![为什么Java程序员也需要了解CPU
Puzzle 3
下面哪个程序跑得更快?
private static long[][] longs; private static long[][] longs;
for (int i = 0; i < DIMENSION_1; i++) { for (int j = 0; j < DIMENSION_2; j++) {
for (int j = 0; j < DIMENSION_2; j++) { for (int i = 0; i < DIMENSION_1; i++){
sum += longs[i][j]; sum += longs[i][j];
} }
} }
Taobao Java Team | zhouchen.zm](https://image.slidesharecdn.com/javacpu-120629212711-phpapp01/85/Java-cpu-12-320.jpg)
![为什么Java程序员也需要了解CPU
Puzzle 4
下面程序有什么问题?
public final static class VolatileLong { longs = new VolatileLong[NUM_THREADS];
public volatile long value = 0L; for (int i = 0; i < longs.length; i++) {
} longs[i] = new VolatileLong();
}
public class FS implements Runnable { // 每个线程更新独立的变量
public void run() { for (int i = 0; i < NUM_THREADS; i++) {
long i = ITERATIONS + 1; threads[i] = new Thread(new FS(i)));
while (0 != --i) { }
longs[arrayIndex].value = i; for (Thread t : threads) {
} t.start();
} }
} for (Thread t : threads) {
t.join();
}
Taobao Java Team | zhouchen.zm](https://image.slidesharecdn.com/javacpu-120629212711-phpapp01/85/Java-cpu-15-320.jpg)
![为什么Java程序员也需要了解CPU
Example
1. class X { $ java -XX:+UnlockDiagnosticVMOptions -XX:PrintAssemblyOptions=hsdis-print-bytes -
2. int a, b; XX:CompileCommand=print,X.f X
3. volatile int v, u;
4. void f() { …
5. int i, j;
6.
7. i = a; // load a
[Verified Entry Point]
8. j = b; // load b 0xb3a1e28c: push %ebp ;...55
9. i = v; // load v 0xb3a1e28d: sub $0x8,%esp ;...81ec0800 0000
10. // LoadLoad 0xb3a1e293: mov 0x10(%ecx),%ebx ;...8b5910 ; i = v
11. j = u; // load u 0xb3a1e296: mov 0x14(%ecx),%edi ;...8b7914 ; j = u
12. // LoadStore 0xb3a1e299: mov %ebx,0x8(%ecx) ;...895908 ; a = i
13. a = i; // store a 0xb3a1e29c: mov %edi,0xc(%ecx) ;...89790c ;b=j
14. b = j; // store b 0xb3a1e29f: mov %ebx,0x10(%ecx) ;...895910 ; v = i
15. // StoreStore 0xb3a1e2a2: mov %edi,0x14(%ecx) ;...897914 ; u = j
16. v = i; // store v 0xb3a1e2a5: lock addl $0x0,(%esp) ;...f0830424 ; memory barrier
17. // StoreStore
0xb3a1e2aa: mov 0x14(%ecx),%ebp ;...8b6914 ; i = u
18. u = j; // store u
19. // StoreLoad
0xb3a1e2ad: mov %ebp,0x8(%ecx) ;...896908 ; a = i
20. i = u; // load u 0xb3a1e2b0: add $0x8,%esp ;...83c408
21. // LoadLoad 0xb3a1e2b3: pop %ebp ;...5d
22. // LoadStore 0xb3a1e2b4: test %eax,0xb78ab000 ;...850500b0 8ab7
23. j = b; // load b
24. a = i; // store a …
25. }
26.}
参考: jdk/hotspot/src/os_cpu/linux_x86/vm/orderAccess_linux_x86.inline.hpp
Taobao Java Team | zhouchen.zm](https://image.slidesharecdn.com/javacpu-120629212711-phpapp01/85/Java-cpu-24-320.jpg)
Java cpu
- 1. 了解CPU 作者：周忱 | 淘宝综合业务 微博：@MinZhou 邮箱：[email protected]
