The request throttler had several problems:
- Throttling was very laggy, but when it reacted it throttled too quickly.
- PartialPieces got returned back to PeerCoordinator,
and then back to PeerState, every few seconds, rather
than continuing steady-state in PeerState
- Holding the PartialPiece in the PeerCoordinator for
several seconds while staying interested is poor and risks snubbing
- Max request queue size was reduced far too quickly,
doing so exponentially when the reduction occurred linearly
- shouldRequest() called the request limiter second in the conditional,
so the estimator was not decayed properly when over down bw limit
- shouldRequest() used overDownBandwidthLimit() rather than _down.offer(),
ensuring a very slow and laggy feedback loop.
- Requests were added too quickly on unthrottle
Fix these by:
- Fixing and slowing max request queue size reduction to simulate
an exponential reduction over time when throttled
- Fixing shouldRequest() to always call offer() for both limiters
- Slow addition of requests after unthrottle to simulate
a linear increase over time
Log tweaks also
The request queue now mostly maintains a steady size,
and properly meters the requests.
The PartialPiece returns to the PeerCoordinator much less often,
and comes back to PeerState much quicker if it does.
Further tuning TODO
- For fast peers:
Do not rerequest, peer is expected to meet its BEP 6 obligation,
and rerequesting needlessly inflates the request bandwidth estimate
- Drop dup requests for pieces peers already have
- Remove old comment about future enhancements we already did
- Fix repeated PartialPiece log about hole filled in
- Log tweaks
Should improve d/l from qbittorrent that aggressively reorders
for large piece sizes, use a buffer instead,
in two spots: creating and rechecking.
Pass the hash instead of the piece to MetaInfo.checkPiece()
Remove now-unused methods.
- Refactor padding file addition, inspired by similar code in I2P+
Now will be called from Storage, not done on-the-fly during
MetaInfo creation, which was the wrong place.
- Skip padding files in PieceLayers constructor based on attributes
- toString() tweaks for logging
- Add method to generate piece hashes from a File
- Make InputStream methods private
- Fix subclass name
- Add alternate construtor for PieceLayers
- Change PieceLayers constructor exception class
and disable XHR when hidden, to prevent nonce failures,
made worse by nonce changes in the last release.
We keep the XHR timer firing for now.
Cancelling/restarting timer is a possible part 2 TBD.
Credit: drzed for assistance
For now, disable it, since we now have good bandwidth limiting
in snark, we don't need the additional throttling/drops in I2CP
causing more streaming issues and making snark/streaming analysis/optimization harder.
The I2CP limiter is also bad and old and needs to be completely rewritten or tossed out.
for faster window size recovery (WIP)
- Fast: increase min from 2 to 16
- Timer: increase min from 2 to 8
- Both: Double Westwood BDP value used in calculation, use rounding
- Add missing synchronization in fast-retx path
- Log tweaks
zzz