Talks & Seminars
Title: Leveraging Plan Re-costing for Online Optimization of Parameterized Queries
Dr. Anshuman Dutt, Microsoft Research, Redmond
Date & Time: January 10, 2018 15:00
Venue: Conference Room, 01st Floor, C Block, Department of Computer Science and Engineering, Kanwal Rekhi (KReSIT) Building
Optimization of parameterized queries is a well-studied problem. For each incoming instance of a parameterized query, the challenge is to decide whether to optimize it afresh or reuse one of the previously stored plans for the same parameterized query. For each mistaken decision, we either pay unnecessary optimization overhead or execute a sub-optimal execution plan. Most existing solutions handle this challenge by identifying and storing a set of execution plans such that at least one of them is close to optimal for any future query instance. Later, when a query instance arrives, they rely on an efficient mechanism to select the most suitable execution plan. All such approaches fall under the term “Parametric Query Optimization” (PQO). Initial research efforts for PQO focused on an offline version of the problem where the set of all required plans are identified and stored even before the first instance arrives. Since the high startup cost of offline solution is not always acceptable, recently there have been attempts to handle this challenge in an online fashion, which is also the focus of this work. An ideal solution to online PQO would ensure that (a) the sub-optimality value for each query instance is guaranteed to be small; (b) only a small fraction of query instances are optimized, and (c) the set of stored plans is small and can adapt to changes in query distribution. Existing solutions to online PQO either provide sub-optimality guarantee or save optimizer overhead. Also, none of these techniques provide a mechanism to control the number of stored plans. In this talk, we start with an overview of existing solutions to PQO and then present our solution technique that perform well on all three metrics. Our solution is based on “plan re-costing”, a technique that enables us to evaluate performance of an existing plan for a new query instance, and at least 10x faster than fresh optimization. We show the effectiveness of our solution with empirical results on industry benchmark and real-world query workloads using a modified version of Microsoft SQL Server optimizer. (Talk based on research paper published in ACM SIGMOD 2017) (Joint work with Vivek Narasayya and Surajit Chaudhuri at Microsoft Research, Redmond.)
Speaker Profile:
Anshuman Dutt is a post-doctoral researcher in DMX group at Microsoft Research, Redmond. Previously, he completed Ph.D. from Dept. of Computer Science and Automation, Indian Institute of Science, Bangalore in 2016.
List of Talks


Faculty CSE IT
Forgot Password
    [+] Sitemap     Feedback