As modern HPC applications generate massive amounts of data, there is an increasing demand for understanding supercomputer system with manycore processor such as Intel Knights Landing. I/O performance analysis and improvement should be made and finding suitable configurations for various environments with KNL processor is needed. In addition, as the cost of high performance storage devices such as NAND-flash becomes more affordable, the need of all-flash parallel file system arises. Since the computing power of manycore system increases significantly, the storage system needs to ensure low latency, high bandwidth I/O for HPC applications with all-flash storage devices.
Our research includes I/O performance analysis and optimization of storage system under manycore environment using KNL processor. We focus on performance improvement of HPC system using all-flash storage system. We analyze the SSD performance and I/O characteristics under various HPC environments. In order to implement a storage system considering the I/O characteristics of HPC applications, the optimization and proper design of I/O path of high-performance SSD is needed. Also we study on I/O scheduling method using the multi-stream feature to handle a wide variety of workloads and maintain high I/O bandwidth.
KNL Processor Architecture
Multi-streamed SSD Architecture