Performance Testing

[lplewa]

Categories of Performance Tests

Performance tests can be divided into two main categories:

1. Artificial Benchmarks
2. Real UMF Use Cases

I intend to begin with Artificial Benchmarks but i'm open to feedback on this approach.

1. Artificial Benchmarks

Objective: Create controlled benchmarks to evaluate UMF configurations under various workloads.

• UMF Configurations: Set up different memory providers, with different memory pools and without.
• Workloads: Specify parameters such as allocation sizes and the ratio of allocations to frees etc....

Current Status:

• We have rally limited amount of benchmark using the ubench framework, which has limited functionality.

Proposal:

• Migrate to Google Benchmark:
  • Offers more features and is "an industry standard".
  • Similar to GTEST, which is already in use.
  • A lot of features which we would to implement while sticking to ubench, is included out of the box.
• Implement varied set of benchmarks which allows to test main functionality of UMF in each possible configuration.
• Stretch Goal:
  • Define benchmarks through configuration files to allow easier benchmarking of multiple cases without code changes.

2. Real Use Benchmarks

Objective: Benchmark UMF in real-world applications to assess performance in practical scenarios.

• Approach:
  • Use applications that utilize UMF directly or through a proxy library.
  • Preferably select applications with existing benchmark suites.

Current Need:

• A list of potential applications is yet to be compiled.
• Request: Suggestions for suitable applications are welcome.

Performance Testing Framework

We plan to employ GitHub Action Benchmark to automate performance testing.

Features:

• Parses test results and generates performance reports.
• Stores archival reports(on GitHub Pages).
• Generates charts displaying performance metrics over time (commits on the X-axis, metrics on the Y-axis). which are available on GitHub Pages.

Testing Strategy

• Pull Requests (PRs):
  • Run a selected set of benchmarks for each PR.
  • Compare performance against the main branch.
  • Optionally fail the workflow if performance degrades beyond a set threshold.
• Main Branch Commits:
  • Run a broader selection of benchmarks on each push to the main branch.
  • Update performance archive to serve as a reference for future PRs.
  • Update gh-page with a performance charts for a new commit.

Next Steps

To implement this performance testing plan, I will begin by migrating existing benchmarks from ubench to Google Benchmark. And integration GitHub Action Benchmark with our GHa CI/CD. When this will be complied, we will start extending list of artificial benchmarks, along this identifying Real Use one.

Along with this performance testing task we are planning to introduce CTL. CTL is an interface for examination and modification - it will be useful to read some internal statistic from providers/pools, which can be used as additional performance counters. More details about ctl will be provided in the separate issue.

Feedback Requested

We welcome any input on the following:

• Suggestions for real-world applications to include in our benchmarks.
• Ideas to enhance the benchmarking and performance testing process.
• Feedback on the proposed migration to Google Benchmark.

[pbalcer]

• Migrate to Google Benchmark:

  • Offers more features and is "an industry standard".
  • Similar to GTEST, which is already in use.
  • A lot of features which we would to implement while sticking to ubench, is included out of the box.

I used nanobench before, and found it very easy to use and quick. Way better than the colossal beast that is google benchmarks.

L0, OpenCL and UR use the benchmark tooling in compute-benchmarks. The benefit there is that everything is in one place and all the results across all the different projects use the same format.