Paper Title: Minimizing Customer Interruptions Due to SSD Failures

Paper Abstract: Availability and performance are critical metrics for hyperscalers selling Virtual Machines (VMs). SSDs are a critical infrastructure component impacting these metrics due to their Annual Failure Rate and performance variability over time. Overallocation of SSD resources to protect from single device failures is an expensive way of resolving this problem. If SSD failures can be predicted, virtual machines can be proactively migrated to healthy nodes without impacting the end user or overallocating SSDs. In order to predict and respond to issues, the internals of the SSD must be made transparent, telemetry engines must periodically collect data, services must reliably pick out a signal, and automated node migration must be enacted. This presentation explores which telemetry data is the most valuable, how it is used, and how failure prediction compares to other solutions seeking to mitigate VM interruption.

Paper Author: Brennan Watt, System Architect, Microsoft

Author Bio: Brennan Watt is a System Architect and Product Planner for Microsoft Azure SSDs. He previously held a similar role at Intel for the better part of a decade. He has helped reshape the storage interface and device architecture through his contributions to Zoned Namespaces and Project Denali. He has brought over a dozen product lines spanning several generations of nand flash technology into mass production and acquired multiple patents for his efforts. He holds an MS from USC in Computer Science and BS from OSU in Electrical & Computer Engineering.