Commit Graph

54 Commits

Author SHA1 Message Date
Matt Joiner f235c8ea51
Reuse piece order state for request ordering 2022-05-12 10:34:28 +10:00
Matt Joiner 23a0e7f7d3
Remove redundant type conversions 2022-05-09 12:51:02 +10:00
Matt Joiner a54d2d81e4
Retain peer local request ordering 2022-05-09 12:05:50 +10:00
Matt Joiner 9a9c7dee00
Add and use typed roaring bitmap 2022-05-09 12:05:50 +10:00
Jonathan McDowell a1a820d3c9
Avoid heap allocation in GetRequestablePieces (#734)
The calculation of whether we should ignore a piece in
GetRequestablePieces ends up doing an allocation for every piece, when
all we really need to do is query the index in the torrent. Provide an
IgnorePiece function instead, which avoids the need for a temporary
allocation.

Observed to cut out 40% of object allocations in some workloads (large
download, lots of seeds).
2022-04-11 14:22:05 +10:00
Matt Joiner 934944e498 Add AjwernerBtree as an alternative btree backing for piece request order
Also add some scans to benchmarks. Make a few changes to reduce allocations using piece request order.
2022-01-22 18:40:33 +11:00
Matt Joiner c54dc22592 Benchmark PieceRequestOrder with varying styles of path hint usage
Add per-piece piece request order path hints
2022-01-22 18:39:36 +11:00
Matt Joiner 62060fb897 Use tidwall/btree generics for piece request ordering 2022-01-22 18:39:36 +11:00
Matt Joiner 1bae62fd22 Dynamic outbound max requests
This might be one solution to https://github.com/anacrolix/torrent/issues/698.
2021-12-24 08:55:57 +11:00
Matt Joiner 506ff8d037 Use relative availabilities to determine piece request order
Most overhead comes from peers that connect that have everything, and we just increment every single piece's availability. There may be some unresolved determinism with torrents that share the same ordering.
2021-12-23 14:00:00 +11:00
Matt Joiner 1668a18859 Some optimizations in PieceRequestOrder.Update and item comparisons 2021-12-15 19:18:16 +11:00
Matt Joiner 04beb8937f More consistent cancellation management 2021-12-12 18:35:02 +11:00
Matt Joiner 30375615b3 Remove unused request strategy code 2021-12-12 18:35:02 +11:00
Matt Joiner 1bc84e316f Don't group Torrent piece request orders without a storage capacity together 2021-12-12 18:35:02 +11:00
Matt Joiner 8c5df03e14 Minimize allocations due to heap escapes 2021-12-12 18:35:01 +11:00
Matt Joiner 1b92249f2a Improve panic message 2021-12-12 18:35:01 +11:00
Matt Joiner 8025d15155 Stop iterating pieces when storage is exhausted 2021-12-12 18:35:01 +11:00
Matt Joiner 135f21fb64 Use interfaces to lazily expose the bare minimum inputs to GetRequestablePieces 2021-12-12 18:35:01 +11:00
Matt Joiner 4088e847f3 Provide mapping from infohash to Torrent in Input 2021-12-12 18:35:01 +11:00
Matt Joiner 94bb5d40ba Implement piece request ordering with retained state 2021-12-12 18:35:01 +11:00
Matt Joiner b99dd505b5 Avoid reflection sorting request pieces 2021-12-12 18:35:01 +11:00
Matt Joiner de2bea502c Use indices to lookup pieces for comparison 2021-12-12 18:35:01 +11:00
Matt Joiner 231301f5d7 Cache piece request orderings 2021-12-12 18:35:01 +11:00
Matt Joiner 67ed5d0032 Partition piece request strategy by storage capacity key 2021-12-12 16:56:01 +11:00
Matt Joiner 8e707c9b15 Remove unused peer stuff in request strategy 2021-11-29 10:14:16 +11:00
Matt Joiner c6ee03f449 gofumpt 2021-11-08 14:47:01 +11:00
Matt Joiner 7d0be0ac65 Optimize chunk calculations in request strategy 2021-10-19 14:08:56 +11:00
Matt Joiner 0f53cbf07e Change peer requesting to spread requests out evenly 2021-10-19 14:08:13 +11:00
Matt Joiner eab111dd84 Switch Peer.PieceAllowedFast and several request strategy inputs to raw roaring.Bitmaps
This is in preparation to support encoding request strategy run inputs for benchmarking.
2021-10-05 20:06:23 +11:00
Matt Joiner 1d2d1a9cde Store peer requests in a bitmap 2021-09-19 15:16:37 +10:00
Matt Joiner 21358ba458 Do peer requests separately for each peer 2021-09-18 20:34:14 +10:00
Matt Joiner cfb23e271f Export request_strategy.GetRequestablePieces 2021-09-18 18:57:50 +10:00
Matt Joiner c6fb0535d4 requesting: Remove some obsoleted condition checks
These shouldn't be necessary since peers were pruned for the ability to allocate chunks to the current piece.
2021-09-18 12:53:22 +10:00
Matt Joiner dd52a1440f request_strategy.Peer.canRequestPiece performance: Reorder conditions 2021-09-18 12:53:22 +10:00
Matt Joiner b757b62cf4 requesting: Filter peers that can't be allocated chunks 2021-09-16 10:47:27 +10:00
Matt Joiner a4697497aa Performance improvements in piece requesting 2021-09-16 00:12:24 +10:00
Matt Joiner 2b443d695b Change torrent capacity to not return a pointer
It's an unnecessary complication for a storage implementer.
2021-09-15 23:33:17 +10:00
YenForYang a8db640c62
Drop bradfitz/iter dependency (#605)
* Drop bradfitz/iter dependency

