300 likes | 439 Vues
Wake Up and Smell the Coffee: Performance Analysis Methodologies for the 21st Century . Kathryn S McKinley Department of Computer Sciences University of Texas at Austin. Shocking News!. In 2000, Java overtook C and C++ as the most popular programming language [TIOBE 2000--2008].
E N D
Wake Up and Smell the Coffee: Performance Analysis Methodologies for the 21st Century Kathryn S McKinley Department of Computer Sciences University of Texas at Austin
Shocking News! In 2000, Java overtook C and C++ as the most popular programming language [TIOBE 2000--2008]
Systems Researchin Industry and Academia ISCA 2006 20 papers use C and/or C++ 5 papers are orthogonal to the programming language 2 papers use specialized programming languages 2 papers use Java and C from SPEC 1 paper uses only Java from SPEC
What is Experimental Computer Science? • An idea • An implementation in some system • An evaluation
The success of most systems innovation hinges on evaluation methodologies. Benchmarks reflect current and ideally, future reality Experimental design is appropriate Statistical data analysis
The success of most systems innovation hinges on experimental methodologies. Benchmarks reflect current and ideally, future reality [DaCapo Benchmarks 2006] Experimental design is appropriate. Statistical Data Analysis [Georges et al. 2006]
Experimental Design • We’re not in Kansas anymore! • JIT compilation, GC, dynamic checks, etc • Methodology has not adapted • Needs to be updated and institutionalized “…this sophistication provides a significant challenge to understanding complete system performance, not found in traditional languages such as C or C++” [Hauswirth et al OOPSLA ’04]
Experimental Design • Comprehensive comparison • 3 state-of-the-art JVMs • Best of 5 executions • 19 benchmarks • Platform: 2GHz Pentium-M, 1GB RAM, linux 2.6.15
Experimental Design First Iteration Second Iteration Third Iteration
Experimental Design Another Experiment • Compare two garbage collectors • Semispace Full Heap Garbage Collector • Marksweep Full Heap Garbage Collector
Experimental Design Another Experiment • Compare two garbage collectors • Semispace Full Heap Garbage Collector • Marksweep Full Heap Garbage Collector • Experimental Design • Same JVM, same compiler settings • Second iteration for both • Best of 5 executions • One benchmark - SPEC 209_db • Platform: 2GHz Pentium-M, 1GB RAM, linux 2.6.15
Marksweep vs Semispace Semispace Marksweep
Experimental Design:Best Practices • Measuring JVM innovations • Measuring JIT innovations • Measuring GC innovations • Measuring Architecture innovations
JVM InnovationBest Practices • Examples: • Thread scheduling • Performance monitoring • Workload triggers differences • real workloads & perhaps microbenchmarks • e.g., force frequency of thread switching • Measure & report multiple iterations • start up • steady state (aka server mode) • never configure the VM to use completely unoptimized code! • Use a modest or multiple heap sizes computed as a function of maximum live size of the application • Use & report multiple architectures
Best Practices Pentium M AMD Athlon SPARC
JIT Innovation Best Practices Example: new compiler optimization • Code quality: Does it improve the application code? • Compile time: How much compile time does it add? • Total time: compiler and application time together • Problem: adaptive compilation responds to compilation load • Question: How do we tease all these effects apart?
JIT Innovation Best Practices Teasing apart compile time and code quality requires multiple experiments • Total time: Mix methodology • Run adaptive system as intended • Result: mixture of optimized and unoptimized code • First & second iterations (that include compile time) • Set and/or report the heap size as a function of maximum live size of the application • Report: average and show statistical error • Code quality • OK: Run iterations until performance stabilizes on “best”, or • Better: Run several iterations of the benchmark, turn off the compiler, and measure a run guaranteed to have no compilation • Best:Replay mix compilation • Compile time • Requires the compiler to be deterministic • Replay mix compilation
Replay Compilation Force the JIT to produce a deterministic result • Make a compilation profiler & replayer Profiler • Profile first or later iterations with adaptive JIT, pick best or average • Record profiling information used in compilation decisions, e.g., dynamic profiles of edges, paths, &/or dynamic call graph • Record compilation decisions, e.g., compile method bar at level two, inline method foo into bar • Mix of optimized and unoptimized, or all optimized/unoptimized Replayer • Reads in profile • As the system loads each class, apply profile +/- innovation • Result • controlled experiments with deterministic compiler behavior • reduces statistical variance in measurements • Still not a perfect methodology for inlining
GC Innovation Best Practices • Requires more than one experiment... • Use & report a range of fixed heap sizes • Explore the space time tradeoff • Measure heap size with respect to the maximum live size of the application • VMs should report total memory not just application memory • Different GC algorithms vary in the meta-data they require • JIT and VM use memory... • Measure time with a constant workload • Do not measure through put • Best: run two experiments • mix with adaptive methodology: what users are likely to see in practice • replay: hold the compiler activity constant • Choose a profile with “best” application performance in order to keep from hiding mutator overheads in bad code.
Architecture Innovation Best Practices • Requires more than one experiment... • Use more than one VM • Set a modest heap size and/or report heap size as a function of maximum live size • Use a mixture of optimized and uncompiled code • Simulator needs the “same” code in many cases to perform comparisons • Best for microarchitecture only changes: • Multiple traces from live system with adaptive methodology • start up and steady state with compiler turned off • what users are likely to see in practice • Wont work if architecture change requires recompilation, e.g., new sampling mechanism • Use replay to make the code as similar as possible
“…improves throughput by up to 41x” “speed up by 10-25% in many cases…” “…about 2x in two cases…” “…more than 10x in two small benchmarks” “speedups of 1.2x to 6.4x on a variety of benchmarks” “can reduce garbage collection time by 50% to 75%” “…demonstrating high efficiency and scalability” “our prototype has usable performance” “sometimes more than twice as fast” “our algorithm is highly efficient” “garbage collection degrades performance by 70%” “speedups…. are very significant (up to 54-fold)” “our …. is better or almost as good as …. across the board” “the overhead …. is on average negligible” Quotes from recent research papers statistics Disraeli benchmarks There are lies, damn lies, and
Conclusions • Methodology includes • Benchmarks • Experimental design • Statistical analysis [OOPSLA 2007] • Poor Methodology • can focus or misdirect innovation and energy • We have a unique opportunity • Transactional memory, multicore performance, dynamic languages, • What we can do • Enlist VM builders to include replay • Fund and broaden participation in benchmarking • Research and industrial partnerships • Funding through NSF, ACM, SPEC, industry or ?? • Participate in building community workloads
Thank you! www.dacapobench.org