Java in High-Performance Computing Dawid Weiss Carrot Search Institute of Computing Science, Poznan University of Technology GeeCon Poznań, 05/2010
Learn from the mistakes of others. You can t live long enough to make them all yourself. Eleanor Roosevelt
Talk outline What is High performance? What is Java? Measuring performance (benchmarking). HPPC library.
Talk outline What is High performance? What is Java? Measuring performance (benchmarking). HPPC library. Crosscutting: (un?)common pitfalls and performance killers. Some HotSpot internals.
Divide-and-conquer style algorithm for (Example e : examples) { e.hasquiz()? e.showquiz() : e.showcode(); e.explain(); e.deriveconclusions(); }
PART I High Performance Computing
High-performance computing (HPC) uses supercomputers and computer clusters to solve advanced computation problems. Wikipedia
Is Java faster than C/C++? The short answer is: it depends. Cliff Click
It s usually hard to make a fast program run faster.
It s usually hard to make a fast program run faster. It s easy to make a slow program run even slower.
It s usually hard to make a fast program run faster. It s easy to make a slow program run even slower. It s easy to make fast hardware run slow.
For now, HPC limited allowed computation time, constrained resources (hardware, memory).
For now, HPC limited allowed computation time, constrained resources (hardware, memory). Good HPC software no (obvious) flaws.
PART II What is Java? (Recall: Is Java faster than C/C++?)
Example 1 public void testsum1() { int sum = 0; for (int i = 0; i < COUNT; i++) sum += sum1(i, i); result = sum; } public void testsum2() { int sum = 0; for (int i = 0; i < COUNT; i++) sum += sum2(i, i); result = sum; }
Example 1 public void testsum1() { int sum = 0; for (int i = 0; i < COUNT; i++) sum += sum1(i, i); result = sum; } public void testsum2() { int sum = 0; for (int i = 0; i < COUNT; i++) sum += sum2(i, i); result = sum; } where the body of sum1 and sum2 sums arguments and returns the result and COUNT is significantly large...
sun-1.6.0-20 VM sum1 sum2
VM sum1 sum2 sun-1.6.0-20 0.04
VM sum1 sum2 sun-1.6.0-20 0.04 2.62 sun-1.6.0-16
VM sum1 sum2 sun-1.6.0-20 0.04 2.62 sun-1.6.0-16 0.04 3.20 sun-1.5.0-18
VM sum1 sum2 sun-1.6.0-20 0.04 2.62 sun-1.6.0-16 0.04 3.20 sun-1.5.0-18 0.04 3.29 ibm-1.6.2
VM sum1 sum2 sun-1.6.0-20 0.04 2.62 sun-1.6.0-16 0.04 3.20 sun-1.5.0-18 0.04 3.29 ibm-1.6.2 0.08 6.28 jrockit-27.5.0
VM sum1 sum2 sun-1.6.0-20 0.04 2.62 sun-1.6.0-16 0.04 3.20 sun-1.5.0-18 0.04 3.29 ibm-1.6.2 0.08 6.28 jrockit-27.5.0 0.18 0.16 harmony-r917296
VM sum1 sum2 sun-1.6.0-20 0.04 2.62 sun-1.6.0-16 0.04 3.20 sun-1.5.0-18 0.04 3.29 ibm-1.6.2 0.08 6.28 jrockit-27.5.0 0.18 0.16 harmony-r917296 0.17 0.35 (averages in sec., 10 measured rounds, 5 warmup, 64-bit Ubuntu, dual-core AMD Athlon 5200).
VM sum1 sum2 sum3 sum4 sun-1.6.0-20 0.04 2.62 1.05 3.76 sun-1.6.0-16 0.04 3.20 1.39 4.99 sun-1.5.0-18 0.04 3.29 1.46 5.20 ibm-1.6.2 0.08 6.28 0.16 14.64 jrockit-27.5.0 0.18 0.16 1.16 3.18 harmony-r917296 0.17 0.35 9.18 22.49 (averages in sec., 10 measured rounds, 5 warmup, 64-bit Ubuntu, dual-core AMD Athlon 5200).
int sum1(int a, int b) { return a + b; } Integer sum2(integer a, Integer b) { return a + b; } Integer sum2(integer a, Integer b) { return Integer.valueOf( a.intvalue() + b.intvalue()); }
int sum3(int... args) { int sum = 0; for (int i = 0; i < args.length; i++) sum += args[i]; return sum; } Integer sum4(integer... args) { int sum = 0; for (int i = 0; i < args.length; i++) { sum += args[i]; } return sum; } Integer sum4(integer [] args) { //... }
Conclusions Syntactic sugar may be costly. Primitive types are fast. Large differences between different VMs.
