On 26/Oct/2009 20:33, Jesse Wilson (JIRA) wrote: > Logging performance improvements > -------------------------------- > > Key: HARMONY-6362 > URL: https://issues.apache.org/jira/browse/HARMONY-6362 > Project: Harmony > Issue Type: Improvement > Components: Classlib > Environment: SVN Revision: 829932 > Reporter: Jesse Wilson > Priority: Minor > > > The logging module performs more synchronization and object > allocation than necessary. In particular, getHandlers() always needs > synchronization and creates an array. By adopting util.concurrent > features (CopyOnWriteArrayList) and coding to the common case (when > there are zero handlers) I saw a 2.5x improvement in throughput for > calls to logger.log(Level, String). In a benchmark that measures calls > to "logger.info", call time improved from 385000ns to 154000ns per message.
The improvement in performance is impressive. What is the contribution made by the use of the CopyOnWriteArrayList? I'll tell you why I am asking... This would be our first dependency on util.concurrent from the logging module, and if the benefit is small then I'd prefer to see reduced coupling between the modules rather than a small improvement in performance. Which is all part of the larger 'meta' question... What is the smallest set of modules that we see forming a useful runtime core? Today, people can reuse our logging implementation even if they don't have an implementation of concurrent. Should we relax that capability and assume that concurrent support is going to be universal? What is the smallest useful set of modules that we should declare are necessary for a runtime, so expect inter-module dependencies are ok; and which ones are always going to be optional/left behind in smaller runtimes? etc. Discuss :-) Tim
