC scalability = parallelism / latency Implementation Upgrading hardware and software optimization * Same Efforts to achieve high Scalability in Big Data systems Scaling out Raising throughput as the number of computing nodes and workload scale, subject to an acceptable latency High scalability is achieved by Maximizing parallelism to raise throughput Exploiting locality to reduce latency * Computing environment is not Inclusive General-purpose environment OS and application software are designed for CPU Other devices are external to the ecosystem Specialized devices are self-functioning Data is segregated from the main system Major hardware producers become catalog vend