Jvm goes big_data_sfjava

JVM goes BigDatasrisatish.ambati AT gmail.comDataStax/OpenJDK4/12/2011@srisatish

Motivation• A compendium of recent jvm scale issues while working with big data.• This talk will not have details on big data.• Thanks Sasa!

Trail AheadsynchronizedNonblocking Hashmap A state transition viewCollectionsSerializationUUIDGarbage Collection The free parameters! Generations, Promotion, Fragmentation OffheapQuestions & asynchronous IO

tools of trade • What the JVM is doing: – dtrace, hprof, introscope, jconsole, visualvm, yourkit, gchisto, zvision • Invasive JVM observation tools: – bci, jvmti, jvmdi/pi agents, logging • What the OS is doing: – dtrace, oprofile, vtune, perf • What the network/disk is doing: – ganglia, iostat, lsof, nagios, netstat, tcpdump

synchronizedunder the hood – Fast path for nocontention thin lock – Bias threads to lock or bulk revoke bias – Store free biasing

JMM: happensbefore, causalityPartial ordervolatilePiggybackingFutureTaskBlockingQueuejsr133

* Java Concurrency in Practice, Brian Goetz

java.util.concurrent also holds locks!

Tomcat under concurrent load!

Nonblocking collections: Amdahls > Moores!State, Actions – key/value pairs! get, put, delete, _resizeByteArray to hold DataConcurrent writes: using CAS No locks, no volatile Much faster than locking under heavy load Directly reach main data array in 1 stepResize as needed Copy Array to a larger Array on demand. Post updates

Death & Taxes: Java Overheads! • Cost of an 8char String? 8b 12b 4b hdr fields ptr A: 56 bytes, or a 7x blowup 8b 4b 16b 4b hdr len data pad • Cost of 100entry TreeMap<Double,Double> ? 48b TreeMap 40b TreeMap$Entry 16b 16b A: 7248 bytes or a ~5x blowup Double Double

yourkit: memory profile

Which collection: Mozart or Bach?Concurrency: Nonblocking HashMap Google CollectionsOverheads Watch out for perelement costs! Primitives can be hard to manage!Sparse collections Average collection size in enterprise is ~3

serializable  java.io.Serializable is S.L..O.…W True to platform  Use “transient”  ObjectSerialField[]  Avro  Google Protocol Buffers,  Externalizable + byte[]  Roll your own

ser+deser smaller is better https://github.com/eishay/jvmserializers.git

avro• Schema – No per datum overheads – Optional code gen• Types are runtime• Untagged data• No manuallyassigned field IdsCons:• Schema mismatches• Runtime only checks

googleprotobuffer• Define message format in .proto file• All data in key/value pairs• Generate sources• .builder for each class with getter/setter

thrift• Type, Transport, Protocol, Version, Processors• Separation of structure from protocol & transport• TCompactProtocol, etc – tag/data, compression• TSocket, TfileTransport, etc• colocated clients & servers

UUIDjava.util.UUID is slow ● dominated by sha_transform costs ● Leachsalz (128bit) Turns out that default PRNG (via SecureRandom)Uses /dev/urandom for seed initialization Djava.security.egd=file:/dev/urandom ● PRNG without file is atleast 20%40% better.Use TimeUUIDs where possible – much faster Alternatives: JUG – java.uuid.generator, com.eaio.uuid ~10x fasterhttp://github.com/cowtowncoder/javauuidgenerator http://jug.safehaus.org/ http://johannburkard.de/blog/programming/java/JavaUUIDgeneratorscompared.htm

Leachsalz UUID /** * Returns a {@code String} object representing this {@code UUID}. * * <p> The UUID string representation is as described by this BNF: * <blockquote><pre> * {@code * UUID = <time_low> "-" <time_mid> "-" * <time_high_and_version> "-" * <variant_and_sequence> "-" * <node> * time_low = 4*<hexOctet> * time_mid = 2*<hexOctet> * time_high_and_version = 2*<hexOctet> * variant_and_sequence = 2*<hexOctet> * node = 6*<hexOctet> * hexOctet = <hexDigit><hexDigit> * hexDigit = * "0" | "1" | "2" | "3" | "4" | "5" | "6" | "7" | "8" | "9" * | "a" | "b" | "c" | "d" | "e" | "f" * | "A" | "B" | "C" | "D" | "E" | "F" * }</pre></blockquote> * * @return A string representation of this {@code UUID} */ public String toString() { return (digits(mostSigBits >> 32, 8) + "-" + digits(mostSigBits >> 16, 4) + "-" + digits(mostSigBits, 4) + "-" + digits(leastSigBits >> 48, 4) + "-" + digits(leastSigBits, 12)); }

