opentracker design notes ======================== Decisions that cost measurement to reach, and the alternatives that were tried and rejected. What the tracker is optimised for --------------------------------- UDP carries most of the traffic. An HTTP request costs about ten times a UDP announce, 26 microseconds of tracker CPU against 2.5. Nearly all of that is the kernel setting up and tearing down a connection rather than tracker code. A connection carries one request and closes. Keep-alive works and is rarely asked for. Statistics are approximate on purpose. Counters are per thread and summed when read, so a /stats read can under-report once. Decisions --------- One UDP socket per worker, where the kernel spreads datagrams over a group bound to one address. Cost per datagram then stays flat as workers are added, 0.71 to 0.72 microseconds from one worker to eight. Sharing one socket grows it eightfold over the same range, 0.71 to 6.07. listen.udp.workers defaults to 1. One worker beats none by 19 to 37 per cent everywhere, because a blocking read costs one system call where the event loop costs two. Beyond one the gain depends on the kernel balancing, so raising it is the operator's decision. A thread parked in a blocking read is woken by a signal, not by closing or shutting down its socket: mechanism Linux FreeBSD macOS close() blocked blocked 0 ms shutdown(SHUT_RD) 1 ms 1 ms blocked SO_RCVTIMEO works works works pthread_kill 1 ms 1 ms 1 ms Only the signal works everywhere. A receive timeout also works, but wakes every thread on every tick for the life of the process to catch something that happens once: at eight workers it doubled idle context switches, 255 to 495 over thirty seconds. The cleaner is the exception, and waits on a pipe nobody reads. A poll timeout is an interval rather than a deadline, so no clock enters into it. Each binary is built in one compiler invocation. Without object files a build cannot link objects compiled with different FEATURES, which was a real failure mode. It costs incremental rebuilds, 0.075 seconds becoming 1.56. A socket is read or written, never both. http_interest_set replaces an add-and-drop pair so that the invariant cannot be broken by accident, and the event layer hears about a change once instead of twice. Accept waits until the client's request has arrived: TCP_DEFER_ACCEPT on Linux, the accf_data filter on FreeBSD. Measured against a flood of silent connections, the filter serves more requests than no filter, not fewer, so it is not the denial of service vector that FreeBSD-SA-02:26 once described. The torrent table is split into 2^OT_BUCKET_COUNT_BITS buckets, 1024 by default, set from the Makefile. Inserting a torrent opens a slot in a sorted vector, so the bytes moved grow with the bucket's occupancy. Insert cost per torrent, Linux: 2^10 2^12 2^14 2^16 200k 466 ns 439 ns 1M 834 ns 579 ns 529 ns 4M 2348 ns 1163 ns 770 ns 674 ns Re-announcing an existing torrent moves by under ten per cent over the same range, so the cost is paid while the torrent population is built and not in steady state. Below a few hundred thousand torrents there is nothing to win. 2^14 captures most of the gain for two per cent more resident memory. The cleaner's wake rate does not follow the bucket count. A cap of ten wakes a second sets the sleep, and a wake cleans as many buckets as it must to finish a sweep in OT_CLEAN_INTERVAL_MINUTES. At the default bucket count that is one bucket every 117 ms. A refused announce is answered over UDP with a well formed announce for an empty swarm rather than with the error action. BEP 15 binds a client to the interval it is handed and says nothing about what a client should do with an error, so an error puts the client on its own retry schedule instead. Over five minutes against one blacklisted torrent, libtorrent 2.0.8 sent seven connects and three announces to a tracker answering with the error and two of each to one answering with an empty swarm, and was still going on a widening backoff. Transmission 3.00 retries twenty seconds after an error, then at five, fifteen, thirty, sixty and finally every hundred and twenty minutes. That last figure sets OT_CLIENT_REJECT_INTERVAL: at anything under two hours Transmission's own backoff would be the cheaper of the two in steady state. Neither client bounds a tracker's interval from above, and libtorrent clamps it from below at five minutes. The reject interval is a constant and not a config entry, because the ordinary announce interval is a constant too and an operator has nothing to measure a different value against. Over HTTP a refusal stays a failure reason, which is the only place a user is told why. Startup and teardown order -------------------------- Nothing that consults the access list runs before the list has been read. The loader used to do its first read inside its own thread, so whatever ran next saw an empty list: a state file read with -l was checked against it and a blacklist build restored torrents it should have refused, a whitelist build restored none at all. Announces arriving in the same window went the same way, 3089 of them refused over 373 ms against a two million entry whitelist. The first read is therefore synchronous in accesslist_init, and the thread only handles reloads. Nothing that adds a peer starts before the state file has been read. That is the TCP and UDP listeners and live sync, so main starts all three after load_state rather than leaving live sync inside trackerlogic_init. A cluster packet arriving in that window created the torrent, and the saved state for it was then dropped on the exactmatch add_torrent_from_saved_state bails on, losing that torrent's completed count. Teardown stops the cleaner first. It is the one worker that calls into other modules, accesslist_cleanup and stats_cleanup, so stopping it after them left it locking a mutex accesslist_deinit had already destroyed: five shutdowns in sixty at a bucket count low enough to make the window easy to hit. Signal handlers are installed before the first thread exists. A thread created earlier runs with whatever disposition a signal still has, and SIGUSR1, the one that wakes a blocked worker, terminates a process by default. Paths not taken --------------- A keyed hash for the bucket index. Info hashes are SHA-1 output and already uniform, so it changes no distribution. It would buy resistance to deliberate bucket flooding and cost the ordering a full scrape depends on: buckets are ordered by hash prefix and sorted within, so walking them emits the sorted dictionary bencode wants, with no sort and no copy. Cache line padding for the per-bucket locks and vector headers. The contention it was meant to remove is not there. On glibc, where a 48-byte mutex straddles a line, padding gains 3 to 8 per cent, but gains as much at one thread where no false sharing can exist, and scaling from one to eight threads is identical either way. On FreeBSD, where eight 8-byte mutexes share a line and the risk should be highest, padding is 15 per cent slower at one thread and neutral to 6 per cent slower at four and eight. accf_http, the HTTP-aware accept filter. It duplicates a parser the tracker has anyway, puts more kernel code in the path, and holds a connection that sends "GET /" and stops, which accf_data hands straight over. recvmmsg and sendmmsg. No measurable gain on either FreeBSD machine, between -4 and +10 per cent, and they do not exist on Darwin. Link time optimisation. It cuts .text by 13 per cent, 93,497 bytes to 81,630, and changes throughput not at all. A tracker waits on memory and system calls, not on instruction fetch: the announce path is 6.6 KB of code against a 32 KB instruction cache at worst. A directory for object files. bmake has no pattern rules and the Makefile builds under FreeBSD's make today. Building each binary in one invocation clears the objects out of the tree without giving that up. Platform differences that cost us code -------------------------------------- SO_REUSEPORT means three different things. Linux spreads datagrams over the group; FreeBSD permits the bind and needs SO_REUSEPORT_LB to spread. Darwin hands every datagram to whichever socket bound last, so a group there starves all but one reader. Waking a blocked reader differs per kernel. see the table above. Deferring accept is TCP_DEFER_ACCEPT, or accf_data, or nothing. make is bmake on FreeBSD and GNU make elsewhere, and the Makefile has to build under both.