Xprof, the forgotten HotSpot profiler
Written on
Did you know that every HotSpot-based JVM includes a profiler? You can activate it on the command line with -Xprof and it dumps its output to stdout.
The information about it is pretty sparse, so I decided to write a short blog post about it.
Getting Started
Let’s dive right in with an example:
public class Fibonacci {
private static long fib(long n) {
if (n <= 0) {
throw new IllegalArgumentException("n must be a natural number");
} else if (n == 1 || n == 2) {
return 1;
} else {
return fib(n - 2) + fib(n - 1);
}
}
public static void main(String[] args) {
System.out.println(fib(50));
}
}
If you compile and run this code with java Fibonacci -Xprof you’ll get the following output after the program has finished:
Flat profile of 45.00 secs (3664 total ticks): main
Compiled + native Method
100.0% 3664 + 0 Fibonacci.fib
100.0% 3664 + 0 Total compiled
Flat profile of 0.00 secs (1 total ticks): DestroyJavaVM
Thread-local ticks:
100.0% 1 Unknown: no last frame
This output reveals already a few things about the profiler:
- Why is
main()not showing up in the profile? The reason is thatXprofis a “flat” profiler. We are normally used to hierarchical profilers that show complete stack traces. However, this profiler will only show the top-most method to avoid the overhead of stack-walking. As this example is only very simple, this is not immediately clear, but we’ll see this in a more complex example later. - It is a sampling profiler. Samples are captured at each “tick”, and a tick occurs every 10 milliseconds (see the creation of FlatProfilerTask and the definition of the delay interval in the JDK sources)
- It prints the captured information only when a thread exits (see
before_exit(JavaThread* thread)in the JDK sources)
If you pay close attention, you’ll see that the number of ticks does not match the reported time period. For 45 seconds, it should capture 4,500 samples, given a sampling frequency of 10 milliseconds. A quick experiment (just running Thread#sleep()) shows that the sampling interval (on my machine) is actually 12 milliseconds.
| Sleep duration [ms] | Total samples |
|---|---|
| 5,000 | 421 |
| 10,000 | 830 |
| 20,000 | 1,665 |
| 40,000 | 3,328 |
Before we move on, let’s see what additional information we can get from Xprof.
We can see more information about JVM threads with -XX:+ProfileVM:
Flat profile of 0.00 secs (3665 total ticks): VM Thread
Compiled + native Method
100.0% 0 + 3665 __psynch_cvwait
100.0% 0 + 3665 Total compiled
We can see the following with -XX:+ProfilerPrintByteCodeStatistics (this will get more interesting with a complex example):
Bytecode ticks:
Global summary of 45.00 seconds:
100.0% 3665 Received ticks
0.0% 1 Unknown code
And finally we can get even low-level info about hot methods with -XX:+ProfilerRecordPC:
Printing compiled methods with PC buckets having more than 15 ticks
===================================================================
[CodeBlob (0x000000010c4b7890)]
Framesize: 8
Could not load hsdis-amd64.dylib; library not loadable; PrintAssembly is disabled
Apparently we need the Java disasssembler library hsdis, which you need to build yourself (see the build instructions for your platform in the JITWatch wiki). With: -XX:+ProfilerRecordPC and export LD_LIBRARY_PATH=/path/to/hsdis we get indeed some assembly. If you prefer Intel assembly syntax instead of AT&T, just add -XX:+UnlockDiagnosticVMOptions -XX:PrintAssemblyOptions=intel.