Example 2 Write once, run anywhere!
But it s the same VM!
It works on my machine!
private static boolean ready; public static void startthread() { new Thread() { public void run() { try { sleep(2000); } catch (Exception e) { /* ignore */ } System.out.println("Marking loop exit."); ready = true; } }.start(); } public static void main(string[] args) { startthread(); System.out.println("Entering the loop..."); while (!ready) { // Do nothing. } System.out.println("Done, I left the loop!"); }
while (!ready) { // Do nothing. }? boolean r = ready; while (!r) { // Do nothing. }
while (!ready) { // Do nothing. }? boolean r = ready; while (!r) { // Do nothing. } In most cases true, from a JMM perspective.
JVM Internals...
C1: fast not (much) optimization C2: slow(er) than C1 a lot of JMM-allowed optimizations
There are hundreds of JVM tuning/diagnostic switches.
My personal favorite:
Conclusions Bytecode is far from what is executed. A lot going on under the (VM) hood. Bad code may work, but will eventually crash. HotSpot-level optimizations are good.
Conclusions Bytecode is far from what is executed. A lot going on under the (VM) hood. Bad code may work, but will eventually crash. HotSpot-level optimizations are good. If there is a bug in the HotSpot compiler...
Any other diversifying factors?
J2ME more VM vendors, hardware diversity, software and hardware quirks.
Non-JVM target platforms Dalvik GWT IKVM
Conclusions There is no single Java performance model. Performance depends on the VM, environment, class library, hardware. Apply benchmark-and-correct cycle.
Benchmarking
Example 3 public void testsum1() { int sum = 0; for (int i = 0; i < COUNT; i++) sum += sum1(i, i); result = sum; } public void testsum1_2() { int sum = 0; for (int i = 0; i < COUNT; i++) sum += sum1(i, i); }
sun-1.6.0-20 VM sum1 sum1_2
VM sum1 sum1_2 sun-1.6.0-20 0.04
VM sum1 sum1_2 sun-1.6.0-20 0.04 0.00
VM sum1 sum1_2 sun-1.6.0-20 0.04 0.00 sun-1.6.0-16 0.04 0.00 sun-1.5.0-18 0.04 0.00 ibm-1.6.2 0.08 0.01 jrockit-27.5.0 0.17 0.08 harmony-r917296 0.17 0.11 (averages in sec., 10 measured rounds, 5 warmup, 64-bit Ubuntu, dual-core AMD Athlon 5200).
java -server -XX:+PrintOptoAssembly -XX:+PrintCompilation...
java -server -XX:+PrintOptoAssembly -XX:+PrintCompilation... - method holder: com/dawidweiss/geecon2010/example03 - access: 0xc1000001 public - name: testsum1_2... 010 pushq rbp subq rsp, #16 # Create frame nop # nop for patch_verified_entry 016 addq rsp, 16 # Destroy frame popq rbp testl rax, [rip + #offset_to_poll_page] # Safepoint: poll for GC 021 ret
Conclusions Benchmarks must be executed to provide feedback. HotSpot is smart and effective at removing dead code.
Example 4 @Test public void testadd1() { int sum = 0; for (int i = 0; i < COUNT; i++) { sum += add1(i); } guard = sum; } public int add1(int i) { return i + 1; } Note add1 is virtual.
switch testadd1 -XX:+Inlining -XX:+PrintInlining 0.04 -XX:-Inlining? (averages in sec., 10 measured rounds, 5 warmup, 64-bit Ubuntu, dual-core AMD Athlon 5200, JRE 1.7b80-debug).
switch testadd1 -XX:+Inlining -XX:+PrintInlining 0.04 -XX:-Inlining 0.45 (averages in sec., 10 measured rounds, 5 warmup, 64-bit Ubuntu, dual-core AMD Athlon 5200, JRE 1.7b80-debug).
Most Java calls are monomorphic.
HotSpot adjusts to megamorphic calls automatically.
