jvm goes big_data_sfjava
Post on 15-Jan-2015
2.550 Views
Preview:
DESCRIPTION
TRANSCRIPT
JVM goes BigData
srisatish.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 Ahead
synchronizedNonblocking 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
synchronized
under 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: Amdahl's > Moore's!
State, Actions – key/value pairs!get, put, delete, _resize
ByteArray to hold DataConcurrent writes: using CAS
No locks, no volatileMuch faster than locking under heavy load
Directly reach main data array in 1 step
Resize as neededCopy Array to a larger Array on demand. Post updates
Death & Taxes: Java Overheads!
• Cost of an 8char String?
• Cost of 100entry TreeMap<Double,Double> ?
8bhdr
12bfields
4bptr
4bpad
8bhdr
4blen
16bdata
A: 56 bytes, or a 7x blowup
48bTreeMap
40bTreeMap$Entry
16bDouble
16bDouble
A: 7248 bytes or a ~5x blowup
yourkit: memory profile
Which collection: Mozart or Bach?
Concurrency: Nonblocking HashMap Google Collections
Overheads Watch out for perelement costs! Primitives can be hard to manage!
Sparse collections Average collection size in enterprise is ~3
java.io.Serializable is S.L..O.…WTrue to platform Use “transient” ObjectSerialField[] Avro Google Protocol Buffers, Externalizable + byte[] Roll your own
serializable
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 faster
http://github.com/cowtowncoder/javauuidgenerator
http://jug.safehaus.org/
http://johannburkard.de/blog/programming/java/JavaUUIDgeneratorscompared.htm
/**
* 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));
}
Leachsalz UUID
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
summary
TimebasedUUIDs 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 couldn't 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:gc
GC 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
edensurvivor spaces
new Object()survivor ratio
jvm allocates
TenuringThreshold
promotion
old gen
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/tool
Full 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 CMS
Don’t promote too often! Frequent promotion causes fragmentation
(avoid never tenure) TenuringThreshold
Size the generations Min GC times are a function of Live Set
Old Gen should host steady state comfortably
Avoid CMS Initiating heuristic XX:+UseCMSInitiationOccupanyOnly
Use Concurrent for System.gc() XX:+ExplicitGCInvokesConcurrent
GC Threads
Parallelize 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; }
Fragmentation
Performance 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 promotion
Small objects still can fit in the holes!Compaction – stop the world.Unsolved on Oracle/Sun Hotspot Azul Systems Pauseless JVM.
JRockit Mission Control
Example
Application suddenly transitions to backtoback full gcs.
Cannot use free mem – too many holes!
Tools
• GCHisto• jconsole• VisualVM/VisualGC• Logs• Thread dumps• yourkit memory profile, snapshots
GCSpy
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 jdk16u21
Workaround:XX:MaxDirectMemorySize=<size> Manually Trigger System.gc() to avoid “leak”
Virtually there!
Ballooning driver for Memory: Disable it!Time (TSC) issue! It's 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
References
Chris Wimmer, Chris Wimmer, http://wikis.sun.com/display/HotSpotInternals/Synchronizationhttp://wikis.sun.com/display/HotSpotInternals/SynchronizationRussel & Detlefs Russel & Detlefs http://www.oracle.com/technetwork/java/biasedlockingoopsla2006wp149958.pdfhttp://www.oracle.com/technetwork/java/biasedlockingoopsla2006wp149958.pdfGoogle Protocol Buffers Google Protocol Buffers http://code.google.com/p/protobufhttp://code.google.com/p/protobufThrift Thrift http://incubator.apache.org/thrift/static/thrift20070401.pdfhttp://incubator.apache.org/thrift/static/thrift20070401.pdfLeachSalz Variant of UUID LeachSalz Variant of UUID http://www.upnp.org/resources/draftleachuuidsguids00.txthttp://www.upnp.org/resources/draftleachuuidsguids00.txtHans Boehm, Hans Boehm, http://www.hpl.hp.com/personal/Hans_Boehm/gc/complexity.htmlhttp://www.hpl.hp.com/personal/Hans_Boehm/gc/complexity.htmlBrian Goetz, JSR133 Brian Goetz, JSR133 http://www.ibm.com/developerworks/java/library/jjtp03304/http://www.ibm.com/developerworks/java/library/jjtp03304/GCSpy GCSpy http://www.cs.kent.ac.uk/projects/gc/gcspy/http://www.cs.kent.ac.uk/projects/gc/gcspy/Understanding GC logs Understanding GC logs http://blogs.sun.com/poonam/entry/understanding_cms_gc_logshttp://blogs.sun.com/poonam/entry/understanding_cms_gc_logs
Cliff Click's http://sourceforge.net/projects/highscalelib/Cliff Click's http://sourceforge.net/projects/highscalelib/
top related