Download presentation
Presentation is loading. Please wait.
1
1 June 3, 2015June 3, 2015June 3, 2015 Resource Advisor for SQL Server Automating DBMS performance prediction Dushyanth Narayanan, Paul Barham, Eno Thereska, Anastassia Ailamaki
2
2 What and why Live system monitoringLive system monitoring –Lightweight, end-to-end tracing –Workload agnostic Automated analysisAutomated analysis –Answering “what-if” questions –Visualization To aid DB adminsTo aid DB admins –Resource upgrade decisions –Identify limiting resource Memory, disk, CPU, locks, …Memory, disk, CPU, locks, …
3
3 Outline InstrumentationInstrumentation –Where, how, and how much Initial ResultsInitial Results –“What if” I bought more memory? Current statusCurrent status –Papers, patents, etc. Future workFuture work –Storage, CPU, locking, … –Adaptive query optimizer
4
4 Instrumentation Resource usage / multiplex pointsResource usage / multiplex points –E.g. buffer touch, transaction start, … Source-levelSource-level –Private copy from SQL Server tree Function call interfaceFunction call interface –Automatically generated stubs Minimally invasiveMinimally invasive –Lightweight, non-blocking ETW events –189 lines modified in 6 files
5
5 Resource models Buffer managerBuffer manager –page reference trace, allocations –cache simulator DiskDisk –analytic model: single spindle, random access –disk params, Q length service time –queue length from throughput, #users CPU scalingCPU scaling –by clock speed, SPECint, …
6
6 Accuracy of “what-if”: throughput
7
7 Accuracy of “what-if”: mean latency
8
8 Outline InstrumentationInstrumentation –Where, how, and how much Initial ResultsInitial Results –“What if” I bought more memory? Current statusCurrent status –Papers, patents, etc. Future workFuture work –Storage, CPU, locking, … –Adaptive query optimizer
9
9 Status Submitted to MASCOTSSubmitted to MASCOTS Patent filedPatent filed –“Predicting database system performance” White paper for SQL ServerWhite paper for SQL Server –Tracing recommendations Potential tech transfer to IndyPotential tech transfer to Indy Collaboration with CMU (ongoing)Collaboration with CMU (ongoing)
10
10 Future Work Simulation of transaction control flowSimulation of transaction control flow –avoid limitations of analytic approach Storage model [with Thereska, Ganger @ CMU]Storage model [with Thereska, Ganger @ CMU] –random + sequential mix, RAID, … LockingLocking –what happens as #users increases? Making commit order deterministicMaking commit order deterministic –simulate the performance impact Resource feedback for query optimizerResource feedback for query optimizer Feedback-driven cohort schedulingFeedback-driven cohort scheduling
11
11 Resource Advisor architecture
12
12 End-to-end visualization Detailed, per-request informationDetailed, per-request information
13
13 Buffer cache model accuracy
14
14 Disk model accuracy
15
15 Changing the transaction rate
16
16 Latency has high variance
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.