Here is a shortened version of the output:
Printing compiled methods with PC buckets having more than 15 ticks
===================================================================
[CodeBlob (0x000000010bf09650)]
Framesize: 8
Loaded disassembler from hsdis-amd64.dylib
Decoding compiled method 0x000000010bf09650:
Code:
[Disassembling for mach='i386:x86-64']
[Entry Point]
[Verified Entry Point]
[Constants]
# {method} {0x0000000124b60b70} 'fib' '(J)J' in 'Fibonacci'
# parm0: rsi:rsi = long
# [sp+0x40] (sp of caller)
0x000000010bf097c0: mov DWORD PTR [rsp-0x14000],eax
0x000000010bf097c7: push rbp
0x000000010bf097c8: sub rsp,0x30
0x000000010bf097cc: mov r10,rsi
0x000000010bf097cf: test rsi,rsi
0x000000010bf097d2: jle 0x000000010bf0989e
0x000000010bf097d8: cmp rsi,0x1
0x000000010bf097dc: je 0x000000010bf097e4
0x000000010bf097de: cmp rsi,0x2
0x000000010bf097e2: jne 0x000000010bf09844
0x000000010bf097e4: mov eax,0x1
0x000000010bf097e9: jmp 0x000000010bf09838
0x000000010bf097eb: mov eax,0x1
0x000000010bf097f0: jmp 0x000000010bf09833
0x000000010bf097f2: mov r11d,0x1
0x000000010bf097f8: mov QWORD PTR [rsp+0x10],r11
0x000000010bf097fd: dec r10
0x000000010bf09800: test r10,r10
0x000000010bf09803: jle 0x000000010bf098f9
0x000000010bf09809: cmp r10,0x1
0x000000010bf0980d: je 0x000000010bf097eb
0x000000010bf0980f: cmp r10,0x2
0x000000010bf09813: je 0x000000010bf097eb
0x000000010bf09815: mov rbp,r10
0x000000010bf09818: data32 xchg ax,ax
0x000000010bf0981b: call 0x000000010bf097c0 ; OopMap{off=96}
;*invokestatic fib
; - Fibonacci::fib@37 (line 10)
; - Fibonacci::fib@43 (line 10)
; {static_call}
0x000000010bf09820: mov rbp,rax
0x000000010bf09823: mov rsi,QWORD PTR [rsp+0x8]
0x000000010bf09828: data32 xchg ax,ax
0x000000010bf0982b: call 0x000000010bf097c0 ; OopMap{off=112}
;*invokestatic fib
; - Fibonacci::fib@43 (line 10)
; - Fibonacci::fib@43 (line 10)
; {static_call}
36.1% [1321]
0x000000010bf09830: add rax,rbp
0x000000010bf09833: add rax,QWORD PTR [rsp+0x10]
0x000000010bf09838: add rsp,0x30
0x000000010bf0983c: pop rbp
0x000000010bf0983d: test DWORD PTR [rip+0xfffffffffded97bd],eax # 0x0000000109de3000
; {poll_return}
27.0% [989]
0x000000010bf09843: ret
0x000000010bf09844: mov r11,rsi
0x000000010bf09847: add r11,0xfffffffffffffffe
0x000000010bf0984b: mov QWORD PTR [rsp+0x8],r11
0x000000010bf09850: test r11,r11
0x000000010bf09853: jle 0x000000010bf098c9
0x000000010bf09855: add rsi,0xfffffffffffffffd
0x000000010bf09859: cmp r11,0x1
0x000000010bf0985d: je 0x000000010bf097f2
0x000000010bf0985f: cmp r11,0x2
0x000000010bf09863: je 0x000000010bf097f2
0x000000010bf09865: mov QWORD PTR [rsp],rsi
0x000000010bf09869: mov rbp,r10
0x000000010bf0986c: mov rsi,r10
0x000000010bf0986f: add rsi,0xfffffffffffffffc
0x000000010bf09873: call 0x000000010bf097c0 ; OopMap{off=184}
;*invokestatic fib
; - Fibonacci::fib@37 (line 10)
; - Fibonacci::fib@37 (line 10)
; {static_call}
0x000000010bf09878: mov QWORD PTR [rsp+0x10],rax
0x000000010bf0987d: mov rsi,QWORD PTR [rsp]
[...]
