#302 - Generate benchmark configurations from profiles

[dleshchev]

Summary

Generate reproducible benchmark configurations from explicit profiles instead of maintaining repetitive Spark, cross-host, multi-queue, and transform YAML copies. The existing PR and discussion are preserved, while hardware-free C++ validation lives in the stacked foundation in #316.

Related issue

Closes #302.

Changes

• Add deterministic RawPairSpec and SocketPairSpec profiles, rendering with constrained overrides, and the installed gen_daqiri_config.py entry point.
• Generate raw DPDK/ibverbs loopback and cross-host roles, Spark multi-queue cells, VLAN/VXLAN/GRE/NVGRE transforms, and UDP/TCP/RoCE client/server roles.
• Preserve benchmark semantics, including explicit raw buffer capacity, configurable tx_eth_src, and the Spark payload sweep's fixed 8064-byte buffers.
• Drive both Spark benchmark harnesses from profiles and remove 17 repetitive generated YAML files plus their two superseded mutation helpers.
• Retain the canonical teaching configs and specialized HDS, reorder, quantize, latency, pacing, dynamic-flow/resource, and application configs.
• Keep focused portable tests on Python rendering, CLI/profile constraints, and generated structure; send the broad valid matrix through daqiri_config_validate from #315 - Add hardware-free configuration validation #316.
• Lead readers through the checked-in examples first, then present the generator as the repeatable path for system-specific, multi-queue, transform, and cross-host variants.
• Describe application-startup validation as the normal path and the standalone validator as an optional hardware-free preflight for CI, batch checks, and development without target hardware.

The normalized deletion audit mapped every removed YAML to its generated replacement. Intentional differences are canonical formatting, explicit role separation, derived queue/flow lists, standardized generated identifiers, explicit transform engines/matches, role-appropriate socket payloads, one-way RoCE roles, and fixed 8064-byte Spark buffers. No unintentional semantic differences remained.

Validation

• Privileged project-container Release configure/build with DAQIRI_BUILD_EXAMPLES=OFF, DAQIRI_ENGINE="dpdk ibverbs", and GPU/hugepage access: passed.
• Installed validator and generator entry points: passed.
• scripts/check_pr.sh: 46 portable tests, all 27 retained example/application configurations, 1 foundation compatibility case, 14 expected-invalid foundation parser/common-semantic regressions, 23 valid generated configurations, strict MkDocs build, HTML links, and documentation-reference checks passed.
• Generated output determinism across independent Python processes: passed.
• Shell syntax, Python compilation, clang-format --dry-run --Werror, DCO trailers, and git diff --check: passed.

The generated RoCE role checks emit nonfatal warnings for memory regions retained for the opposite role; validation still succeeds. This does not change benchmark behavior.

Dependencies or follow-up

Depends on PR #316. This PR is temporarily based on feat/315-config-validator so its review diff contains only generator/configuration migration work. After #316 lands, this branch can be rebased onto main without adding an independent parser, validator implementation, or JSON Schema.

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Replaces hand-written benchmark configs with a generator script.

The PR appears safe to merge based on the reviewed changes.

Summary

The PR replaces repetitive benchmark YAMLs with deterministic raw and socket profiles, migrates the Spark runners to generated configurations, and adds validation and documentation. Subsequent changes add raw-ibverbs destination-MAC resolution and reorganize setup guidance.

_{Reviews (5) · Last reviewed commit: "#302 - Align cross-host generator guide ..."}