`range iter.N` looks nice and doesn't allocate, but unfortunately using a `range` expression blocks a function from being inlined wherever it's used (for now). It's not that we need inlining in all cases, but I do think a C-style for loop looks just as nice and is probably clearer to the majority. There also aren't any clear disadvantages to changing (unless you just happen to dislike the look of C)

* Update misc_test.go

* Update rlreader_test.go

* Update torrent_test.go

* Update bench_test.go

* Update client_test.go

* Update iplist_test.go

* Update mse_test.go

* Update peerconn_test.go

* Update peerconn.go

* Update order_test.go

* Update decoder_test.go

* Update main.go

* Update bench-piece-mark-complete.go

* Update main.go

* Update torrent.go

* Update iplist_test.go

* Update main.go
2021-09-14 13:46:50 +10:00
Matt Joiner 6c9a7fec39 Optimize peersForPieces allocations for make and sorting 2021-09-10 23:07:10 +10:00
Matt Joiner e9193968a3 Disable ensureValidSortedPeersForPieceRequests
Should approximately half the contribution of the parent function to overhead.
2021-09-10 23:07:10 +10:00
Matt Joiner 016bf1b07c Optimize sortPeersForPiece in allocatePendingChunks 2021-09-10 23:07:10 +10:00
Matt Joiner 7096da9f03 Minimize allocations and resizing in request computation 2021-08-01 22:02:40 +10:00
Matt Joiner f50b8fc2fa Fix issue with duplicate preallocated requests
Likely introduced by allowing actual and next request states to differ.
2021-06-07 13:01:40 +10:00
Matt Joiner 1a27e14391 Don't always force piece reallocation
The balancing for unforced pieces may not be exactly what we want yet.
2021-06-07 13:01:39 +10:00
Matt Joiner ad298364aa Add client-level max unverified bytes 2021-06-07 13:01:39 +10:00
Matt Joiner 63b0e42731 Get max unverified bytes working 2021-06-07 13:01:39 +10:00
Matt Joiner 307d6d178f Prepare to implement max unverified bytes 2021-06-07 13:01:39 +10:00
Matt Joiner 8e9cb9f2be Do checks for preallocated requests too
Otherwise we reserve requests with the assumption that they can be assigned later, and they actually might not be.
2021-06-07 13:01:39 +10:00
Matt Joiner 9dd85e5bd2 Fix non-deterministic test 2021-06-07 13:01:39 +10:00
Matt Joiner b80e2b08b5 Get request rebalancing working optimally! 2021-06-07 13:01:39 +10:00