AbstractsComputer Science

NAND flash has risen to become a popular storage medium in recent years due to its significant bandwidth, access latency and power advantages over traditional hard disks. However, in modern Big Data systems, simply replacing disks with flash does not yield proportional performance gains. This is because of bottlenecks in various levels of the system stack: I/O interface, network, file system and processor. Introduced in 2012, BlueDBM is a novel Big Data flash storage platform that includes a dedicated low latency sideband storage network between flash controllers, reconfigurable fabric for implementing in-store hardware accelerators and a flash-aware file system. While results are promising, the original BlueDBM platform used old flash boards that were merely 16GB in capacity running at 80MB/s. This work presents BlueFlash, a revamped and improved storage device for the BlueDBM platform. We present the design of a new 0.5TB flash board, a new ONFI-compliant FPGA-based flash controller with ECC and we explore the characteristics of the flash board. We demonstrate that BlueFlash scales well with multiple buses and multiple chips per bus, reaching 1.2GB/s (75% of theoretical max) read and 1.0GB/s (91% theoretical max) write bandwidth while consuming only 6W of power.