Example 5 abstract class Superclass { abstract int call(); } class Sub1 extends Superclass { int call() { return 1; } } class Sub2 extends Superclass { int call() { return 2; } } class Sub3 extends Superclass { int call() { return 3; } } Superclass[] mixed = initwithrandominstances(10000); Superclass[] solid = initwithsub1instances(10000); @Test public void testmonomorphic() { int sum = 0; int m = solid.length; for (int i = 0; i < COUNT; i++) sum += solid[i % m].call(); guard = sum; } @Test public void testmegamorphic() { int sum = 0; int m = mixed.length; for (int i = 0; i < COUNT; i++) sum += mixed[i % m].call(); guard = sum; }
VM monomorphic megamorphic sun-1.6.0-20 0.19 0.32 sun-1.6.0-16 0.19 0.34 sun-1.5.0-18 0.18 0.34 ibm-1.6.2 0.20 0.30 jrockit-27.5.0 0.22 0.29 harmony-r917296 0.27 0.32 (averages in sec., 10 measured rounds, 5 warmup, 64-bit Ubuntu, dual-core AMD Athlon 5200).
Example 6 @Test public void testbitcount1() { int sum = 0; for (int i = 0; i < COUNT; i++) sum += Integer.bitCount(i); guard = sum; } @Test public void testbitcount2() { int sum = 0; for (int i = 0; i < COUNT; i++) sum += bitcount(i); guard = sum; } /* Copied from * {@link Integer#bitCount} */ static int bitcount(int i) { // HD, Figure 5-2 i = i - ((i >>> 1) & 0x55555555); i = (i & 0x33333333) + ((i >>> 2) & 0x33333333); i = (i + (i >>> 4)) & 0x0f0f0f0f; i = i + (i >>> 8); i = i + (i >>> 16); return i & 0x3f; }
VM testbitcount1 testbitcount2 sun-1.6.0-20 0.43 0.43 sun-1.7.0-b80 0.43 0.43 (averages in sec., 10 measured rounds, 5 warmup, 64-bit Ubuntu, dual-core AMD Athlon 5200).
VM testbitcount1 testbitcount2 sun-1.6.0-20 0.43 0.43 sun-1.7.0-b80 0.43 0.43 (averages in sec., 10 measured rounds, 5 warmup, 64-bit Ubuntu, dual-core AMD Athlon 5200). VM testbitcount1 testbitcount2 sun-1.6.0-20 0.08 0.33 sun-1.7.0-b83 0.07 0.32 (averages in sec., 10 measured rounds, 5 warmup, 64-bit Windows 7, Intel I7 860).
... -XX:+PrintInlining...
... -XX:+PrintInlining...... Inlining intrinsic _bitcount_i at bci:9 in..example06::testbitcount1 Inlining intrinsic _bitcount_i at bci:9 in..example06::testbitcount1 Inlining intrinsic _bitcount_i at bci:9 in..example06::testbitcount1 Example06.testBitCount1: [measured 10 out of 15 rounds] round: 0.07 [+- 0.00], round.gc: 0.00 [+- 0.00]... @ 9 com.dawidweiss.geecon2010.example06::bitcount inline (hot) @ 9 com.dawidweiss.geecon2010.example06::bitcount inline (hot) @ 9 com.dawidweiss.geecon2010.example06::bitcount inline (hot) Example06.testBitCount2: [measured 10 out of 15 rounds] round: 0.32 [+- 0.01], round.gc: 0.00 [+- 0.00]...
... -XX:+PrintOptoAssembly...
... -XX:+PrintOptoAssembly... {method} - klass: {other class} - method holder: com/dawidweiss/geecon2010/example06 - name: testbitcount1... 0c2 B13: # B12 B14 <- B8 B12 Loop: B13-B12 inner stride:... 0c2 movl R10, RDX # spill... 0e1 movl [rsp + #40], R11 # spill 0e6 popcnt R8, R8... 0f5 addl R9, #7 # int 0f9 popcnt R11, R11 0fe popcnt RCX, R9
Conclusions Benchmarks must be statistically sound. averages, variance, min, max, warm-up phase Account for HotSpot optimisations. Account for hardware differences. test-on-target Use domain data and real scenarios. Inspect suspicious output with debug JVM. See more: Cliff Click, http://java.sun.com/javaone/2009/articles/rockstar_click.jsp.
