Description
Currently many metrics, including RpcMetrics and RpcDetailedMetrics, use a synchronized counter to be updated by all handler threads (multiple hundreds in large production clusters). As andrew.wang suggested, it'd be more efficient to use the LongAdder library which dynamically create intermediate-result variables.
Assigning to xkrogen who has already done some investigation on this.
Attachments
Attachments
Issue Links
- is duplicated by
-
HADOOP-13615 Convert uses of AtomicLong for counter metrics to LongAdder
- Resolved
- relates to
-
HADOOP-13782 Make MutableRates metrics thread-local write, aggregate-on-read
- Resolved