On Thu, 2005-01-20 at 10:12, Ceki Gülcü wrote: > Elias, > > Thank you for your proposal. Would it possible for you to backup your > proposal with prototype code? I'd be very interested in a comparative > study between an Appender using the current concurrency model and an > Appender following your improved model. A performance test involving > several threads would be quite appropriate. > > (Appender is an interface so one can easily swap implementations.) > > How does that sound?
Sounds fantastic.
I wrote some simple tests. Here's some of the code using the old interface and the new one.
ConcurrentAppender ca = new ConcurrentAppender() { protected void append(LoggingEvent event) { try { Thread.sleep(1); } catch (InterruptedException e) {} } protected void internalClose() {} };
AppenderSkeleton as = new AppenderSkeleton() { protected void append(LoggingEvent event) { try { Thread.sleep(1); } catch (InterruptedException e) {} } public void close() {} };
Using 5 threads with 100 log statements, here's what I got (a representative sample):
[java] AppenderSkeleton [java] Appender class org.apache.log4j.concurrent.PerformanceTest$2 [java] Took 1066ms for 100 logs * threads 5
[java] ConcurrentAppender [java] Appender class org.apache.log4j.concurrent.PerformanceTest$1 [java] Took 288ms for 100 logs * threads 5
It's about 3-4 times as efficient.
What do these results prove, that you can clobber the output 3-4 times faster than outputting logs correctly? There are two side to this problem, the performance side and the safety side. Maybe you have the required grasp of both sides of the issue, but without the source code for what you have done, it's hard to tell.
Is there some place to put a patch?
Please send your code to this mailing list.
I also wrote new ConsoleAppender, WriterAppender versions. They perform about the same as the old ones on a single processor machine, which isn't a surprising finding, but it's good to know they are just as fast.
One thing I noticed was that Layouts were not thread-safe. Most of the code time is spent in that class *. I changed the PatternConverter class signature from:
protected abstract void convert(LoggingEvent event);
to:
protected abstract void convert(StringBuffer sb, LoggingEvent event);
and rewrote Layout in such a way that it is thread safe. Obviously, there would be compatibility issues, which is why I am thinking of creating a separate Layout class to be used in this way.
* According to my last run 87% of the CPU time was spent formatting. So if you have 2 CPUs, 87% of the other CPU's time is waiting!
Since what you are doing is not clear, your enthusiasm, although nice to see, is hard to share.
-- Ceki G�lc�
The complete log4j manual: http://www.qos.ch/log4j/
--------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