A more complex example
For this example I have started Elasticsearch 5.5.0 with -Xprof -XX:+ProfileVM -XX:+ProfilerPrintByteCodeStatistics and ran a macrobenchmark that stresses indexing. Here is one example:
Flat profile of 573.60 secs (31156 total ticks): elasticsearch[vWpZ2JV][bulk][T#2]
Interpreted + native Method
0.0% 0 + 10 java.lang.Thread.yield
0.0% 9 + 0 org.apache.lucene.index.DefaultIndexingChain$PerField.invert
0.0% 0 + 6 sun.nio.ch.FileDispatcherImpl.force0
0.0% 5 + 0 org.apache.lucene.document.Field.tokenStream
0.0% 4 + 0 org.elasticsearch.index.shard.IndexShard.shouldFlush
0.0% 4 + 0 org.elasticsearch.common.util.concurrent.KeyedLock$ReleasableLock.<init>
0.0% 3 + 0 org.apache.lucene.util.BytesRef.<init>
0.0% 3 + 0 org.apache.lucene.index.DefaultIndexingChain.getOrAddField
0.0% 3 + 0 org.apache.lucene.index.DefaultIndexingChain.processDocument
0.0% 3 + 0 java.util.function.BinaryOperator$$Lambda$1634.<init>
0.0% 3 + 0 org.elasticsearch.action.support.replication.TransportWriteAction$WritePrimaryResult.respond
0.0% 0 + 2 sun.nio.ch.FileChannelImpl.unmap0
0.0% 2 + 0 org.elasticsearch.indices.IndexingMemoryController.getIndexBufferRAMBytesUsed
0.0% 2 + 0 org.apache.lucene.codecs.blocktree.BlockTreeTermsWriter$PendingTerm.<init>
0.0% 2 + 0 org.apache.lucene.index.FreqProxTermsWriterPerField$FreqProxPostingsArray.<init>
0.0% 2 + 0 java.util.concurrent.locks.AbstractQueuedSynchronizer.unparkSuccessor
0.0% 2 + 0 org.elasticsearch.action.bulk.BulkResponse.toXContent
0.0% 2 + 0 org.elasticsearch.common.util.concurrent.AsyncIOProcessor.put
0.0% 2 + 0 org.elasticsearch.index.mapper.FieldNamesFieldMapper.parseCreateField
0.0% 2 + 0 java.util.concurrent.ThreadPoolExecutor.runWorker
0.0% 2 + 0 org.elasticsearch.index.mapper.BaseGeoPointFieldMapper.parse
0.0% 2 + 0 org.apache.lucene.index.DefaultIndexingChain.writeDocValues
0.0% 0 + 1 sun.nio.fs.UnixNativeDispatcher.open0
0.0% 1 + 0 org.apache.lucene.util.packed.PackedInts$Format.longCount
0.0% 1 + 0 org.elasticsearch.index.mapper.DocumentParser.parsedDocument
0.4% 101 + 22 Total interpreted (including elided)
Compiled + native Method
6.0% 1695 + 1 org.apache.lucene.index.DefaultIndexingChain.processField
5.2% 1476 + 0 org.apache.lucene.index.TermsHashPerField.add
5.1% 1460 + 0 org.apache.lucene.util.BytesRefHash.add
4.6% 1308 + 0 org.apache.lucene.index.DefaultIndexingChain$PerField.invert
2.2% 624 + 0 org.apache.lucene.codecs.compressing.LZ4.compress
1.8% 525 + 0 org.apache.lucene.document.Field$StringTokenStream.incrementToken
1.7% 478 + 0 org.apache.lucene.index.DefaultIndexingChain.processDocument
1.3% 358 + 0 org.apache.lucene.util.BytesRefHash.findHash
1.2% 335 + 20 org.elasticsearch.index.mapper.FieldNamesFieldMapper.parseCreateField
1.1% 324 + 0 org.apache.lucene.util.MSBRadixSorter.radixSort
1.1% 309 + 0 org.apache.lucene.util.IntroSorter.quicksort