- 2. 为什么Java程序员也需要了解CPU 关于我 • 花名:周忱(chén) • 真名:周敏 • 微博: @MinZhou • Twitter: @minzhou • 2010年6月加入淘宝 • 曾经淘宝Hadoop&Hive研发 组Leader • 目前专注分布式实时计算 • Hive Contributor • 自由、开源软件热爱者 Taobao Java Team | zhouchen.zm
- 3. 为什么Java程序员也需要了解CPU 关于我 • 花名:周忱(chén) • 真名:周敏 • 微博: @MinZhou • Twitter: @minzhou • 2010年6月加入淘宝 • 曾经淘宝Hadoop&Hive研发 组Leader • 目前专注分布式实时计算 • Hive Contributor • 自由、开源软件热爱者 Taobao Java Team | zhouchen.zm
- 4. 为什么Java程序员也需要了解CPU 什么是volatile? Taobao Java Team | zhouchen.zm
- 5. 为什么Java程序员也需要了解CPU Puzzle 1 下面哪个程序跑得更快? private static long value; private static volatile long value; public static void increment() { public static void increment() { while (value < 100000000L) { while (value < 100000000L) { value++; value++; } } } } 如果换成AtomicLong呢? Taobao Java Team | zhouchen.zm
- 6. 为什么Java程序员也需要了解CPU Puzzle 2 下面哪个程序跑得更快? AtomicIntegerFieldUpdater state = … AtomicIntegerFieldUpdater state= … public void lock() { public void lock() { while(state.getAndSet(true)) {} ; //spin while (true) { } while(state.get()) {}; //spin if(!state.getAndSet(true)) return; } } 如果换成多线程呢? Taobao Java Team | zhouchen.zm
- 7. 为什么Java程序员也需要了解CPU Intel 十年计划:Tick-Tock Taobao Java Team | zhouchen.zm
- 8. 为什么Java程序员也需要了解CPU CPU微架构变革 • Hyper-threading • 32KB L1d 与32KB L1i • 256KB L2 Cache • 4–12 MB L3 cache • QPI 代替FSB • PCI E和DMI整合入 处理器, 北桥消失 • 支持 2/3通道DDR3 Taobao Java Team | zhouchen.zm
- 9. 为什么Java程序员也需要了解CPU 现代CPU微架构 Taobao Java Team | zhouchen.zm
- 10. 为什么Java程序员也需要了解CPU Intel Nehalem 排布 Taobao Java Team | zhouchen.zm
- 11. 为什么Java程序员也需要了解CPU Intel Sandy Bridge Taobao Java Team | zhouchen.zm
- 12. 为什么Java程序员也需要了解CPU Puzzle 3 下面哪个程序跑得更快? private static long[][] longs; private static long[][] longs; for (int i = 0; i < DIMENSION_1; i++) { for (int j = 0; j < DIMENSION_2; j++) { for (int j = 0; j < DIMENSION_2; j++) { for (int i = 0; i < DIMENSION_1; i++){ sum += longs[i][j]; sum += longs[i][j]; } } } } Taobao Java Team | zhouchen.zm
- 13. 为什么Java程序员也需要了解CPU CPU Cache cat /sys/devices/system/cpu/cpu0/cache/index0/* Taobao Java Team | zhouchen.zm
- 14. 为什么Java程序员也需要了解CPU Cache Coherence & MESI协议 • M: 修改 • E : 独占 • S : 共享 • I : 失效 发音: messy Taobao Java Team | zhouchen.zm
- 15. 为什么Java程序员也需要了解CPU Puzzle 4 下面程序有什么问题? public final static class VolatileLong { longs = new VolatileLong[NUM_THREADS]; public volatile long value = 0L; for (int i = 0; i < longs.length; i++) { } longs[i] = new VolatileLong(); } public class FS implements Runnable { // 每个线程更新独立的变量 public void run() { for (int i = 0; i < NUM_THREADS; i++) { long i = ITERATIONS + 1; threads[i] = new Thread(new FS(i))); while (0 != --i) { } longs[arrayIndex].value = i; for (Thread t : threads) { } t.start(); } } } for (Thread t : threads) { t.join(); } Taobao Java Team | zhouchen.zm
