Enabling what-if explorations in a distributed storage system
Speaker(s) : Eno Thereska (Carnegie Mellon University)
Abstract: Systems should be self-predicting. They should continuously monitor themselves and provide quantitative answers to What...if questions about hypothetical workload or resource changes. Self-prediction would significantly simplify administrators’ planning challenges, such as performance tuning and acquisition decisions, by reducing the detailed workload and internal system knowledge required. This talk describes and evaluates support for self-prediction in a cluster-based storage system and its application to What...if questions about data distribution selection.
Bio: Eno Thereska is a fifth-year PhD student at Carnegie Mellon University, working with Prof. Greg Ganger. His current research interests are in building real distributed systems that are easy to manage. An approach he is currently pursuing puts sufficient instrumentation and modeling within the system, enabling it to answer several important what-if questions without outside intervention. He is interested in applying methods from queuing analysis (for components build from scratch) and machine learning (for legacy components) to this problem. As a testbed he is using Ursa Minor, a cluster-based storage system being deployed at Carnegie Mellon for researching system management issues. Concrete what-if questions in this system are about the effect of resource upgrades, service migration and data distribution.
In a previous life he worked on Freeblock Scheduling, an advanced disk scheduling mechanism that utilizes otherwise-wasted disk head rotational latencies and allows background I/O bound applications (e.g., defragmentation, backup, cache write-backs) to complete with no impact on foreground applications, even when there is no idle time.