OpenHFTBench: A Benchmark for Temporal Structure in Stateful Event Processing
Abstract
Public systems benchmarks still under-cover event-driven, stateful, latency-sensitive workloads. Existing serving and stream suites capture important domains, but often miss the combination of non-stationary event composition, long-lived mutable state, and architecture-facing observability needed for in-memory event processing pipelines. This paper presents OpenHFTBench, a benchmark framework for temporal structure in stateful event processing. It provides a reproducible 32-byte workload abstraction, a configurable family of benchmarks, and native plus gem5 execution paths with integrated profiling hooks. We instantiate the framework across multiple liquid Nasdaq symbols and market conditions using both a fully public LOBSTER-backed path and a richer Databento-backed path. The central result is that, for the stateful event-processing workloads studied here, preserving temporal structure is necessary for reliable architecture-facing conclusions: across regimes and symbols, OpenHFTBench preserves distinct state trajectories rather than collapsing them into one replay pattern, and workload-fidelity choices materially change measured behavior. Most importantly, a deterministic stationary proxy that preserves aggregate event counts still cuts final active state by about 50% and collapses the linear–ring instructions-per-cycle (IPC) gap from 4.75% to 0.57%. The public path reproduces the key trends and supports proxy-robustness checks. Together, these results position OpenHFTBench as a practical systems benchmark for studying temporal structure beyond simplified proxies.