Details
-
Bug
-
Status: Closed
-
Blocker
-
Resolution: Not A Problem
-
2.0-beta7, 2.0-beta8
-
Windows 7 64 bit, JDk 1.6.0_43
Description
Hello,
we encountered a severe performance loss with log4j 2. Unfortunately we have no exact error description such as a stacktrace but I'll give it may best shot.
We are developing an application working with mass data. During the development we are running tests such as importing 15,000 files via FTP or creating 500,000 business objects and workiong with them.
Switching from log4j 1.2.17 to log4j 2 we ran into really severe performance issues.
With the beta7 we even get an OutOfMemoryError. Using a profiler we see that there was a huge amount of string objects (that were log messages according to a control sample).
Switching to the beta8 the profiler shows there are only 10 % of the string objects left that we had before but it still was a big number. Especially the strings live for a really long time (> 100 generations).
Using the RollingFileAppender and ConsoleAppender, (not a FlumeAppender for which a memory has been fixed in beta8) and default garbage collection we had the following behaviour with beta8:
Starting the application (each time exactly the same test scenario) the machine fastly reaches the memory limit given for the JVM. Afterwards periodically garbage collection is invoked, causing a long "Stop the World Phase"
(8 minutes in another scenario), freeing memory down a "base level" and the application continues until the memory is full again. The base level remains the same for next invocations of the garbage collection.
This may relate to tickets: https://issues.apache.org/jira/browse/LOG4J2-323, https://issues.apache.org/jira/browse/LOG4J2-322, https://issues.apache.org/jira/browse/LOG4J2-304.
Furthermore when looking into one string object being not GCed (372 KB) I noticed that it contained a lot of log messages being concatenated. The application contains no log statement that logs such a big message.
Looking at the string towards the end I found even complete log events including the log level and so on not only the message. This feels being related to https://issues.apache.org/jira/browse/LOG4J2-290, somehow log events /
strings get concatenated / corrupted.
Maybe it is even a problem with the ParameterizedMessage which we use a lot of.
Switching to the AsyncAppender (FastRollingFile) asynhronously the application performs better than with the RollingFile but still bad. Running the same scenario with the FastRollingFile "immediateFlush" and "append" parameters ommitted in the configuration the FastRollingFile crashed due to a IllegalArgumentExeption in class ByteBuffer (stacktrace is lost).
The blocker priority may be at bit extremen but currently we are thinking of going back to log4j 1.2.17, throwing away several days of work and requiring a few days for adjusting our code. Hence this is a blocker for us as we could not release our software in the current state.