amao992的专栏

1. 介绍我们常通过CPU 进程绑定（binding or affinity）的方法来提高MPI 程序的性能。通过CPU 进程绑定，可以避免进程在CPU 核之间切换带来的开销，可以减轻cache 争抢现象。特别是当进程数为CPU 总核数一半左右时，有时会发现测试结果不稳定，时好时坏，很可能是因为进程切换造成的，这时不妨尝试进行进程CPU 绑定。MPI 程序的进程CPU 绑定可以

http://emonkey.blog.sohu.com/166546157.html参与排序的MPI进程首先对输入文件进行划分，以确定每个进程计算的文件偏移量和长度，然后通过MPI-IO读取本进程需要排序的部分每个进程对要处理的数据用qsort进行本地排序，然后收集本进程的最小键值和最大键值，min_key, max_key，通过MPI_Allreduce()调用，每个进程获得全局的

http://emonkey.blog.sohu.com/166467283.html1.基本概念进程私有文件指针：每个参与IO的进程都有一个私有的文件指针，进程间对该指针完全独立进程共享文件指针：所有参与IO的进程共享一个文件指针，任何进程对该指针的操作，都会同步更新其他进程中该指针文件视图：每个参与IO的进程都可以定义自有的文件视图，以表示本进程需读写文件哪些部分，以及这些

I was always confused with the parameters used in ./configure command. I did not know how to find the exact meaning of these parameters. Yesterday, I finally found the secrete: --help. After typin

MPI_FILE_OPEN(comm, filename, amode, info, fh)IN　　　comm　　　　组内通信域IN　　　filename　　将打开的文件名IN　　　amode　　　 打开方式IN　　　info　　　　传递给运行时的信息OUT　　 fh　　　　　返回的文件句柄int MPI_File_open(MPI_Comm comm, char * filename

/** * int MPI_File_open(MPI_Comm comm, char *filename, int amode, MPI_Info info, MPI_File *fh); * int MPI_File_close(MPI_File *fh); * * int MPI_File_write_shared(MPI_File fh, void *buf, int count,

借助MPI功能可以提供一种分布式数据服务。多个服务节点通过MPI_COMM形式组成一个逻辑服务器，维护一个“逻辑”上的数据文件，数据文件可能会分布存储在每个服务节点本地盘上，客户端可以从服务节点群获取数据！好处：每个服务节点可以利用本地磁盘和内存，提高响应时间适合：数据只读操作，e.g.,按关键字浏览相关函数：MPI_Comm_s

#include#include#include#include"mpi.h"#include #include /** * fh is shared by all processes. * In case of reading, after one process finishes, and the fh moves to the new position, * t

#include#include#include"mpi.h"/** * int MPI_File_read_at(MPI_File fh, MPI_Offset offset, void * buf, int count, MPI_Datatype datatype, MPI_Statye * status); * int MPI_File_write_at(MPI_File fh,

#include#include#include#include"mpi.h"/** * all process read/write at the same offset * int MPI_File_read_at_all(MPI_File fh, MPI_Offset offset, void *buf, int count, MPI_Datatype datatype, MP

#include#include#include#include"mpi.h"/** * all process read/write at the same offset * int MPI_File_iread_at(MPI_File fh, MPI_Offset offset, void * buf, int count, MPI_Datatype datatype, MPI_

#include#include#include#include"mpi.h"/** * non-blocking read/write in the context of fileview in two stage * because *_end equals to MPI_Wait(&request, &status), MPI_Wait is not necessary *

#include#include#include#include"mpi.h"/** * non-blocking read/write in the context of fileview * int MPI_File_iread(MPI_File fh, void * buf, int count,Datatype datatype, MPI_Request * request)

#include#include#include#include"mpi.h"/** * all process exe blocking read/write in the context of fileview * int MPI_File_read_all(MPI_file fh, void * buf, int count, MPI_Datatype datatype, MPI

#include#include#include#include"mpi.h"/** * blocking read/write in the context of a specific process's fileview * * int MPI_File_read(MPI_file fh, void * buf, int count, MPI_Datatype datatype,

