ydb/library/cpp, branch fix-header-ident-bloomMirror of YDB github reposhttps://code.mastervirt.ru/ydb/atom?h=fix-header-ident-bloom2023-08-30T23:51:25Z2023-08-30T23:51:25Zpgpg@yandex-team.com2023-08-30T23:33:19Zurn:sha1:13004290f7ae5603115f0cd10c6a0e5d0b9c5be9finally, allow more than one listener thread2023-08-30T18:27:36Zkulikovkulikov@yandex-team.com2023-08-30T17:53:24Zurn:sha1:3f2ddee8b16d59a940bc71a44b6384e6382ad1a4
With single poller thread on incoming connections, every OS scheduler latency on this thread wakeup directly affects requests timings. With oneshot poll events, we can poll on the same poller from many threads, and, if one thread has stalled for some reason, some other will take it's work on the next incoming event. So:
- make vector of listener threads instead of single one;
- add nListenerThreads option;
- stop request queues and listening sockets from the last finished thread;
- check incoming options and set OneShotPoll if needed.
There is a problem around removing connections on MaxConnections limit or ExpirationTimeout. There is no simple way to safely remove items from epoll (https://lwn.net/Articles/520012/) if it has raw pointers in event data. Try to handle it via postponed deletion of connection objects, wait until all listener threads are ready to reenter poller wait and there are no threads where deleted object can be used:
- close socket immediately after remove from poller, but instead of immediate TClientConnection destruction, put it to "pending delete" list;
- add cleanup state with thread mask, each bit stating that corresponding thread should reenter poller;
- call Cleanup routine before each poller wait, it will switch to 0 current thread's bit for each pending connection;
- when thread mask becomes all zero, really delete the connection;
- force there is a timeout for poller wait, ensure that all threads do reenter;
- add more configurations for some tests.
There is no significant changes or overhead for standard case with single listener thread, cleanup and pending deletion are just skipped. Also there is no overhead for common case where removing connections is rare. Here is the same review with nListenerThreads = 4 by default https://a.yandex-team.ru/review/4413226.
Fix input variable missprint2023-08-30T17:50:06Zsvidyuksvidyuk@yandex-team.com2023-08-30T17:31:54Zurn:sha1:c76aaf823d18acf064939d806427b39cea1bbe16All .ll files support in LLVM_BC2023-08-30T09:42:20Zsvidyuksvidyuk@yandex-team.com2023-08-30T09:12:06Zurn:sha1:cb3da9494c53283f0230ad37e4e8d0ea61b7d8fc2023-08-30T02:04:37Zpgpg@yandex-team.com2023-08-30T01:40:44Zurn:sha1:d41c3a928314a3cbea37acf3cc44efcafe2fd5dcmore preparations for more than one listener thread2023-08-29T23:24:59Zkulikovkulikov@yandex-team.com2023-08-29T18:38:44Zurn:sha1:6b340452cae7be442a49905343f8b76f1a801bf6
- apply one shot poll (under option) for listening socket too;
- some code rearrangements.
sample_by_rate filter: add volume limit on data stream in filter2023-08-29T20:57:27Zgitnabgitnab@yandex-team.com2023-08-29T15:06:47Zurn:sha1:68766cbb191fb70ae7eb497e3dbc7029f92ee11fPretty printing for TNode2023-08-28T19:44:55Zermolovdermolovd@yandex-team.com2023-08-28T19:21:54Zurn:sha1:83bf33d776ac03c37162fa324f659ce191776308Removed the last typedefs in the codebase2023-08-28T15:41:07Zsabdenovchsabdenovch@yandex-team.com2023-08-28T15:11:46Zurn:sha1:4d9954dde66b7a3383ca8cfa5540bf5bfbb19c54
removed a lot of typedefs and turned a couple of std::pairs into
something more meaningful
[fetcher] support unistat2023-08-28T08:01:49Zkgershovkgershov@yandex-team.com2023-08-28T07:47:37Zurn:sha1:02bb32604c0d20f5af0ce6c7e2d294206153045d