1.1% 302 + 3 org.apache.lucene.codecs.PushPostingsWriterBase.writeTerm
1.0% 280 + 2 org.apache.lucene.util.BytesRefHash.rehash
1.0% 273 + 0 org.apache.lucene.analysis.standard.StandardTokenizerImpl.getNextToken
0.9% 242 + 0 org.apache.lucene.index.TermVectorsConsumerPerField.start
0.8% 216 + 0 org.elasticsearch.index.mapper.DocumentParser.internalParseDocument
0.8% 214 + 0 org.apache.lucene.analysis.standard.StandardTokenizer.incrementToken
0.7% 202 + 0 org.apache.lucene.analysis.FilteringTokenFilter.incrementToken
0.7% 194 + 8 org.elasticsearch.index.mapper.FieldMapper.parse
0.7% 195 + 1 org.apache.lucene.document.Field.tokenStream
0.7% 161 + 32 org.apache.lucene.index.DocumentsWriter.updateDocument
0.6% 176 + 0 org.apache.lucene.index.DefaultIndexingChain.getOrAddField
0.6% 158 + 1 java.util.DualPivotQuicksort.sort
0.6% 156 + 1 org.apache.lucene.codecs.lucene54.Lucene54DocValuesConsumer.addNumericField
0.5% 138 + 1 org.apache.lucene.index.FreqProxTermsWriterPerField.addTerm
55.2% 15335 + 364 Total compiled (including elided)
Stub + native Method
20.8% 0 + 5906 java.lang.Object.notifyAll
9.3% 0 + 2643 sun.reflect.Reflection.getCallerClass
5.5% 0 + 1573 sun.reflect.Reflection.getCallerClass
1.6% 0 + 458 java.lang.Thread.yield
1.5% 0 + 418 sun.nio.ch.FileDispatcherImpl.write0
1.1% 0 + 308 java.lang.Thread.isAlive
1.1% 0 + 306 sun.nio.ch.NativeThread.current
0.8% 0 + 214 java.lang.String.intern
0.7% 0 + 213 sun.nio.ch.FileDispatcherImpl.force0
0.6% 0 + 167 sun.nio.fs.UnixNativeDispatcher.open0
0.5% 0 + 140 java.security.AccessController.getStackAccessControlContext
0.3% 24 + 66 java.security.AccessController.doPrivileged
0.2% 0 + 57 sun.nio.ch.FileDispatcherImpl.close0
0.1% 0 + 37 java.io.UnixFileSystem.canonicalize0
0.1% 0 + 27 sun.nio.ch.FileChannelImpl.unmap0
0.1% 0 + 16 sun.nio.fs.UnixNativeDispatcher.stat0
0.0% 0 + 11 sun.nio.fs.UnixNativeDispatcher.unlink0
0.0% 0 + 8 sun.nio.ch.FileDispatcherImpl.size0
0.0% 0 + 5 java.lang.Object.hashCode
0.0% 0 + 4 java.lang.System.arraycopy
0.0% 0 + 4 java.lang.Thread.isInterrupted
0.0% 0 + 4 sun.misc.Unsafe.copyMemory
0.0% 0 + 3 sun.nio.ch.FileChannelImpl.map0
0.0% 0 + 2 sun.nio.ch.KQueueArrayWrapper.interrupt
0.0% 0 + 2 java.lang.Object.clone
44.4% 24 + 12598 Total stub (including elided)
Thread-local ticks:
8.7% 2712 Blocked (of total)
Remember that this shows a flat profile. So every method printed here was at some point in time the top-most method on the stack. We do not see anything about the caller-callee relationship of these methods.
We can see that the profiler differentiates between interpreted, compiled and JVM stub samples. For an application in steady state, we’d expect the majority of samples in compiled code (which is the case here). Most samples also show Lucene, which makes sense considering that Lucene is responsible for indexing the data in the end.
