并行編程實(shí)驗(yàn)資料報(bào)告材料

1、爭(zhēng)申科技衣茅課程實(shí)驗(yàn)報(bào)告課程名稱：并行編程計(jì)算機(jī)科學(xué)與技術(shù)學(xué)院目錄實(shí)驗(yàn)一 21. 實(shí)驗(yàn)?zāi)康呐c要求 22. 實(shí)驗(yàn)容 23. 實(shí)驗(yàn)結(jié)果 2實(shí)驗(yàn)二 31. 實(shí)驗(yàn)?zāi)康呐c要求 32. 算法描述 33. 實(shí)驗(yàn)方案 34. 實(shí)驗(yàn)結(jié)果與分析 5實(shí)驗(yàn)三 61. 實(shí)驗(yàn)?zāi)康呐c要求 62. 算法描述 63. 實(shí)驗(yàn)方案 64. 實(shí)驗(yàn)結(jié)果與分析 7實(shí)驗(yàn)四 81. 實(shí)驗(yàn)?zāi)康呐c要求 82. 算法描述 83. 實(shí)驗(yàn)方案 84. 實(shí)驗(yàn)結(jié)果與分析 11實(shí)驗(yàn)五 121實(shí)驗(yàn)?zāi)康呐c要求 122. 算法描述 123. 實(shí)驗(yàn)方案 124. 實(shí)驗(yàn)結(jié)果與分析 14PROJECT.2 16AIM: 16HYPOTHESIS.: 16METHODS

2、.: 16RESULT: 16DICUSSION&CONCLUSI.O.N 17REFERENC.E 18實(shí)驗(yàn)一1. 實(shí)驗(yàn)?zāi)康呐c要求become familiar with the parallel development environments, and the basic principles and methods of parallel programming and performance optimization by using tools and frameworks like pthread, OpenMP, MPI under Linux system.2. 實(shí)驗(yàn)

3、容熟悉并行開發(fā)環(huán)境，掌握并行編程用到的工具如線程、 OpenMP,、MPI 等。3. 實(shí)驗(yàn)結(jié)果通過上機(jī)操作熟悉了各種命令，編寫了簡(jiǎn)單的程序熟悉開發(fā)環(huán)境。實(shí)驗(yàn)二1. 實(shí)驗(yàn)?zāi)康呐c要求a) master the basic principles and methods of parallel programming design and performance optimization using pthreadb) understand the basic method for data partition and task decomposition in parallel programmin

4、gc) implement the parallel algorithm of calculating the value of pi using pthreadd) then carries on the simple analysis and summary of the program execution results2. 算法描述采用蒙特卡洛方法計(jì)算圓周率， 利用單位圓與邊長(zhǎng)為 1 的正方形面積之比 計(jì)算圓周率的近似值。比值的計(jì)算采用蒙特卡羅方法的隨即投點(diǎn)思想， 在正方形中隨機(jī)投入很多 點(diǎn)，使所投點(diǎn)在正方形中每一個(gè)位置的機(jī)會(huì)均等， 然后考察有多少個(gè)點(diǎn)落 在扇形，落在扇形的點(diǎn)的個(gè)數(shù)與投

5、點(diǎn)總數(shù)之比就是該比例的近似值。 每一個(gè)線程完成一次投點(diǎn)， n 個(gè)線程同時(shí)計(jì)算。3. 實(shí)驗(yàn)方案開發(fā)與運(yùn)行環(huán)境：使用筆記本電腦登錄實(shí)驗(yàn)室服務(wù)器。實(shí)驗(yàn)代碼如下：#include <pthread.h>#include <stdlib.h>#include <stdio.h>#define MaxThreadNum 100#define kSamplePoints 10000000#define kSpace 1 void *compute_pi(void *);int total_hits, hitsMaxThreadNumkSpace; int sample_p

6、oints_per_thread, num_threads;int main(void)int i;pthread_t p_threadsMaxThreadNum;pthread_attr_t attr;pthread_attr_init(&attr);pthread_attr_setscope(&attr, PTHREAD_SCOPE_SYSTEM); printf("Enter num_threadsn");scanf("%d", &num_threads);total_hits = 0; sample_points_per_

7、thread = kSamplePoints / num_threads;for(i=0; i<num_threads; i+) hitsi0 = i; pthread_create(&p_threadsi, &attr, compute_pi, (void *)&hitsi);for(i=0; i<num_threads; i+) pthread_join(p_threadsi, NULL); total_hits += hitsi0;double pi = 4.0 * (double)total_hits / kSamplePoints; printf(

