Affects Version/s: None
Hadoop Flags:Incompatible change
Release Note:There is a possibility of file descriptor leakage issue under high workload. Please upgrade to the latest version of JVM or the version that has a fix for this bug (http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=7118373)
NIOServerCnxnFactory is single threaded, which doesn't scale well to large numbers of clients. This is particularly noticeable when thousands of clients connect. I propose multi-threading this code as follows:
- 1 acceptor thread, for accepting new connections
- 1-N selector threads
- 0-M I/O worker threads
Numbers of threads are configurable, with defaults scaling according to number of cores. Communication with the selector threads is handled via LinkedBlockingQueues, and connections are permanently assigned to a particular selector thread so that all potentially blocking SelectionKey operations can be performed solely by the selector thread. An ExecutorService is used for the worker threads.
On a 32 core machine running Linux 2.6.38, achieved best performance with 4 selector threads and 64 worker threads for a 70% +/- 5% improvement in throughput.
This patch incorporates and supersedes the patches for
New classes introduced in this patch are:
- ExpiryQueue (from
ZOOKEEPER-1444): factor out the logic from SessionTrackerImpl used to expire sessions so that the same logic can be used to expire connections
- RateLogger (from
ZOOKEEPER-517): rate limit error message logging, currently only used to throttle rate of logging "out of file descriptors" errors
- WorkerService (also in
ZOOKEEPER-1505): ExecutorService wrapper that makes worker threads daemon threads and names then in an easily debuggable manner. Supports assignable threads (as used by CommitProcessor) and non-assignable threads (as used here).