Visite de Gohar Irfan

Gohar Irfan PhD student at MIT, accueilli dans l'équipe KrakOS au LIG

Abstract:
Storage disaggregation is widely adopted in today’s datacenters because it significantly improves disk utilization. However, efficiently disaggregating SSDs presents a unique challenge due to their inherent high performance variability. We propose Sandook, a rack-scale disaggregated storage system that dynamically adapts to SSDs’ varying performance. It leverages real-time performance monitoring to adaptively steer I/O requests to the most suitable SSDs on a microsecond timescale. Its logically centralized architecture allows for the simultaneous integration of multiple scheduling policies, including profile-driven load balancing, read/write segregation, and SSD congestion avoidance. Our evaluation demonstrates that Sandook is able to unlock the full performance potential of SSDs, achieving a 1.7× increase in storage IOPS while maintaining sub-millisecond tail latency and a 12–94% improvement in end-to-end application performance without requiring code changes.