- 16. 为什么Java程序员也需要了解CPU False Sharing public final static class VolatileLong { public volatile long value = 0L; } => public final static class VolatileLong { public volatile long value = 0L; public long p1, p2, p3, p4, p5, p6; } Taobao Java Team | zhouchen.zm
- 17. 为什么Java程序员也需要了解CPU False Sharing public final static class VolatileLong { public volatile long value = 0L; } => public final static class VolatileLong { public volatile long value = 0L; public long p1, p2, p3, p4, p5, p6; } Taobao Java Team | zhouchen.zm
- 18. 为什么Java程序员也需要了解CPU False Sharing public final static class VolatileLong { public volatile long value = 0L; } => public final static class VolatileLong { public volatile long value = 0L; public long p1, p2, p3, p4, p5, p6; } Taobao Java Team | zhouchen.zm
- 19. 为什么Java程序员也需要了解CPU Cache Consistency & 乱序执行 • CPU指令重排序 • 编译重排序 Taobao Java Team | zhouchen.zm
- 20. 为什么Java程序员也需要了解CPU Memory Ordering • 程序顺序(Program ordering): 读写和指令编制一致, 也称强顺序. • 处理器顺序(Processor ordering): 加快指令执行速度, 例如读可跨越它之前 的buffered写 • Intel386及之前内存顺序模型: 程序顺序 • Pentium和Intel486内存顺序模型:大多数情况下遵循程序顺序 • P6及以后内存顺序模型：处理器顺序, 同时提供其它加强或弱化内存顺序 模型的特殊指令 Taobao Java Team | zhouchen.zm
- 21. 为什么Java程序员也需要了解CPU 内存屏障 • Load Buffer • Store Buffer • CPU串行化指令 – CPUID – SFENCE – LFENCE – MFENCE • Lock系指令 Taobao Java Team | zhouchen.zm
- 22. 为什么Java程序员也需要了解CPU Java内存模型 • Happens-Before // Lock • StoreStore pthread_mutex_lock(&lock); • LoadStore sequence = i; pthread_cond_signal(&condition); • StoreLoad pthread_mutex_unlock(&lock); • LoadLoad • Full Fence // Soft Barrier • 锁和synchronized块 asm volatile(“” ::: “memory”); • volatile 关键字 sequence = i; • final 关键字 // Fence asm volatile(“” ::: “memory”); sequence = i; asm volatile(“lock addl $0x0,(%rsp)”); Taobao Java Team | zhouchen.zm
- 23. 为什么Java程序员也需要了解CPU Java内存模型 需要内存屏 重排序 次操作 次操作 障 Normal Volatile Volatile Volatile Load Load Store Normal Normal Load Volatile 首操作 首操作 Normal Monitor Monitor Load Store MonitorEnt Store er Store Enter Exit Normal Normal Load LoadStore Load No Normal Normal Store StoreStore Store Volatile Load Volatile Load No No No LoadLoad LoadStore LoadLoad LoadStore MonitorEnter Monitor Enter Volatile store Volatile Store No No StoreLoad StoreStore Monitor MonitorExit Exit Taobao Java Team | zhouchen.zm