When we terminate the application, we’ll see a global summary. Also the bytecode ticks summary shows some statistics now:
Bytecode ticks:
47 0 = ldc
1 0 = ldc_w
1 0 = ifgt
3 0 = if_icmplt
46 0 = goto
2 0 = ireturn
2 0 = areturn
2 0 = return
8 0 = getstatic
1 0 = putstatic
3 0 = getfield
3 0 = putfield
5 0 = invokevirtual
2 0 = invokespecial
6 0 = invokestatic
2 0 = invokeinterface
4 0 = invokedynamic
29 1 = new
91 0 = newarray
106 0 = anewarray
2 0 = checkcast
2 0 = instanceof
2 0 = monitorenter
2 0 = monitorexit
In order to understand the output, you need to glance at the source code. The left column shows the number of ticks outside of stub code and the right column the number of ticks in stub code.
However, you may find the VM summary more useful as we can get a feeling how much time the JVM has spent for compilation, GC and other internal operations.
Global summary of 592.86 seconds:
100.0% 34399 Received ticks
4.6% 1571 Received GC ticks
21.1% 7253 Compilation
0.3% 86 Other VM operations
0.0% 1 Class loader
Flat profile of 0.00 secs (34300 total ticks): VM Thread
0.3% 99 Lost ticks
Compiled + native Method
99.2% 0 + 34034 __psynch_cvwait
0.2% 0 + 76 SafepointSynchronize::begin()
0.2% 0 + 60 PromotionInfo::promoted_oops_iterate_nv(ParScanWithoutBarrierClosure*)
0.0% 0 + 15 ThreadSafepointState::examine_state_of_thread()
0.0% 0 + 13 PromotionInfo::nextDisplacedHeader()
0.0% 0 + 11 SymbolTable::buckets_unlink(int, int, int*, int*, unsigned long*)
0.0% 0 + 9 __psynch_cvsignal
0.0% 0 + 8 RelocIterator::next()
0.0% 0 + 5 JNIHandleBlock::weak_oops_do(BoolObjectClosure*, OopClosure*)
0.0% 0 + 5 __psynch_mutexwait
0.0% 0 + 4 ObjectSynchronizer::deflate_idle_monitors()
0.0% 0 + 4 pthread_getspecific
0.0% 0 + 3 Monitor::IWait(Thread*, long)
0.0% 0 + 3 CodeHeap::find_start(void*) const
0.0% 0 + 3 nmethod::get_deopt_original_pc(frame const*)
0.0% 0 + 2 __vfprintf
0.0% 0 + 2 RelocIterator::advance_over_prefix()
0.0% 0 + 2 __mmap
0.0% 0 + 2 ThreadSafepointState::roll_forward(ThreadSafepointState::suspend_type)
0.0% 0 + 2 _platform_memmove$VARIANT$Haswell
0.0% 0 + 2 ReceiverTypeData::clean_weak_klass_links(BoolObjectClosure*)
0.0% 0 + 2 __mprotect
0.0% 0 + 2 _pthread_cond_wait
0.0% 0 + 2 TemplateInterpreter::notice_safepoints()
0.0% 0 + 2 OopMapSet::find_map_at_offset(int) const
100.0% 0 + 34300 Total compiled (including elided)
Summary
So when should you use -Xprof? It has a couple of advantages:
- Due to its low overhead (it only captures a flat profile so there is no need to walk the stack) you can enable it in production
- It is already available on every (HotSpot) JVM
However, I think the lack of information about it on the Internet is telling. In my opinion it has serious disadvantages:
- You can neither activate it on an already running JVM, nor can you force that captured samples are dumped.
- There is no flag that allows you to write the output to a dedicated file.
- It only provides a very limited view: output is dumped per-thread and the profile is only flat, so you’ll have no clue which code paths called a hot method.
I’d always prefer a “proper” profiler like Java Flight recorder, Honest Profiler or Yourkit during development. If you need to quickly analyze a problem in production, periodically taking thread dumps is usually helpful (either manually or automated, e.g. with stcap). An alternative to the latter may be -Xprof, especially if you are in an extremely limited environment and want to quickly get at least some information out of your application.