#include#include#include#include"mpi.h"/** * int MPI_File_seek(MPI_File fh, MPI_Offset offset, int whence) * here whence can be: * MPI_SEEK_SET, point to offset * MPI_SEEK_CUR, point to c

fileview + seek  能组合出非常强大的功能，首先由fileview来抽象进程自己的文件视图，然后再在新的fileview逻辑视图中seek想要达到的位置。而不关心物理位置。#include#include#include#include"mpi.h"/** * int MPI_File_seek(MPI_File fh, MPI_Offset offset,

#include#include#include#include"mpi.h"/** * int MPI_File_set_view(MPI_File fh, MPI_Offset disp, MPI_Datatype etype, MPI_Datatype filetype, char * datarep, MPI_Info info); * etype = basic uni

#include#include#include#include"mpi.h"/** * two phase non-blocking read and write * * int MPI_File_read_at_all_begin(MPI_File fh, MPI_Offset offset, void * buf, int count, MPI_Datatype dataty

#include#include#include#include"mpi.h"#include #include /** * fh is shared by all processes. * In case of reading, after one process finishes, and the fh moves to the new position, * t

todo

todo

1. 程序源代码#include#include"mpi.h"int main(int argc, char *argv[]){ int totalTaskNum, rankID; int rt = MPI_Init(&argc, &argv); if(rt != MPI_SUCCESS){ pri

1. 根据Prog1改写 #include #include"mpi.h" int main(int argc, char *argv[]){ int totalTaskNum, rankID; int rt = MPI_Init(&argc, &argv);

1. 源码#include#include"mpi.h"int main(int argc, char *argv[]){ int totalTaskNum, rankID; int rt = MPI_Init(&argc, &argv); if(rt != MPI_SUCCESS){

1. 非阻塞(non-blocking)通信(1)非阻塞的sender(a)几乎可以立刻返回去干别的事情。不管数据是否从application buffer到了system buffer， 或者数据到了receiver端的application buffer或者system buffer.(b)发送操作由MPI lib选择合适的时间去完成。(c)sender最好不要去立刻修

All or None:Collective communication must involve all processes in the scope of a communicator. All processes are by default, members in the communicator MPI_COMM_WORLD.It is the programmer'

1.将openmpi-1.5.tar.gz 文件拷贝到一个临时的目录里面（如tem）2.解压文件：$tar -zxvf openmpi-1.5.tar.gz3.进入解压后的目录：$cd openmpi-1.54.$ ./configure  --prefix=/home/bjwang/software/openmpi-1.5 CC=icc CXX=icpc F77=ifort FC

1.配置环境(1) 在两台机器上分别配置/etc/hosts  (root用户)192.168.100.37   amao991    （主节点）192.168.100.50   amao992      (从节点)(2)  在两台机器上分别创建amao用户(3)  配置NFS从amao992上mount主节点amao991上/home目录，设为

1. MPI_Reduce函数int MPI_Reduce(void *sendBuf, void *receiveBuf, int count, MPI_Datatype dataType,                                MPI_Op operator, int root, MPI_Comm comm)每个进程从sendBuf向root进程的rec

1. 函数 MPI_ScatterMPI_Scatter (&sendbuf,sendcnt,sendtype,&recvbuf,recvcnt,recvtype,root,comm)Distributes distinct messages from a single source task to each task in the group.2. 举例#in

1. 连续的构造数据类型(1)把同一内置数据类型／连续的多个元素组合成一个构造数据类型2. 举例#include#include"mpi.h"#define SIZE 4int main(int argc, char *argv[]){ int totalNumTasks, rankID; float sendBuf[SIZE][S

1. Blocking send / Blocking receive(1)阻塞发送(blocking send)是指sender把消息放入application buffer后，不能返回继续其他操作。直到这个application buffer可修改后。这个时刻有两种可能：(a)异步(asynchronous)阻塞发送:  消息从sender的application bu

todo

General ConceptsTypes of Point-to-Point Operations:MPI point-to-point operations typically involve message passing between two, and only two, different MPI tasks. One task is performing a send o