Learned scheduling for database management systems
Name
Ukyab-tsu-meng-eecs-2021-thesis.pdf
Description
Thesis PDF
Size
470.76 KB
Format
Adobe PDF
Checksum (MD5)
22f7e5fbed3ff1c04f0db53f5f0305d9
Author(s)
Ukyab, Tenzin Samten
Advisor(s)
Kraska, Tim
Date Issued
June 2021
Publisher
Massachusetts Institute of Technology
Abstract
Parallel database management systems need efficient job scheduling. Currently systems use simple heuristics ignoring the characteristics of database workloads. Therefore, we created an effective scheduler that uses machine learning techniques, such as reinforcement learning and neural networks, and does not require human intervention beyond an objective, such as reducing average job completion time. We use existing training techniques for job schedulers with dependency constraints. However, the model is specialized for database workloads using features specific to database queries, such as node operator type. In addition, we represent pipelining scheduling opportunities between operator tasks. With further training time our learned scheduler will be able to improve the average job completion time in comparison to heuristic schedulers, such as FIFO and fair scheduling.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright MIT
Persistent DSpace Link
