Семинар 1 (23) Программирование сопроцессора Intel Xeon Phi

Transcription

1 Семинар () Программирование сопроцессора Intel Xeon Phi Михаил Курносов WWW: Цикл семинаров «Основы параллельного программирования» Институт физики полупроводников им. А. В. Ржанова СО РАН Новосибирск, 6

2 Intel Xeon Phi Intel Teraflops Research Chip (Polaris, x-7 гг.) 8 cores, 8 routers, self-correction system Single-Chip Cloud Computer (SCC, 9) 48 P54C Pentium cores, 4x6 D-mesh of tiles ( cores) Intel Larrabee (X- гг.) GPGPU: x86, cache coherency, 4-bit ring bus, 4-way SMT, разработка отменена в г. Larrabee GPU (SIGGRAPH, 8) Intel MIC (Intel Many Integrated Core Architecture) кеш-когерентный мультипроцессор с общей памятью, аппаратная многопоточность (4-way SMT), широкие векторные регистры (5 бит) Knights Ferry (): in-order cores, 4-way SMT, ring bus, 75 GFLOPS Knights Corner/ Xeon Phi (): nm, >= 5 cores Knights Landing ( nd gen., 4 nm, ): 74 Airmont (Atom) cores, bootable Knights Hill ( rd gen., nm)

3 Intel Knights Corner micro-architecture Больше 5 ядер Pentium P54C: x86 ISA, in-order, 4-way SMT, 5-bit SIMD units Cache: KB L data/i-cache, coherent L cache (5 KB), двусторонняя кольцевая шина Кольцевая шина: двусторонняя Подключается через шину PCI Express cnmin (oak): Intel Xeon Phi A Cores: 57 4-way SMT (core #57 for OS only!): 4 threads RAM: GDDR5 6 GiB Intel Compiler: $ source /opt/intel_cluster/bin/iccvars.sh intel64

4 Подсчет простых чисел (serial version) int is_prime_number(int n) int limit = sqrt(n) + ; for (int i = ; i <= limit; i++) if (n % i == ) return ; return (n > )? : ; Число операций O( n) int count_prime_numbers(int a, int b) int nprimes = ; if (a <= ) /* Count '' as a prime number */ nprimes = ; a = ; if (a % == ) /* Shift 'a' to odd number */ a++; /* Loop over odd numbers: a, a +, a + 4,..., b */ for (int i = a; i <= b; i += ) if (is_prime_number(i)) nprimes++; return nprimes; Проверка всех нечетных чисел в интервале [a, b]

5 Подсчет простых чисел (OpenMP) int count_prime_numbers_omp(int a, int b) int nprimes = ; /* Count '' as a prime number */ if (a <= ) nprimes = ; a = ; /* Shift 'a' to odd number */ if (a % == ) a++; #pragma omp parallel /* Loop over odd numbers: a, a +, a + 4,..., b */ #pragma omp for schedule(dynamic, ) reduction(+:nprimes) for (int i = a; i <= b; i += ) if (is_prime_number(i)) nprimes++; return nprimes;

6 Подсчет простых чисел (OpenMP) int count_prime_numbers_omp(int a, int b) int nprimes = ; /* Count '' as a prime number */ if (a <= ) nprimes = ; a = ; /* Shift 'a' to odd number */ if (a % == ) a++; cnmic (oak.cpct.sibsutis.ru) System board: ASUS ZPE-D8 WS (Dual CPU) CPU: x Intel Xeon E5-6v (.4GHz, Haswell, 6 cores) RAM: 64 GiB, DDR4 Mhz (4x8 GiB, 4x8 GiB) Coprocessor: Intel Xeon Phi A (Cores: 56 4-way SMT, RAM: GDDR5 6 GiB) $ icc -std=c99 -Wall -O -fopenmp -c primes.c -o primes.o $ icc -o primes primes.o -lm -fopenmp #pragma omp parallel /* Loop over odd numbers: a, a +, a + 4,..., b */ #pragma omp for schedule(dynamic, ) reduction(+:nprimes) for (int i = a; i <= b; i += ) return nprimes; if (is_prime_number(i)) nprimes++; $ export OMP_NUM_THREADS= $./primes Count prime numbers in [, ] Result (host serial): 686 Result (host parallel): 686 Execution time (host serial):.76 Execution time (host parallel):.7 Speedup host_serial/host_omp:.7

7 Подсчет простых чисел: Xeon Phi offload (serial) #include <offload.h> double run_phi_serial() #ifdef INTEL_OFFLOAD printf("intel Xeon Phi devices: %d\n", _Offload_number_of_devices()); #endif int n; double t = wtime(); #pragma offload target(mic) out(n) n = count_prime_numbers_phi(a, b); t = wtime() - t; printf("result (phi serial): %d\n", n); return t; Структурный блок выгружается (offload) для выполнения на сопроцессор Требуется указать: объекты, которые необходимо скопировать в память сопроцессора перед выполнением блока (in) объекты, которые необходимо скопировать из памяти сопроцессора после выполнением блока (out)

