> On Mar 13, 2017, at 11:32 AM, Daniel Fuchs <[email protected]> wrote: > > Hi Ralph, > > On 13/03/2017 04:25, Ralph Goers wrote: >> Sorry for not getting back sooner but I finally found some time to follow up. >> >> I took a look at >> https://www.sitepoint.com/deep-dive-into-java-9s-stack-walking-api/ >> <https://www.sitepoint.com/deep-dive-into-java-9s-stack-walking-api/> and >> modified the benchmarks that were used there to add a few more use cases. I >> also created a small set of benchmarks for Java 8 to compare it against. >> The resulting project is at >> https://github.com/rgoers/stackwalker-vs-Reflection_getCallerClass/tree/master/java9 >> >> <https://github.com/rgoers/stackwalker-vs-Reflection_getCallerClass/tree/master/java9>. >> I’ve summarized the results in >> https://issues.apache.org/jira/browse/LOG4J2-1359 >> <https://issues.apache.org/jira/browse/LOG4J2-1359>, but for your >> convenience here are the bullet points: >> >> 1. Walking the Throwable StackTraceElements is significantly faster in Java >> 8 than Java 9. I am not sure what benchmark Brent used but my results differ. >> 2. Using StackWalker to get the StackTraceElements in Java 9 is almost twice >> as slow as walking the Throwable in Java 8. (Log4j relied on this pre-Java9 >> and apparently will have to continue to do so, but it will now be slower). >> 3. Using StackWalker to search for the caller's class is about twice as slow >> as sun.reflect.Reflection.getCallerClass() was (Log4j requires this and it >> is going to hurt performance). >> 4. sun.reflect.Reflection.getCallerClass is about 10 times faster than using >> StackWalker.getCallerClass to obtain the Class object of the immediate >> caller. >> In short it appears that the performance of StackWalker means that we are >> going to want to avoid using it. >> >> Ralph > > StackWalker can give you some performance improvements if > you walk the stack from i=0 to i=StackTraceElements[].length, > and stop somewhere before reaching the end. > > If you attempt to walk the stack backwards, starting from > i=StackTraceElements[].length and decreasing i, then I would > not expect as much improvement as you will need to walk the > whole stack anyway. > > best regards, > > — daniel
The “search” that uses StackWalker probably walks it to the end because I had to use test() to determine the StackFrame to return. If you can provide a better algorithm to use I’d be happy to retest it. If I=0 is the current StackFrame, then that is what we do. Please take a look at the stackwalkerSearch method at https://github.com/rgoers/stackwalker-vs-Reflection_getCallerClass/blob/master/java9/src/main/java/org/github/arnaudroger/StackWalkerGetCallerClass.java <https://github.com/rgoers/stackwalker-vs-Reflection_getCallerClass/blob/master/java9/src/main/java/org/github/arnaudroger/StackWalkerGetCallerClass.java>. Ralph > >> >> >> >>> On May 18, 2016, at 11:24 AM, Brent Christian <[email protected]> >>> wrote: >>> >>> Hi, >>> >>> I compared 8u65 and 9b116 performance on a simple Throwable.getStackTrace() >>> JMH benchmark that I have, using a variety of call stack depths. >>> Performance looks very similar between the two; if anything, 9b116 has a >>> slight edge for larger stack depths. >>> >>> So I don't think the difference is due to the walking of the stack itself, >>> at least based on what I measured. >>> >>> HTH, >>> -Brent >>> >>> On 5/10/16 9:49 AM, Ralph Goers wrote: >>>> I just ran one of the Log4j performance tests that specifically captures >>>> location information. To run the test I do >>>> >>>> java -jar log4j-perf/target/benchmarks.jar >>>> ".*AsyncAppenderLog4j2LocationBenchmark.*" -f 1 -wi 10 -i 20 -t 4 -si true >>>> >>>> And the results are: >>>> >>>> java version "1.7.0_80 >>>> >>>> Benchmark Mode >>>> Samples Score Error Units >>>> o.a.l.l.p.j.AsyncAppenderLog4j2LocationBenchmark.throughputSimple thrpt >>>> 20 124819.285 ± 3003.918 ops/s >>>> >>>> java version "1.8.0_65" >>>> >>>> Benchmark Mode >>>> Samples Score Error Units >>>> o.a.l.l.p.j.AsyncAppenderLog4j2LocationBenchmark.throughputSimple thrpt >>>> 20 123209.746 ± 3064.672 ops/s >>>> >>>> >>>> java version "9-ea" >>>> Java(TM) SE Runtime Environment (build 9-ea+116) >>>> >>>> Benchmark Mode >>>> Samples Score Error Units >>>> o.a.l.l.p.j.AsyncAppenderLog4j2LocationBenchmark.throughputSimple thrpt >>>> 20 96090.261 ± 4565.763 ops/s >>>> >>>> >>>> This tells me that Java 9 is about 23% slower than previous versions in >>>> walking the stack trace elements. >>>> >>>> Ralph >>> >> > >