PerfTop: 1485 irqs/sec kernel:18.6% exact: 0.0% [1000Hz cycles], (all, 8 CPUs) samples pcnt function DSO _______ _____ ________________________________________________________________ 1882.00 26.3% intel_idle [kernel.kallsyms] 1678.00 23.5% os::javaTimeMillis() libjvm.so 382.00 5.3% SpinPause libjvm.so 335.00 4.7% Timer::ImplTimerCallbackProc() libvcllx.so 291.00 4.1% gettimeofday /lib/libc2.12.1.so 268.00 3.7% hpet_next_event [kernel.kallsyms] 254.00 3.6% ParallelTaskTerminator::offer_termination(TerminatorTerminator*) libjvm.so PerfTop: 1656 irqs/sec kernel:59.5% exact: 0.0% [1000Hz cycles], (all, 8 CPUs) samples pcnt function DSO _______ _____ ________________________________________________________________ 6980.00 38.5% sha_transform [kernel.kallsyms] 2119.00 11.7% intel_idle [kernel.kallsyms] 1382.00 7.6% mix_pool_bytes_extract [kernel.kallsyms] 437.00 2.4% i8042_interrupt [kernel.kallsyms] 416.00 2.3% hpet_next_event [kernel.kallsyms] 390.00 2.2% extract_buf [kernel.kallsyms] 376.00 2.1% ThreadInVMfromNative::~ThreadInVMfromNative() libjvm.so 321.00 1.8% T.3542 libjvm.so 298.00 1.6% __ticket_spin_lock [kernel.kallsyms] 296.00 1.6% Timer::ImplTimerCallbackProc() libvcllx.so 255.00 1.4% Unsafe_GetInt libjvm.so

summaryTimebasedUUIDs vs. UUIDsuse ~4 times less kernel time on creation!No SHA library calls!optimized toString()Much faster than standard java.util.UUID Better Instructions per clocks as well. If on EC2: Watch out for noncacheable file access to /dev/urandom!

String theory of Java!byte[] vs. char[]If ver > jdk16u21 try XX:+UseCompressedStringsAppend performance (gc) differs: Strings vs. StringBufferscom.google.common.base.Joiner • Join text for cheap, • skipNulls or useForNulls()com.google.common.base.Splitter

“Null References: A billion dollar mistake” C.A.R Hoare“I call it my billiondollar mistake. It was the invention of the null reference in 1965. At that time, I was designing the first comprehensive type system for references in an object oriented language (ALGOL W). My goal was to ensure that all use of references should be absolutely safe, with checking performed automatically by the compiler. But I couldnt resist the temptation to put in a null reference, simply because it was so easy to implement. This has led to innumerable errors, vulnerabilities, and system crashes, which have probably caused a billion dollars of pain and damage in the last forty years.” qconlondon, 09

Best Practices: Garbage Collection

verbose:gcGC Logs are cheap even in production Xloggc:gc.log XX:+PrintGCDetails XX:+PrintGCTimeStamps XX:+PrintTenuringDistribution A bit expensive/obscure ones: XX:PrintFLSStatistics=2 XX:CMSStatistics=1 XX:CMSInitiationStatistics XX:+PrintFLSCensus

Three free parameters Allocation Rate: your workload!Size: defines runway! Live Set, memoryPause times: Stoppages!

Four free parameters Allocation Rate: your application load!Size: defines runway! Live Set, system memoryPause times: Stoppages! (fourth: Overheads of GC – Space & CPU.)

Part I: Sizingto be Xmx == Xms or not?Young generation:Use Xmn for predictable performance new Object() survivor ratio eden survivor spaces promotion Tenuring Threshold old gen jvm allocates

Part II: Pick a collector!Serial GC – Serial new + Serial OldParallel GC (default) Parallel Scavenge + Serial OldUseParallelOldGC : Parallel Scavenge + Parallel OldUseConcurrentMarkSweep: ParNew, CMS Old, Serial OldG1/Experimental

Reading GC logs – a topic/toolFull GC is STWInitial Mark, Rescan/WeakRef/Remark are STWLook for promotion failuresLook for concurrent mode failures

... 995.330: [CMSconcurrentmark: 0.952/1.102 secs] [Times: user=3.69 sys=0.54, real=1.10 secs] 995.330: [CMSconcurrentprecleanstart]995.618: [CMSconcurrentpreclean: 0.279/0.287 secs] [Times: user=0.90 sys=0.20, real=0.29 secs] 995.618: [CMSconcurrentabortableprecleanstart]995.695: [GC 995.695: [ParNew (promotion failed)Desired survivor size 41943040 bytes, new threshold 1 (max 1) age 1: 29826872 bytes, 29826872 total: 720596K>703760K(737280K), 0.4710410 secs]996.166: [CMS996.317: [CMSconcurrentabortablepreclean: 0.218/0.699 secs] [Times: user=1.39 sys=0.10, real=0.70 secs] (concurrent mode failure): 4100132K>784070K(5341184K), 4.7478300 secs] 4780154K>784070K(6078464K), [CMS Perm : 17033K>17014K(28400K)], 5.2191410 secs] [Times: user=5.70 sys=0.01, real=5.22 secs]...