8、" Pi: %lfn", pi);return 0;void *compute_pi(void * s)unsigned int seed;int i;int *hit_pointer;double rand_no_x, rand_no_y; hit_pointer = (int *)s; seed = *hit_pointer;/ int local_hits = 0;for(i=0; i<sample_points_per_thread; i+) rand_no_x = (double)(rand_r(&seed)/(double)(RAND_MAX);r

9、an d_no_y = (double)(ra nd(&seed)/(double)(RAND_MAX); if(ra nd_no_x - 0.5)*(ra nd_no_x - 0.5) + (ran d_no_y - 0.5) (ran d_no_y - 0.5) < 0.25)一 一 (*hit_poi nter)+;seed *= i; pthread_exit(0); _4.實(shí)驗(yàn)結(jié)果與分析實(shí)驗(yàn)結(jié)果符合預(yù)期：pppuseEnter num_thread5Pi: 3.152斗日2實(shí)驗(yàn)三1. 實(shí)驗(yàn)?zāi)康呐c要求a) master the basic principles and m

10、ethods of parallel programming design and performance optimization using OpenMPb) implement the parallel algorithm of calculating the value of pi using OpenMPc) carries on the simple analysis and summary of the program execution resultsd) compare it with the results of Lab22. 算法描述與實(shí)驗(yàn)二相似，同樣采用蒙特卡羅方法計(jì)算

11、 pi 值，算法不再詳細(xì)描述。3. 實(shí)驗(yàn)方案實(shí)驗(yàn)環(huán)境：使用筆記本電腦登錄實(shí)驗(yàn)室服務(wù)器。實(shí)驗(yàn)代碼如下：#include <omp.h>#include <stdio.h>#include <stdlib.h>#include <math.h>#define SEED 35791246main(int argc, char* argv)int numiter=0;/loop timesdouble x,y,z,pi;int i,count; /* # of points in the 1st quadrant of unit circle */ pr

12、intf("Enter the number of iterations used to estimate pi: "); scanf("%d",&niter);/* initialize random numbers */ srand(SEED);count=0;int chunk;/ sizechunk = 1;#pragma omp parallel shared(chunk) private(i,x,y,z) reduction(+:count) #pragma omp for schedule(dynamic,chunk)for ( i

13、=0; i<niter; i+) x = (double)rand()/RAND_MAX; y = (double)rand()/RAND_MAX;z = x*x+y*y; if (z<=1) coun t+; pi=(double)co un t/niter*4; printf("# loop times= %d , estimate of pi is %g n",niter,pi);Openmp自動(dòng)將for循環(huán)分解成多個(gè)線程并行執(zhí)行。4.實(shí)驗(yàn)結(jié)果與分析實(shí)驗(yàn)結(jié)果如下：pppuserSSOtfinodeiSl 1訪豈$. / 3En±er the nu

14、mber of iterations lib已d to estImate pl:10&M# IcFctp 七i mES=f est i.rnat e erf pi. i s 3.133&Time： D + Epppu«r2S&gnode231 lsb3JJ . /3二n七已r th亡 number o4 iteration£ used to estimate pi: 'i lopp time5=t 好耳tinrat亡 of pi i旨 3.14159rimen 4i7與實(shí)驗(yàn)二相比本實(shí)驗(yàn)是使用ope nmp自動(dòng)分解多線程并行，精度并無明顯 區(qū)別

15、，與所選算法有關(guān)。循環(huán)次數(shù)越多，得到的結(jié)果就越精確。實(shí)驗(yàn)四1. 實(shí)驗(yàn)?zāi)康呐c要求a) master the basic principles and methods of parallel programming design and performance optimization using MPIb) implement the parallel algorithm of calculating the value of pi using MPIc) carries on the simple analysis and summary of the program execution re