8 Подсчет простых чисел: Xeon Phi offload (serial) attribute ((target(mic))) int count_prime_numbers_phi(int a, int b) int nprimes = ; /* Count '' as a prime number */ if (a <= ) nprimes = ; a = ; /* Shift 'a' to odd number */ if (a % == ) a++; Функции и переменные, доступ к которым осуществляется на сопроцессоре, помечаются атрибутом attribute ((target(mic))) (гетерогенная компиляция) /* Loop over odd numbers: a, a +, a + 4,..., b */ for (int i = a; i <= b; i += ) if (is_prime_number(i)) nprimes++; return nprimes;

9 Подсчет простых чисел: Xeon Phi offload (serial) attribute ((target(mic))) int is_prime_number(int n) int limit = sqrt(n) + ; for (int i = ; i <= limit; i++) if (n % i == ) return ; return (n > )? : ; Функции и переменные, доступ к которым осуществляется на сопроцессоре, помечаются атрибутом attribute ((target(mic))) (гетерогенная компиляция)

10 Подсчет простых чисел: Xeon Phi offload (serial) int main(int argc, char **argv) printf("count prime numbers in [%d, %d]\n", a, b); double thost_serial = run_host_serial(); double thost_par = run_host_parallel(); double tphi_serial = run_phi_serial(); printf("execution time (host serial): %.6f\n", thost_serial); printf("execution time (host parallel): %.6f\n", thost_par); printf("execution time (phi serial): %.6f\n", tphi_serial); printf("ratio phi_serial/host_serial: %.f\n", tphi_serial / thost_serial); printf("speedup host_serial/host_omp: %.f\n", thost_serial / thost_par); return ; На данном тесте ядро Intel Xeon Phi A в ~5 раз медленнее ядра Intel Xeon E5-6 v $ export OMP_NUM_THREADS= $./primes Count prime numbers in [, ] Result (host serial): 686 Result (host parallel): 686 Intel Xeon Phi devices: Result (phi serial): 686 Execution time (host serial):.8 Execution time (host parallel):.8 Execution time (phi serial): 8.96 Ratio phi_serial/host_serial: 4.86 Speedup host_serial/host_omp:.74

11 Подсчет простых чисел: Xeon Phi offload (OpenMP) double run_phi_parallel() #ifdef INTEL_OFFLOAD printf("intel Xeon Phi devices: %d\n", _Offload_number_of_devices()); #endif int n; double t = wtime(); #pragma offload target(mic) out(n) n = count_prime_numbers_phi_omp(a, b); t = wtime() - t; printf("result (phi parallel): %d\n", n); return t;

12 Подсчет простых чисел: Xeon Phi offload (OpenMP) attribute ((target(mic))) int count_prime_numbers_phi_omp(int a, int b) int nprimes = ; /* Count '' as a prime number */ if (a <= ) nprimes = ; a = ; /* Shift 'a' to odd number */ if (a % == ) a++; /* Loop over odd numbers: a, a +, a + 4,..., b */ #pragma omp parallel #pragma omp for schedule(dynamic, ) reduction(+:nprimes) for (int i = a; i <= b; i += ) if (is_prime_number(i)) nprimes++; return nprimes;

13 Подсчет простых чисел: Xeon Phi offload (OpenMP) int main(int argc, char **argv) printf("count prime numbers in [%d, %d]\n", a, b); double thost_serial = run_host_serial(); double thost_par = run_host_parallel(); double tphi_serial = run_phi_serial(); double tphi_par = run_phi_parallel(); printf("execution time (host serial): %.6f\n", thost_serial); printf("execution time (host parallel): %.6f\n", thost_par); printf("execution time (phi serial): %.6f\n", tphi_serial); printf("execution time (phi parallel): %.6f\n", tphi_par); printf("ratio phi_serial/host_serial: %.f\n", tphi_serial / thost_serial); printf("speedup host_serial/host_omp: %.f\n", thost_serial / thost_par); printf("speedup host_omp/phi_omp: %.f\n", thost_par / tphi_par); printf("speedup host_serial/phi_omp: %.f\n", thost_serial / tphi_par); printf("speedup phi_serial/phi_omp: %.f\n", tphi_serial / tphi_par); return ;

