|
Eno ThereskaMicrosoft Research Ltd Email: etheres AT microsoft.com |
[ Bio ]
[ Resume ]
[ Publications ]
News:
- I am serving on the PC for SIGMETRICS 2009. Please consider submitting a paper
Research interests and selected projects:
I am a systems person, currently with a focus on file systems and storage technologies and high-performance data centers. I also have great interest in applying machine learning and queuing analysis to help simplify and automate system management.Storage systems
- Everest (MS internal only for now, watch out for our OSDI'08 paper) is a short-term distributed store that is used to temporarily off-load high I/O bursts in a data center environment. Everest is transparent to and usable by unmodified applications, because it does not change the persistence and consistency semantics of the storage.
- Ursa Minor is a versatile distributed file system [FAST'05]. It allows each client/application to encode its files differently to meet diverse performance, availability and confidentiality goals
- I have worked on Freeblock Scheduling. There are many maintenance applications, such as backup, cache write backs, data migration, etc., that annoy us especially when they interfere with foreground applications. It is surprising, but many of these maintenance tasks can run unnoticed, even when the disks are fully utilized! Freeblock scheduling interleaves low priority disk activity with the normal workload by replacing many foreground rotational latency delays with useful background media transfers.
What it takes to design self-predicting systems
- Robust math models from queuing [ICAC'06] and statistical theory [SIGMETRICS'08]
- End-to-end measurements [SIGMETRICS'06]
- Inherently more predictable system algorithms (e.g., performance insulation [FAST'07])
- We built Ursa Minor to be self-predicting from the beginning and converted a legacy system, SQL Server, to be self-predicting.
Black-box modeling and optimizations
- 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. A recent project explores optimization techniques and utility functions to help with storage system provisioning [FAST'08].