16、sultsd) compare it with the results of Lab2 and Lab32. 算法描述本實(shí)驗(yàn)采用與實(shí)驗(yàn)一實(shí)驗(yàn)二相同的蒙特卡羅算法實(shí)現(xiàn) pi 值得計(jì)算，即利 用單位圓與邊長(zhǎng)為 1 的正方形面積之比計(jì)算圓周率的近似值。 比值的計(jì)算采用蒙特卡羅方法的隨即投點(diǎn)思想， 在正方形中隨機(jī)投入很多 點(diǎn)，使所投點(diǎn)在正方形中每一個(gè)位置的機(jī)會(huì)均等， 然后考察有多少個(gè)點(diǎn)落 在扇形，落在扇形的點(diǎn)的個(gè)數(shù)與投點(diǎn)總數(shù)之比就是該比例的近似值。3. 實(shí)驗(yàn)方案Mpi 是一種基于消息傳遞的并行編程技術(shù)， 各個(gè)進(jìn)程有獨(dú)立的堆棧和代碼 段，進(jìn)程之間的信息交互通過調(diào)用通信函數(shù)完成?；镜?API 如下：in

17、t MPI_Init(int *argc, char *argv)MPInit是MPI程序的第一個(gè)調(diào)用，它完成MPI程序的所有初始化 工作， 啟動(dòng) MPI 環(huán)境，標(biāo)志并行代碼的開始。int MPI_Finalize(void)MPI_Finalize 是 MPI 程序的最后一個(gè)調(diào)用，它結(jié)束 MPI 程序的運(yùn) 行，標(biāo) 志并行代碼的結(jié)束， 結(jié)束除主進(jìn)程外其它進(jìn)程。 其之后串行代碼 仍可在 主進(jìn)程 (rank = 0) 上繼續(xù)運(yùn)行。int MPI_Comm_size(MPI_Comm comm, int *size);獲取進(jìn)程個(gè)數(shù) p。int MPI_Comm_rank(MPI_Comm comm,

18、 int *rank);MPI獲取當(dāng)前進(jìn)程的 RANK rank值取址圍是 0p-1 , RANKfi唯一的表示了進(jìn)程的ID，其中Rank=0的為主進(jìn)程int MPI_Send(void* buf, int count, MPI_Datatype datatype, int dest,int tag,MPI_Comm comm);發(fā)送函數(shù)：當(dāng)前進(jìn)程將以 buf 為初始地址，長(zhǎng)度為 count 且元素類型 為 datatype 的信息發(fā)動(dòng)給 rank 值為 dest 的進(jìn)程，這條消息的標(biāo)識(shí)符 為 tag 。其中datatype有MPINT, MPI_FLOA等常用類型，Tag的作用是用于 區(qū)分一

19、對(duì)進(jìn)程之間發(fā)送的不同信息int MPI_Recv(void* buf, int count, MPI_Datatype datatype, int source, int tag, MPI_Comm comm, MPI_Status *status);接受函數(shù)：從 rank 值為 source 的進(jìn)程接受標(biāo)識(shí)符為 tag 的信息，存 入以 buf 為初始地址， 長(zhǎng)度為 count 的存儲(chǔ)區(qū)域中， 類型為 datatype. 實(shí)驗(yàn)環(huán)境：使用筆記本電腦登錄實(shí)驗(yàn)室服務(wù)器。具體代碼如下：#include<stdio.h>#include<stdlib.h>#include<

20、;math.h>#include<time.h>#include<mpi.h>void read_num(long long int *num_point,int my_rank,MPI_Comm comm);void compute_pi(long long int num_point,long long int* num_in_cycle,long long int* local_num_point,int comm_sz,long long int *total_num_in_cycle,MPI_Comm comm,int my_rank);int main

21、(int argc,char* argv)longlongintnum_in_cycle,num_point,total_num_in_cycle,local_num_point;int my_rank,comm_sz;MPI_Comm comm;MPIni t(NULL,NULL);/初始化comm=MPI_COMM_WORLD;MPI_Comm_size(comm,&comm_sz);/得到進(jìn)程總數(shù)MPI_Comm_ra nk(comm,&my_ran k);得到進(jìn)程編號(hào) read_num(&num_point,my_rank,comm);/ 讀取輸入數(shù)據(jù)compu

22、te_pi(num_point,&num_in_cycle,&local_num_point,comm_sz,&total_num_in _cycle,comm,my_rank);MPI_Finalize();return 0;void read_num(long long int* num_point,int my_rank,MPI_Comm comm) if(my_rank=0)printf("please input num in sqaure n"); scanf("%lld",num_point);/*廣播函數(shù)int M

23、PI_Bcast(void* data_p /in/outint count /inMPI_Datatype datatype /in int source_proc /in MPI_Comm comm /in) */MPI_Bcast(num_point,1,MPI_LONG_LONG,0,comm);void compute_pi(long long int num_point,long long int* num_in_cycle,long long int* local_num_point,int comm_sz,long long int *total_num_in_cycle,MP