Tuning CMSDon’t promote too often! Frequent promotion causes fragmentation (avoid never tenure) TenuringThresholdSize the generations Min GC times are a function of Live Set Old Gen should host steady state comfortablyAvoid CMS Initiating heuristic XX:+UseCMSInitiationOccupanyOnlyUse Concurrent for System.gc() XX:+ExplicitGCInvokesConcurrent

GC ThreadsParallelize on multicores XX:ParallelGCThreads=4 (default: derived from # of cpus on system) *8 + (n5)/8 XX:ParallelCMSThreads=4 (default: derived from # of parallelgcthreads)Strategy A: Tune min gcs & let appl data in eden

Did someone ask about defaults? if (FLAG_IS_DEFAULT(ParallelGCThreads)) { assert(ParallelGCThreads == 0, "Default ParallelGCThreads is not 0"); // For very large machines, there are diminishing returns // for large numbers of worker threads. Instead of // hogging the whole system, use a fraction of the workers for every // processor after the first 8. For example, on a 72 cpu machine // and a chosen fraction of 5/8 // use 8 + (72 8) * (5/8) == 48 worker threads. unsigned int ncpus = (unsigned int) os::active_processor_count(); return (ncpus <= switch_pt) ? ncpus : (switch_pt + ((ncpus switch_pt) * num) / den); } else { return ParallelGCThreads; }

FragmentationPerformance degrades over timeInducing “Full GC” makes problem go awayFree memory that cannot be used Round off errorsReduce occurrenceUse a compacting collectorPromote less oftenUse uniform sized objects

Not enough large contiguous space for promotionSmall objects still can fit in the holes!Compaction – stop the world.Unsolved on Oracle/Sun Hotspot Azul Systems Pauseless JVM.

JRockit Mission Control

ExampleApplication suddenly transitions to back toback full gcs.Cannot use free mem – too many holes!

Tools• GCHisto• jconsole• VisualVM/VisualGC• Logs• Thread dumps• yourkit memory profile, snapshots

Gone 0xff the heap !!ByteBuffer.allocateDirect(16 * 1024 * 1024)Also can be mapped memory of a file regionStore longlived objects outside jvm Managed by native i/o ops.JNA: dynamically load & call native libraries without compile time decl like JNIWorks for limited use cases in the lab. Ex: Terracotta, Hbase, Cassandra

Gone 0xff the heap ?Issues to consider:No clear api to deallocate from this region ● See jbellis patch to JNA179 for FreeableBufferObject cleanup relegated to finalization Single finalizer thread, Bug ID: 4469299Behind WeakReference processing in jdk16u21Workaround:XX:MaxDirectMemorySize=<size> Manually Trigger System.gc() to avoid “leak”

Virtually there! Ballooning driver for Memory: Disable it!Time (TSC) issue! Its relative!Scheduling when # of threads > # of vcpus.. Tickless _nohz kernelGC Thread starvation = STW pauseslarge ec2 instances are not all equal..DirectPathIO & vtd, rvi – Watch out for Sockets!Tools: Performance counters still not virtualized!

summary• JVM is still the most popular platform for deployment for the new languages!• JVM heartburn around scale! – Serialization – UUID – Object overhead – Garbage Collection – Hypervisor

ReferencesChris Wimmer, http://wikis.sun.com/display/HotSpotInternals/SynchronizationRussel & Detlefs http://www.oracle.com/technetwork/java/biasedlockingoopsla2006wp149958.pdfGoogle Protocol Buffers http://code.google.com/p/protobufThrift http://incubator.apache.org/thrift/static/thrift20070401.pdfLeachSalz Variant of UUID http://www.upnp.org/resources/draftleachuuidsguids00.txtHans Boehm, http://www.hpl.hp.com/personal/Hans_Boehm/gc/complexity.htmlBrian Goetz, JSR133 http://www.ibm.com/developerworks/java/library/jjtp03304/GCSpy http://www.cs.kent.ac.uk/projects/gc/gcspy/Understanding GC logs http://blogs.sun.com/poonam/entry/understanding_cms_gc_logsCliff Clicks http://sourceforge.net/projects/highscalelib/

Jvm goes big_data_sfjava

by srisatish ambati

Accessibility

Categories

Upload Details

Usage Rights

Statistics

Jvm goes big_data_sfjava Presentation Transcript