"In the past decade, all-flash storage systems have been replacing HDD-arrays to handle the performance requirements of modern enterprise workloads. However, deploying all-flash storage systems has encountered a couple of challenges. As an example, SSDs are much more expensive than HDDs, and with increasing demands for large capacities, the cost problem becomes more severe. As another example, recent enterprise applications (e.g., in the machine learning domain) have been demanding extreme performance which cannot be address by typical SSD-arrays. Replacing SSD-arrays with faster memories would increase the cost by orders of magnitude, and make it impractical.

In this talk, I would go over two of our recent research work at HPDS Research, and show how we address the above mentioned problems. Specifically, I present two approaches that we followed: (1) a scalable data reduction architecture that improves the performance per cost in all-flash systems by 57x, (2) a scalable I/O caching architecture that can boost the performance by over an order of magnitude.

"