24、I_Comm comm,int my_rank)*num_in_cycle=0; *local_num_point=num_point/comm_sz; double x,y,distance_squared; srand(time(NULL);for(long long int i=0;i< *local_num_point;i+) x=(double)rand()/(double)RAND_MAX; x=x*2-1; y=(double)rand()/(double)RAND_MAX; y=y*2-1; distance_squared=x*x+y*y; if(distance_sq

25、uared<=1) *num_in_cycle=*num_in_cycle+1; /* 全局函數(shù) MPI_Reduce( void* input_data_p /in void* output_data_p /out int count /in MPI_Datatype datatype /in MPI_Op oprtator /in int dest_process /in MPI_Comm comm /in) */MPI_Reduce(num_in_cycle,total_num_in_cycle,1,MPI_LONG_LONG,MPI_SUM,0,co mm);if(my_rank

26、=0)double pi=(double)*total_num_in_cycle/(double)num_point*4; printf("the estimate value of pi is %lfn",pi);4.實(shí)驗(yàn)結(jié)果與分析實(shí)驗(yàn)結(jié)果如下：pppuser?3-0gnode231 lati斗玉 f斗 input square nuntjer:1600©the estimate of pi is 久"56W Time: 9.19pppuser23flnode 151 lat4f . f4- inpLi七 square number :±he

27、 estimate 口F pi. is 3.1413&0 lime: 0.13由結(jié)果可知循環(huán)次數(shù)越多得到的pi值就越精確。與實(shí)驗(yàn)二和實(shí)驗(yàn)三相比, 相同循環(huán)次數(shù)下采用mpi運(yùn)算速度更快。實(shí)驗(yàn)五1實(shí)驗(yàn)?zāi)康呐c要求1.understand deeply the architecture of GPGPU and master the CUDA programming model2.implement the parallel algorithm of calculating the value of pi using CUDA3. carries on the simple analysis

28、and summary of the program execution results4. propose optimization solution based on the execution results and hardware environment5. compare it with the results of Lab2 ,Lab3 and Lab42. 算法描述采用積分法計(jì)算 pi 值：積分法計(jì)算pi值的基本思想是利用1/(1+xA2)的原函數(shù)為arctanx，再利用 積分的基本步驟：分割，求和，取極限。將函數(shù)圖形與Y軸和直線X=1圍成的面積盡可能細(xì)分，然后求和，最后乘以相

29、應(yīng)常數(shù)，可以得到一個(gè)非常近似 的 pi 值。3. 實(shí)驗(yàn)方案CUDA在執(zhí)行的時(shí)候是讓host里面的一個(gè)一個(gè)的kernel按照線程網(wǎng)格的概 念在顯卡硬件(GPU上執(zhí)行。每一個(gè)線程網(wǎng)格又可以包含多個(gè)線程塊( block )，每一個(gè)線程塊中又可以包含多個(gè)線程( thread )?；?API 如下：cudaError_t cudaMalloc (void *devPtr,size_tsize );在設(shè)備端分配 size 大小的空間，起始地址為 devPtrcudaError_t cudaMemcpy (void * dst, const void * src,size_tcount,enum cuda

30、MemcpyKind kind);將以 src 為地址長(zhǎng)度為 count 的數(shù)據(jù)賦值到 dst 為起始地址的 存區(qū)域中，常用的 kind 有 cudaMemcpyHostToDevice， cudaMemcpyDeviceToHostcudaError_t cudaFree (void *devPtr);在設(shè)備端清理以 devPtr 為起始地址的存空間將任務(wù)合理的分配到 grid 和 thread 中，有助于提升程序的性能：grid of thread ：Kemtl Opel onGlX*60-> CORESIJunber af Bloc ks ?513rnre;)dlrt>i、

31、32 Threads vvra sizeHIOCH1loochdi'- x-'DicctOim x具體實(shí)驗(yàn)代碼如下:#in elude <stdio.h>global_ void kern el(double *gpu_p) int gpu_co unt = 0;for (i nt gpu_i = 1; gpu_i <= 1000; gpu_i+) if (gpu_cou nt % 2 = 0)gpu_pgpu_i= gpu_pgpu-1 - 4 / (double)(2*gpu-1); else gpu_pgpu_i=gpu_pgpu_i-1 + 4 / (

