This release-candidate cut captures further fine-tuning of the DCPerf v2 benchmarks since v2.0-rc1, most notably FeedSim v2, as well as other bug fixes and feature improvements. ./benchpress_cli.py --version and the benchmark result JSON reports now report v2.0-rc2.
FeedSim v2 fine-tuning
This release further fine-tunes FeedSim v2 to more closely model the performance and workload characteristics of a real-world ranking and aggregation service. The changes tighten the match to production along several axes:
- Dependency updates: updates libtorch to the current latest version (2.13.0) and makes dependency versions overridable via environment variables.
- Performance stability: a more stable and reproducible throughput search (adaptive driver pipeline depth and measured-QPS search bounds) plus deterministic per-request RNGs, so relative performance across CPU generations tracks production more consistently run to run.
- Performance and microarchitecture behavior: a memory-streaming stride sweep adds tunable DRAM traffic per extractor call to bring the memory-bandwidth, cache, and read:write profile closer to production.
- Instruction mix: revised feature-extractor helper math runs to increase integer operations and reduce scalar floating point calculations, making the instruction mix integer-heavy like production feed extraction; changed DLRM batch size to have SIMD composition match better with production workloads.
- Hot function composition: transport encryption (driver-to-leaf and mock-services TLS with hardware AES-GCM) and a calibrated workload recipe shift the CPU time distribution across extraction, inference, compression, and serialization toward the production profile.