[ http://jira.qos.ch/browse/LBCLASSIC-254?page=com.atlassian.jira.plugin.syste... ] Ceki Gulcu commented on LBCLASSIC-254: -------------------------------------- Hi Michael, After looking at your code, I got an idea and was wondering if it would meet your needs. Instead of keeping two maps, we keep a single map but copy it only if necessary. All we need for efficient/timely copying is a variable keeping track of the last operation. A copy is necessary on 'put' or 'remove' but only if the last operation was a 'get'.Get operations never necessitate a copy nor successive 'put/remove' operations, only a get followed by a 'put/remove' requires copying the map. With that in mind, let (m, 'op') designate the tuple consisting of hashMap 'm' and last operation 'op'. So, for example, "put: (m1, 'get') -> (m2, 'put')" states that in response to a put operation the map m1 is transformed into a new map m2. The lastOperation variable is changed from 'get' to 'put'. Here is an illustration of changes into the MDC structure with this notation: put: (null, null) -> (m0, 'put') put: (m0, 'put') -> (m0, 'put) get: (m0, 'get') -> (m0, 'get') get: (m0, 'get') -> (m0, 'get') remove: (m0, 'get') -> (m1, 'remove') get: (m1, 'remove') -> (m1, 'get') get: (m1, 'get') -> (m1, 'get') remove: (m1, 'get') -> (m2, 'remove') ... and so on Note that 8 operations were performed by only 3 maps were created. What do you think?
Performance improvement for LogbackMDCAdapter ---------------------------------------------
Key: LBCLASSIC-254 URL: http://jira.qos.ch/browse/LBCLASSIC-254 Project: logback-classic Issue Type: Improvement Components: Other Affects Versions: 0.9.28 Reporter: Michael Franz Assignee: Ceki Gulcu Attachments: LogbackMDCAdapter.java
During performance analysis of our application the LogbackMDCAdapter showed up as a performance hotspot. This is because the application does relatively often replaces multiple entries in the MDC at ones, e.g. 6 removes/writes without any intermediate log statement. Also actual log messages that come through filtering before creating the LoggingEvent are relatively rare in production environments. I have reworked the implementation to improve the performance. The main idea is to defer cloning the internal Map as long possible. This patch increased overall application performance by about 10% in that test. Other application types were MDC changes are small and many LoggingEvent objects are created (calls to getPropertyMap()) should not be affected significantly. Note: I haven't check if my patch would reintroduce the #LBCLASSIC-183, but my subclasses of InheritableThreadLocal does not override the initialValue() method, so it could work.
-- This message is automatically generated by JIRA. - If you think it was sent incorrectly contact one of the administrators: http://jira.qos.ch/secure/Administrators.jspa - For more information on JIRA, see: http://www.atlassian.com/software/jira