Falcon: Scaling IO Performance in Multi-SSD Volumes

With the high throughput offered by solid-state drives (SSDs), multi-SSD volumes have become an attractive storage solution for big data applications. Unfortunately, the IO stack in current operating systems imposes a number of volume-level limitations, such as per-volume based IO processing in the block layer, single flush thread per volume for buffer cache management, locks for parallel IOs on a file, all of which lower the performance that could otherwise be achieved on multi-SSD volumes. To address this problem, we propose a new design of per-drive IO processing that separates two key functionalities of IO batching and IO serving in the IO stack. Specifically, we design and develop Falcon1 that consists of two major components: Falcon IO Management Layer that batches the incoming IOs at the volume level, and Falcon Block Layer that parallelizes IO serving on the SSD level in a new block layer. Compared to the current practice, Falcon significantly speeds up direct random file read and write on an 8-SSD volume by 1.77× and 1.59× respectively, and also shows strong scalability across different numbers of drives and various storage controllers. In addition, Falcon improves the performance of a variety of applications by 1.69×.