32、double)(2*gpu_i-1);gpu_co unt = gpu_co unt + 1; 一 一_syn cthreads();int main (i nt argc, char *argv) int i;int coun t=0;double p100;double *g_p;int block_size = 32; const int N = 1000;int n_blocks = N/block_size + (N%block_size = 0 ? 0:1); p0 = 0;g_p= (double *)malloc(1000 *sizeof(double); cudaMalloc

33、(void *) &g_p,1000 * sizeof(double);cudaMemcpy(g_p, p,1000 * sizeof(double), cudaMemcpyHostToDevice);kernel <<< n_blocks, block_size >>>( g_p);cudaMemcpy(p, g_p,1000 * sizeof(double), cudaMemcpyDeviceToHost);cudaFree(g_p);for (i = 1; i <= 1000; i+) printf( "%2.4ft"

34、, pi);printf( "blocks = %d block size= %dt",n_blocks, block_size ); system("pause");return 0;4. 實(shí)驗(yàn)結(jié)果與分析實(shí)驗(yàn)結(jié)果如下：-3,1547-3.14B4-3.154B-/=叫呂孑J.1349-Hl4呂2-5.13 5-6'3.14753.1357-3.1474-3+135&-3.1474-3.1358-3,1473-3.13593.1468-3.1364-3.1457-3+1365-3.1467-3.1365-3.1466-3.13663.14

35、52-3.1370-3.1462-3.1370-3.14&1-3*1371-3.1461-3.13713.1457-3.1375-3.1457*3.1375-3.1457-S.1375-3.1456-3,13763.1454-3.137S-3-1453-3.1379-3.1453-3.1379-3.1453-3.13793.14-51-3.13S1-3.1459-3.13B2-3.1450-3.1382-3.1450-5.13B23,1440-3.13&4-3.1440-3*1364-3.1447-3.1364(1447-3.13553.144&-耳1386-3.144

36、5-3*13B71445-3*1387-31445-3.13073.1444-3.1MB-3,1443-3>130&-314斗了-3.1389-3.1443-3.13593.1442-3.1442-3.1390-3.1442-3.1390-3.1442-3.13M3.1440-3.1392-3.144&-3.1392-3.1440-3.1392-3.1446-3,13923.1439-3.139i-3.1435-3.1393-1.1459-3.1439-3.1393-3,1394-3.1453-3.1394-3.1430-3.1394-3,1433-5-13943.143

37、7'3.1395-3.1437-3+1395-.1437-3.1395-3.1436-3.13953.143&-3,143&-3.1395-3,1436-5.1396-3.1435-3.15963.1435-3.1397-3.1435-3*1397-3.1435-3*1397-3.1435-3.13973.1434-3.1593-3.1434*3.1393-3.1434-1.1398-3.1434*3.139B3.1433-3.139&-3,. 1433-3.1399-3 1433-3.1399-3.1433-3.13993.1433-3.1599-3.1435

38、-3.1599-3.1452-3.1393-竄 1432-3.13953,1432-3.14W-3.1432-3+1400-3,1432-3.14W-3.1432-3.14&e3.1431-3.1401-31431-3.1491-31431-3.1461-3.1431-31401弓.1431-3.1401-3.1431-3+1401-3.14S1-3.14C1-3.1431-3,14013.1430-3.14&2-3.143ft-3,14023.1430-3.14C2-3.1430-3.14023.1436*3.1402-3.1430-3.1402-3.143®-3.

39、1462-3.143®-3.14023.14-29-3.14C3-3.1429-3.1403-3.1429-3.14C3-3.1429-3.14033,1+29-3.1403-3.1429-3*1403-3.1429-3.1405-5,1429-3,14«351429-3.1W-3.1423'5.1 4W-3.1428-5il4«3-3h142S-孟昭如3.142S-3.1444-3,1428-3+1404-3.1420-3,1464-3.1428-3.14A43.1428-3.1404-3.1428-3.1404-3.1428-3.1464-3.1428

40、-3,14&43.1427-3.1404-3.1427-3.1404-3.1427-3.14©4-3.1427-3.14*43.1427-3.1405-3.1427-3.1斗匪-3.1427-3.1427-5.1斗05與前三次實(shí)驗(yàn)采用的蒙特卡羅法相比，積分法的精度更高，使用cuda計(jì)算pi值，速度更快。PR0JECT2AIM:master the two typical parallel program developme nt tools (Ope nMPand MPI)un dersta nd the similarities and differe nces betwee n the two toolsduri ng the process of parallel program des