An Empirical Evaluation of SystemC Transaction Data Storage Formats
TL;DR
The findings demonstrate that modern analytical storage techniques can significantly improve SystemC transaction tracing workflows and show that storage layout is a first-order determinant of footprint and query performance at scale.
Abstract
: SystemC transaction-level modeling (TLM) is widely used for system integration, virtual platform simulation, and early software bring-up, yet storage and analysis of TLM traces remains unstandardized and largely driven by convenience or legacy tooling. This paper presents a comparative empirical evaluation of storage formats for SystemC TLM-2.0 transaction traces and, to our knowledge, the first peer-reviewed evaluation of columnar (Apache Parquet) and array-oriented (Zarr) formats in this context. We evaluate CSV, SQLite, Parquet, and Zarr across workloads ranging from deterministic microbenchmarks to full-system Linux boot, measuring write-time overhead, storage footprint, and post-simulation query performance. Results show that storage layout is a first-order determinant of footprint and query performance at scale. Parquet achieves order-of-magnitude reductions in trace size and substantial speedups for scan-heavy and temporal aggregation queries, while SQLite consistently outperforms alternatives for selective, record-oriented access. No single format dominates across all workloads; however, the findings demonstrate that modern analytical storage techniques can significantly improve SystemC transaction tracing workflows.