14 Подсчет простых чисел: Xeon Phi offload (OpenMP) int main(int argc, char **argv) printf("count prime numbers in [%d, %d]\n", a, b); double thost_serial = run_host_serial(); double thost_par = run_host_parallel(); double tphi_serial = run_phi_serial(); double tphi_par = run_phi_parallel(); $ cat./launch.sh #!/bin/sh # Host OpenMP export OMP_NUM_THREADS= # Xeon Phi OpenMP export MIC_ENV_PREFIX=MIC printf("execution time (host serial): %.6f\n", thost_serial); export MIC_OMP_NUM_THREADS=4 printf("execution time (host parallel): %.6f\n", thost_par); printf("execution time (phi serial): %.6f\n", tphi_serial);./primes printf("execution time (phi parallel): %.6f\n", tphi_par); printf("ratio phi_serial/host_serial: %.f\n", tphi_serial / thost_serial); printf("speedup host_serial/host_omp: %.f\n", thost_serial / thost_par); printf("speedup host_omp/phi_omp: %.f\n", thost_par / tphi_par); printf("speedup host_serial/phi_omp: %.f\n", thost_serial / tphi_par); $./launch.sh printf("speedup phi_serial/phi_omp: %.f\n", tphi_serial / tphi_par); Count prime numbers in [, ] Execution time (host serial):.78 return ; Execution time (host parallel):.466 Execution time (phi serial): 8.6 Execution time (phi parallel):.7784 Ratio phi_serial/host_serial: 4.85 Phi-OpenMP-версия на 4 потоках в раза быстрее Speedup host_serial/host_omp:.6 последовательной Phi-версии Speedup host_omp/phi_omp:.4 Speedup host_serial/phi_omp:.58 Phi-OpenMP-версия медленее Host-OpenMP-версии в ~7 раз Speedup phi_serial/phi_omp:.4

15 Xeon Phi: привязка потоков к ядрам (affinity) $ cat./launch.sh #!/bin/sh # Host OpenMP export OMP_NUM_THREADS= # Xeon Phi OpenMP export MIC_ENV_PREFIX=MIC export MIC_OMP_NUM_THREADS=56 export MIC_KMP_AFFINITY=verbose Execution time (host serial):.9 Execution time (host parallel):.86 Execution time (phi serial): 8.78 Execution time (phi parallel):.546 Ratio phi_serial/host_serial: 4.86 Speedup host_serial/host_omp:. Speedup host_omp/phi_omp:. Speedup host_serial/phi_omp:.4 Speedup phi_serial/phi_omp:../primes OMP: Info #4: KMP_AFFINITY: decoding xapic ids. OMP: Info #5: KMP_AFFINITY: cpuid leaf not supported - decoding legacy APIC ids. OMP: Info #49: KMP_AFFINITY: Affinity capable, using global cpuid info OMP: Info #54: KMP_AFFINITY: Initial OS proc set respected:,,,4,5,6,7,8,9,,,,,4,5,6,7,8,9,,,,,4,5,6,7,8,9,,,,,4,5,6,7,8,9,4,4,4,4,44,45,46,47,48,49,5,5,5,5,54,55,56,57,58,59,6,6,6,6,64,65,66,67,68,69,7,7,7,7,74,75,76,77,78,79,8,8,8,8,84,85,86,87,88,89,9,9,9,9,94,95,96,97,98,99,,,,,4,5,6,7,8,9,,,,,4,5,6,7,8,9,,,,,4,5,6,7,8,9,,,,,4,5,6,7,8,9,4,4,4,4,44,45,46,47,48,49,5,5,5,5,54,55,56,57,58,59,6,6,6,6,64,65,66,67,68,69,7,7,7,7,74,75,76,77,78,79,8,8,8,8,84,85,86,87,88,89,9,9,9,9,94,95,96,97,98,99,,,,,4,5,6,7,8,9,,,,,4,5,6,7,8,9,,,,,4 OMP: Info #56: KMP_AFFINITY: 4 available OS procs OMP: Info #57: KMP_AFFINITY: Uniform topology OMP: Info #59: KMP_AFFINITY: packages x 56 cores/pkg x 4 threads/core (56 total cores)

16 Xeon Phi: привязка потоков к ядрам (affinity) OMP: Info #6: KMP_AFFINITY: OS proc to physical thread map: OMP: Info #7: KMP_AFFINITY: OS proc maps to package core thread OMP: Info #7: KMP_AFFINITY: OS proc maps to package core thread OMP: Info #7: KMP_AFFINITY: OS proc maps to package core thread OMP: Info #7: KMP_AFFINITY: OS proc 4 maps to package core thread OMP: Info #7: KMP_AFFINITY: OS proc 5 maps to package core thread OMP: Info #7: KMP_AFFINITY: OS proc 6 maps to package core thread OMP: Info #7: KMP_AFFINITY: OS proc 7 maps to package core thread OMP: Info #7: KMP_AFFINITY: OS proc 8 maps to package core thread OMP: Info #7: KMP_AFFINITY: OS proc 9 maps to package core thread OMP: Info #7: KMP_AFFINITY: OS proc maps to package core thread OMP: Info #7: KMP_AFFINITY: OS proc maps to package core thread OMP: Info #7: KMP_AFFINITY: OS proc maps to package core thread... OMP: Info #7: KMP_AFFINITY: OS proc maps to package core 55 thread OMP: Info #7: KMP_AFFINITY: OS proc maps to package core 55 thread OMP: Info #7: KMP_AFFINITY: OS proc maps to package core 55 thread OMP: Info #7: KMP_AFFINITY: OS proc 4 maps to package core 55 thread Core Core Core Core 55

