Description
Extended TCP metrics were added as a part of IGNITE-6868 ticket. Unfortunately, they causes ~5-10% performance drop for cache operations when working in in-memory mode. Causes:
1) Contention on shared data structures (LongAdder, CMH)
2) A lot of string comparisons because we use class names to track particulat message types
Suggested fix:
1) Move all processing to NIO threads
2) Introcduce thread-local state and update it from NIO threads
3) Metrics readers should merge state from all NIO threads
4) User Message type ID instead of class names
5) When returning final result we should use Class.getName() instead of Class.getSimpleName()
Attachments
Issue Links
- is related to
-
IGNITE-6868 Implement new JMX metrics for TcpCommunicationSpi monitoring
- Resolved
- links to