Benchmark results for the NETPBM reader


Parent ↗

Summary

For iterated, i.e. cache fill time not relevant, unordered is faster than sequential, regardless of format.
Conversely text format is faster than binary, likely because of the in-memory cache.
For shots, where cache fill time is measured as part of the operation binary has no difference to iterated, as expected, as binary is without a cache.
For text format however the time to fill the cache is substantial, up to 8 seconds. Still around 6 seconds for thread count >= actual core count.

Look into ideas for a faster cache implementation than the current (row) vector cache, especially faster fill.
- Veccache uses quite a lot of locking between the contending threads while incrementally filling individual rows (i.e. separate allocations)
- Could a variant of the cache op be better, with a single thread once streaming the file into a single memory block ?
Maybe with incremental release of filled rows to the requesting threads interleaved with the fill itself ?
Look into caching binary too.
Look into other means of reading from a file instead of (serial) through a file descriptor.
- VIPS seems to support mapping of a file directly into process memory ?!