Eno Thereska

Current research interests: understanding system management problems

• With a large percentage of total system cost going to system administration tasks, self-management remains a difficult and important goal in systems. As a step towards the self-management vision, I have developed a framework to enable systems to be self-predicting and answer ``what-if'' questions about their behavior with little or no administrator involvement. I have built a Resource Advisor inside two real systems: Microsoft's SQL Server database and the Ursa Minor storage system at CMU. The Resource Advisor helps with upgrade and data placement decisions and provides what-if interfaces to external administrators (and internal tuning modules). The Resource Advisor is based on efficient system behavioral models that enable robust predictions in multi-tier systems. The models are robust in that they discover regions of operation where the prediction confidence is high and regions for which they choose not to predict. The models handle performance anomalies by pinpointing their likely cause (e.g., system misconfigurations) and continuously collect historical information and refine themselves to account for unforeseen (and hence not programmed-in) workload-system interactions.

• In the context of self-managing systems, I have also worked on ABLE: Black-box learning of file usage characteristics, which allows a storage system to make educated guesses on how clients will use their files at file creation time and Continuous disk layout reorganization, which attempts to continuously refine the data layout on disks based on user access observations. Go to my publications page.

Previous research interests

• I have worked on Freeblock Scheduling. Most systems utilize many background, I/O bound services that balance load, re-encode data and repair corrupted data. I developed a clean API and framework for these background applications to tap into free bandwidth from busy disks. For background tasks, this framework uses idle time and freeblock scheduling, a new approach to utilizing more of disks' potential media bandwidths. Specifically, by interleaving low priority disk activity with the normal workload, one can replace many foreground rotational latency delays with useful background media transfers. With this framework, maintenance applications such as backup, cache write backs, data migration, etc., can make steady forward progress with little-to-no impact on foreground application workloads.