17 Xeon Phi: привязка потоков к ядрам (affinity) OMP: Info #4: KMP_AFFINITY: pid 4 thread bound to OS proc set OMP: Info #4: KMP_AFFINITY: pid 4 thread bound to OS proc set 5 OMP: Info #4: KMP_AFFINITY: pid 4 thread bound to OS proc set 9 OMP: Info #4: KMP_AFFINITY: pid 4 thread bound to OS proc set... OMP: Info #4: KMP_AFFINITY: pid 4 thread 54 bound to OS proc set 7 OMP: Info #4: KMP_AFFINITY: pid 4 thread 55 bound to OS proc set Cores (HW threads): Core Core Core Core Core 4 Core 5 Core 55 OMP threads: 55 Default Affinity

18 Xeon Phi: привязка потоков к ядрам (affinity) $ cat./launch.sh #!/bin/sh # Host OpenMP export OMP_NUM_THREADS= # Xeon Phi OpenMP export MIC_ENV_PREFIX=MIC export MIC_OMP_NUM_THREADS=56 export MIC_KMP_AFFINITY=granularity=fine,balanced Execution time (host serial):.5 Execution time (host parallel):.67 Execution time (phi serial): Execution time (phi parallel):.4887 Ratio phi_serial/host_serial: 4.86 Speedup host_serial/host_omp:.7 Speedup host_omp/phi_omp:. Speedup host_serial/phi_omp:.5 Speedup phi_serial/phi_omp: 7.4./primes fine, balanced Cores (HW threads): Core Core Core Core Core 4 Core 5 Core 55 OMP threads: 55

19 Xeon Phi: привязка потоков к ядрам (affinity) $ cat./launch.sh #!/bin/sh # Host OpenMP export OMP_NUM_THREADS= # Xeon Phi OpenMP export MIC_ENV_PREFIX=MIC export MIC_OMP_NUM_THREADS=56 export MIC_KMP_AFFINITY=granularity=fine,compact./primes Execution time (host serial):.5 Execution time (host parallel):.78 Execution time (phi serial): Execution time (phi parallel):.4787 Ratio phi_serial/host_serial: 4.86 Speedup host_serial/host_omp:.7 Speedup host_omp/phi_omp:.7 Speedup host_serial/phi_omp:.8 Speedup phi_serial/phi_omp:.9 fine, compact Cores (HW threads): Core Core Core Core Core 4 Core 5 Core OMP threads:

20 Xeon Phi: привязка потоков к ядрам (affinity) $ cat./launch.sh #!/bin/sh # Host OpenMP export OMP_NUM_THREADS= # Xeon Phi OpenMP export MIC_ENV_PREFIX=MIC export MIC_OMP_NUM_THREADS=56 export MIC_KMP_AFFINITY=granularity=fine,scatter./primes Execution time (host serial):.4 Execution time (host parallel):.878 Execution time (phi serial): Execution time (phi parallel):.55 Ratio phi_serial/host_serial: 4.86 Speedup host_serial/host_omp:.5 Speedup host_omp/phi_omp:. Speedup host_serial/phi_omp:. Speedup phi_serial/phi_omp:.7 fine, scatter Cores (HW threads): Core Core Core Core Core 4 Core 5 Core 55 OMP threads: 55

21 Xeon Phi: привязка потоков к ядрам (affinity) NUM_THREADS = fine, scatter: core_id = thread_id % ncores Cores (HW threads): Core Core Core Core Core 4 Core 5 Core 55 OMP threads: nodes: 8x + 48x = 55 fine, balanced: Cores (HW threads): Core Core Core Core Core 4 Core 5 Core 55 OMP threads: nodes: 8x + 48x = 8 9 fine, compact: Cores (HW threads): Core Core Core Core Core 4 Core 5 Core 9 OMP threads: nodes: x4 =

22 Задание Реализуйте вычисление определенного интеграла методом прямоугольников на Intel Xeon Phi (шаблон размещен в каталоге _task_integrate):. Выгрузите выполнение функции integrate_phi_omp на сопроцессор. Функцию integare_phi_omp реализуйте на OpenMP