Measured first: one libmicrohttpd thread with 64 KiB a connection held uploads to
13 MB/s, and four connections to 15.3, on a link where FTP reached 25 / 30.5 MB/s.
PS5-Suite's design was reviewed (persistent connection pool, small files per
connection, large files as pwrite ranges, big socket buffers) and an engine built
from scratch on the WebUI's HTTP (src/webui_transfer.{h,cpp}, docs/WEBUI.md):
- a pool of four server threads, 512 KiB a connection, 4 MiB socket buffers, writes
in 1 MiB blocks; non-transfer routes under one lock as on the single thread;
- PUT /api/upload/batch: many small files in one request (a framed stream), each
renamed into place when whole, existing names skipped and named unless replace;
- upload sessions (/api/upload/session, /part, /commit): a large file's ranges in
any order on any number of connections, committed once covered;
- Range downloads (206); every upload still atomic, never over an existing name;
- the page plans batches, sessions and single uploads onto six lanes.
Evidence: evidence/webui-transfer-engine (base PS5 13.40, PC on Wi-Fi 6). Uploads
26.0 MB/s on one connection and 29.9 MB/s for one file in six parts (FTP: 25 /
30.5); 400 x 32 KiB at 215 files/s (FTP 105). The page against the console: 605
files in 20.1 s, all equal by FTP readback, duplicates refused, a cancelled 200 MiB
file left nothing. ThreadSanitizer under a mixed parallel load: no report.
Command: build/tmp/transfer/bench.py; node tests/webui_transfer_browser.cjs;
tools/verify.sh: all five PASS.