- 24. 为什么Java程序员也需要了解CPU Example 1. class X { $ java -XX:+UnlockDiagnosticVMOptions -XX:PrintAssemblyOptions=hsdis-print-bytes - 2. int a, b; XX:CompileCommand=print,X.f X 3. volatile int v, u; 4. void f() { … 5. int i, j; 6. 7. i = a; // load a [Verified Entry Point] 8. j = b; // load b 0xb3a1e28c: push %ebp ;...55 9. i = v; // load v 0xb3a1e28d: sub $0x8,%esp ;...81ec0800 0000 10. // LoadLoad 0xb3a1e293: mov 0x10(%ecx),%ebx ;...8b5910 ; i = v 11. j = u; // load u 0xb3a1e296: mov 0x14(%ecx),%edi ;...8b7914 ; j = u 12. // LoadStore 0xb3a1e299: mov %ebx,0x8(%ecx) ;...895908 ; a = i 13. a = i; // store a 0xb3a1e29c: mov %edi,0xc(%ecx) ;...89790c ;b=j 14. b = j; // store b 0xb3a1e29f: mov %ebx,0x10(%ecx) ;...895910 ; v = i 15. // StoreStore 0xb3a1e2a2: mov %edi,0x14(%ecx) ;...897914 ; u = j 16. v = i; // store v 0xb3a1e2a5: lock addl $0x0,(%esp) ;...f0830424 ; memory barrier 17. // StoreStore 0xb3a1e2aa: mov 0x14(%ecx),%ebp ;...8b6914 ; i = u 18. u = j; // store u 19. // StoreLoad 0xb3a1e2ad: mov %ebp,0x8(%ecx) ;...896908 ; a = i 20. i = u; // load u 0xb3a1e2b0: add $0x8,%esp ;...83c408 21. // LoadLoad 0xb3a1e2b3: pop %ebp ;...5d 22. // LoadStore 0xb3a1e2b4: test %eax,0xb78ab000 ;...850500b0 8ab7 23. j = b; // load b 24. a = i; // store a … 25. } 26.} 参考: jdk/hotspot/src/os_cpu/linux_x86/vm/orderAccess_linux_x86.inline.hpp Taobao Java Team | zhouchen.zm
- 25. 为什么Java程序员也需要了解CPU Puzzle 5 为什么有volatile还需要AtomicXX及 Unsafe.compareAndSwapXX? Taobao Java Team | zhouchen.zm
- 26. 为什么Java程序员也需要了解CPU 原子指令 • CompareAndSwap: LOCK XCHG • LOCK XADD • ABA问题 Taobao Java Team | zhouchen.zm
- 27. 为什么Java程序员也需要了解CPU 让步锁 public class BackoffLock { if (!state.getAndSet(true)) { private AtomicBoolean state = new return; AtomicBoolean(false); } else { private static final int MIN_DELAY = ...; backoff.backoff(); private static final int MAX_DELAY = ...; } } public void lock() { } Backoff backoff = new Backoff(MIN_DELAY, MAX_DELAY); public void unlock() { while (true) { state.set(false); while (state.get()) {}; } } Taobao Java Team | zhouchen.zm
- 28. 为什么Java程序员也需要了解CPU 队列锁 • BackOff锁的问题 – Cache-coherence Traffic – 临界区利用率 – 公平性问题 – 饥饿 • 基于数组的队列 • CLH队列 • MCS队列 • NUMA-Aware变种 https://blogs.oracle.com/dave/entry/flat_combining_numa_locks Taobao Java Team | zhouchen.zm
- 29. 为什么Java程序员也需要了解CPU Puzzle 6 为什么j.u.c或者jvm里面在加锁的时候, 喜欢用先 CAS spin尝试N次, 如果失败则yield多次(可选), 如果 最后还是失败则park, 同时把线程加到队列里? Taobao Java Team | zhouchen.zm
- 30. 为什么Java程序员也需要了解CPU 工具 • top • vmstat • lscpu • perf • Valgrind tools suite • OProfile • SystemTap • numactl • Intel Vtune • MAT Taobao Java Team | zhouchen.zm
- 31. 为什么Java程序员也需要了解CPU 推荐读物 • What every programmer should know about memory • Intel® 64 and IA-32 Architectures Software Developer Manuals • The Art of Multiprocessor Programming • The JSR-133 Cookbook for Compiler Writers (Java Memory Model) • 本人博客: http://coderplay.javaeye.com Taobao Java Team | zhouchen.zm
- 32. 为什么Java程序员也需要了解CPU Q&A 作者：周忱 | 淘宝综合业务 微博：@MinZhou 邮箱：[email protected] Taobao Java Team | zhouchen.zm