Skip to content

[performance] Normalize methods to benchmark DSS #1557

Description

@the-glu

Describe the bug

When running performance tests to validate and debug issues, we noticed a significant difference in results between machines, even with identical parameters.

The CPU type and architecture create major variations, and the local machine load likely also plays a role.

For instance, running FlightInSubs on my machine (Intel(R) Core(TM) Ultra 9 185H) with DSS_IMAGE=interuss/dss:v0.22.0, NUM_USS=7, NUM_NODES=3, and INTER_USS_NETEM_CONF="delay 36ms 40ms 50% distribution paretonormal loss 0.25% 15%" is extremely slow- getting about half the q/s of a Mac M1. Furthermore, running the same test with datastore_max_open_conns set to 30 is impossible on my machine, whereas it actually increases performance on the M1.

Expected behavior

A solution would be to define a specific cloud-based VM instance and OS (and probably write a deployment script) to establish a common baseline and ensure comparable load test results.

Metadata

Metadata

Assignees

No one assigned

    Labels

    P1High priority

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions