[MXS-3700] Benchmark the overhead of qlafilter Created: 2021-07-29  Updated: 2021-10-04  Resolved: 2021-09-21

Status: Closed
Project: MariaDB MaxScale
Component/s: Documentation
Affects Version/s: 6.2.0
Fix Version/s: 6.2.0

Type: Task Priority: Major
Reporter: markus makela Assignee: Niclas Antti
Resolution: Fixed Votes: 0
Labels: None

Sprint: MXS-SPRINT-138, MXS-SPRINT-140

 Description   

The qlafilter provides valuable information but it comes at some performance cost. This task is for measuring that cost and determining whether it is reasonable enough to allow it to be enabled on production systems.

As a part of this benchmarking, an evaluation of strategies for reducing the performance cost can be done. This includes technical improvements to the logging mechanism, sparser sampling of clients, conditional sampling etc..

 Comments   
Comment by Niclas Antti [ 2021-09-21 ]

The QLA logging has been enhanced in version 6.2.0 (see "Fix Version" for accurate version).

The benchmark tests were conducted on a few different machines:

  • Storage was M.2 and SSD
  • Heavy continuous loads utilizing all available cores and HW-threads
  • Very fast connections to the database instances
  • Capturing all available data-points.

With the test setup(s) above the performance degradation is up to 6%. With normal loads and networks, the performance degradation can be expected to be less than 5%, or even insignificant if the system is never under heavy load.

Configuring log flushing had very little impact. In some cases it even improves overall performance. However, if storage is slow the performance is likely to be worse. Slow (hdd, nfs, etc.) storage was not tested.

Generated at Thu Feb 08 04:23:19 UTC 2024 using Jira 8.20.16#820016-sha1:9d11dbea5f4be3d4cc21f03a88dd11d8c8687422.