- Codes for Storage with Queues for Access
- Project Year:
2018
- REU Student (s):
Dalton Burke | University of Colorado Denver/Anschutz Medical Campus CO
| Elise Catania | University of Rochester NY
- Student 1 Institution:
University of Colorado Denver/Anschutz Medical Campus
- Student 2 Institution:
University of Rochester
- Project Mentor:
Emina Soljanin
- Project Mentor Area:
Electrical and Computer Engineering
- Project Abstract:
With the rise of big data and machine learning, computers are used frequently to process and compute large quantities of data. An incoming job is parallelized, divided among different servers. Parallelizing introduces the issue of stragglers, the slowest workers. It is known that MDS coding schemes can be utilized so that the last servers to finish do not slow down the entire job. The known hierarchy of job scheduling policies from most efficient to least efficient is JSQ (join shortest queue), power-of-d, round robin, and random. The service time of balanced and incomplete block design (BIBD) is compared to the service time of other schemes in the hierarchy. BIBD is also tested with other models for which the hierarchy is unknown. At first, simplifying assumptions are made and then, a more realistic model of service time including both job specific randomness and server specific randomness is discussed. As a result, in certain scenarios, BIBD outperforms the other schemes in the hierarchy.
