Performance optimizations for file reader
synCRON:
legacy references: CDIFHD-147
affected version:
Table of contents
Changes to the product
The performance goal is to get close to native NVMe-performance levels as far as the file read in isolation is concerned (not accounting for repackaging to samples on the ADTF side) - at least for streams with comparatively large (512kB buffer size or more) samples.
#113 sets a benchmark for write performance, and the reader has to be at least also as fast to allow real-time playback of e.g. uncompressed video streams and alike.
Benefit for customer
Real Time playback
Additional information
- Async read-ahead (io_uring / OVERLAPPED with unbuffered IO)
- Reduce syscall overhead for reads (registered user-space buffers)
- Check for unnecessary copies
- Check for unnecessary re-allocations
Edited by Florian Roth