Skip to main content

Agent Scale Test

Overview​

Agent Scale Tests stress FBOSS features at production-like scale — for example programming the maximum number of ECMP groups/members and the maximum number of ACL table entries. They share the Agent Hw Test framework (see Agent Hw Test) but link only the scale test cases, so they are a focused way to validate that an ASIC and SDK can hold the resource counts FBOSS expects in the fleet.

Setup​

Like agent hw tests, scale tests run on a standalone DUT with no cabling required; dataplane traffic is exercised via front-panel ports in loopback. They use the same per-platform agent config under fboss/oss/hw_test_configs.

Test Cases​

All scale test cases can be found at:

https://github.com/facebook/fboss/tree/main/fboss/agent/test/agent_hw_tests (see AgentEcmpScaleTests.cpp and AgentAclScaleTests.cpp)

Run Modes​

FBOSS agent runs in one of two modes, and scale tests can run in either:

  • multi_switch (recommended) — the SwSwitch control plane and the HwAgent (ASIC) run as separate processes communicating over Thrift, exactly as in production. run_test.py brings up the hw_agent service automatically. Binary: multi_switch_agent_scale_test.
  • mono (deprecated) — SwSwitch and HwSwitch run in a single process. Binary: sai_agent_scale_test-sai_impl.

Prefer multi_switch — it mirrors the production process model. mono is deprecated and will be removed in the future. If --agent-run-mode is not specified, run_test.py defaults to multi_switch.

Steps to Run Agent Scale Test​

  1. Build and package FBOSS following the build guide and copy the package to the switch per Packaging and Running FBOSS HW Tests on Switch. You must be in the /opt/fboss directory when running tests.

  2. Run the full scale suite (recommended). run_test.py automatically discovers and runs all scale tests; pass --skip-known-bad-tests to skip tests known not to pass for your ASIC/SDK (key format vendor/sdk/sdk/asic/mode):

    ./bin/run_test.py sai_agent_scale \
        --config ./share/hw_test_configs/$CONFIG \
        --agent-run-mode multi_switch \
        --skip-known-bad-tests "<vendor>/<sdk>/<sdk>/<asic>/<mode>"

  3. Run a single test (optional). --filter is only needed to narrow the run; the test name below is for illustration only — use any valid gtest name:

    # multi_switch (recommended)
    ./bin/run_test.py sai_agent_scale \
        --config ./share/hw_test_configs/$CONFIG \
        --agent-run-mode multi_switch \
        --filter AgentEcmpTest.CreateMaxEcmpGroups

    # mono
    ./bin/run_test.py sai_agent_scale \
        --config ./share/hw_test_configs/$CONFIG \
        --agent-run-mode mono \
        --filter AgentEcmpTest.CreateMaxEcmpGroups

Common flags:

  1. --filter: GTEST filter passed through to the test binary (doc).
  2. --agent-run-mode: mono or multi_switch (default multi_switch).
  3. --num-npus {1,2}: number of NPUs to run in multi_switch mode (default 1).
  4. --skip-known-bad-tests: ASIC/SDK key used to skip known-bad/unsupported scale tests. Scale tests use the dedicated scale known-bad list fboss/oss/hw_known_bad_tests/sai_agent_scale_known_bad_tests.materialized_JSON (auto-synced from Meta's internal config); unsupported tests still come from fboss/oss/sai_hw_unsupported_tests.