HPPC High Performance Primitive Collections
Motivation Primitive types: fast and memory-friendly. Optional assertions. Single-threaded. No fail-fast. Fast, fast, fast iterators, with no GC overhead. Open internals (explicit implementation). Programmers know what they re doing.
Why not JCF? public interface List<E> extends Collection<E> { boolean contains(object o); // [-] contract-enforced methods Iterator<E> iterator(); // [-] iterators over primitive types? Object[] toarray(); // [-] troublesome covariants...
Friendly Competition fastutil PCJ GNU Trove Apache Mahout (ported COLT) Apache Primitive Collections All of these have pros and cons and deal with JCF compatibility somehow.
Iterators in fastutil or PCJ interface IntIterator extends Iterator<Integer> { // Primitive-specific method int nextint(); }
Iterators in HPPC public final class IntCursor { public int index; public int value; } public class IntArrayList extends Iterable<IntCursor> { Iterator<IntCursor> iterator() {... } }
Iterating over list elements in HPPC for (IntCursor c : list) { System.out.println(c.index + ": " + c.value); }
Iterating over list elements in HPPC for (IntCursor c : list) { System.out.println(c.index + ": " + c.value); }...or list.foreach(new IntProcedure() { public void apply(int value) { System.out.println(value); } });
Iterating over list elements in HPPC for (IntCursor c : list) { System.out.println(c.index + ": " + c.value); }...or list.foreach(new IntProcedure() { public void apply(int value) { System.out.println(value); } });...or final int [] buffer = list.buffer; final int size = list.size(); for (int i = 0; i < size; i++) { System.out.println(i + ": " + buffer[i]); }
The fastest one?
What s in HPPC?
Open implementation is good.
/** * Applies a supplemental hash function to a given * hashcode, which defends against poor quality * hash functions. [...] */ static int hash(int h) { // This function ensures that hashcodes that differ only by // constant multiples at each bit position have a bounded // number of collisions (approximately 8 at default load factor). h ^= (h >>> 20) ^ (h >>> 12); return h ^ (h >>> 7) ^ (h >>> 4); } HashMap rehashes your (carefully crafted) hash code.
HPPC approach (example): public class LongIntOpenHashMap implements LongIntMap { //... public LongIntOpenHashMap(int initialcapacity, float loadfactor, LongHashFunction keyhashfunction, IntHashFunction valuehashfunction) { //... } Defaults: LongMurmurHash, IntHashFunction.
Example 7 Frequency count of character bigrams in a given text.
HPPC: final char [] CHARS = DATA; final IntIntOpenHashMap counts = new IntIntOpenHashMap(); for (int i = 0; i < CHARS.length - 1; i++) { counts.putoradd((chars[i] << 16 CHARS[i + 1]), 1, 1); } JCF, boxed integer types. final Integer currentcount = map.get(bigram); map.put(bigram, currentcount == null? 1 : currentcount + 1); JCF, with IntHolder (mutable value object). GNU Trove map.adjustorputvalue(bigram, 1, 1); fastutil, OpenHashMap and LinkedOpenHashMap map.put(bigram, map.get(bigram) + 1); PCJ, OpenHashMap and ChainedHashMap
Is Java faster than C/C++? The short answer is: it depends. Cliff Click
Example 8 The same algorithm for building a DFSA automaton accepting a set of strings. Input: 3 565 575 strings, 158M of text.
Example 8 The same algorithm for building a DFSA automaton accepting a set of strings. Input: 3 565 575 strings, 158M of text. real user sys gcc -O2 java 1.6.0_20-64
Example 8 The same algorithm for building a DFSA automaton accepting a set of strings. Input: 3 565 575 strings, 158M of text. gcc -O2 real 63.850s user 63.110s sys 0.240s java 1.6.0_20-64
Example 8 The same algorithm for building a DFSA automaton accepting a set of strings. Input: 3 565 575 strings, 158M of text. gcc -O2 java 1.6.0_20-64 real 63.850s 43.197s user 63.110s 46.370s sys 0.240s 0.840s
Summary and Conclusions
Performance checklist (sanity check) Algorithms, algorithms, algorithms. Proper data structures. Spurious GC activity. Memory barriers in tight loops. CPU cache utilization. Low-level, hotspot-specific code structuring.
HPPC and junit-benchmarks are at: http://labs.carrotsearch.com