Performance¶
The project measures whole-file reads against asammdf's batched select() API. On the recorded September 2026 comparison, the geometric mean for the 12 files over 1 MB was 4.5×, with a slowest included file at 1.8×. Those figures come from a release build, one warm-up, and the median of three measured runs on an Apple-silicon Mac; only files with equal decoded sample counts are included.
The two largest fixtures were asammdf-written repetitions of J1939 logs, so they measure size and decompression volume more than vendor structural variety. Vendor-written compressed files were last measured near parity in an earlier audit and have not been remeasured with the current inflater. See the full comparison and raw reports for file sizes, compression, memory, and sample-count exceptions.
Memory and backends¶
| Backend | Use when | Constraint |
|---|---|---|
| Memory mapped (default) | A completed local file stays unchanged during reading | Truncation by another process can raise SIGBUS |
| Buffered | A file may change, be replaced, or live on a network share | Copies bytes as it reads; memory use varies by file |
The ordinary read path assembles a data group's records before decoding them. Peak memory therefore scales with the largest data group, not simply the requested channel. signals_chunks reads in bounded windows for large recordings, including unsorted groups.
Do not assume that buffered I/O always uses less memory. The project's measured files show different outcomes at different sizes; the backend choice should primarily follow whether the underlying file can change.
Reproduce the comparison¶
The repository holds the benchmark harness and generated summaries:
The recorded comparison used asammdf 8.7.2. The second command requires a Python environment with asammdf; a run with no corpus or no asammdf reports a skip, not a benchmark result.