mpi

MPI

用途

指导 Agent 进行 MPI（消息传递接口）编程：点对点和集合通信、非阻塞操作、子通信器、MPI + OpenMP 混合模式、使用 mpirun 启动进程、调试技术、MPI-IO 以及常见性能问题。

适用场景

• 在多个节点或插槽间进行并行化
• 实现分布式算法（矩阵分解、FFT）
• 将 MPI 进程并行与 OpenMP 线程并行相结合
• 使用 Slurm/PBS + mpirun 运行 HPC 作业
• 调试死锁和消息不匹配问题
• 使用 MPI-IO 进行并行文件 I/O

工作流

1. 最简 MPI 程序

#include <mpi.h>
#include <stdio.h>

int main(int argc, char **argv) {
    MPI_Init(&argc, &argv);

    int rank, size;
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Comm_size(MPI_COMM_WORLD, &size);

    printf("Hello from rank %d of %d\n", rank, size);

    MPI_Finalize();
    return 0;
}

mpicc -o hello hello.c
mpirun -np 4 ./hello
# or
mpiexec -n 4 ./hello

2. 点对点通信

if (rank == 0) {
    int data = 42;
    MPI_Send(&data, 1, MPI_INT, 1, 0, MPI_COMM_WORLD);
} else if (rank == 1) {
    int recv;
    MPI_Recv(&recv, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
    printf("rank 1 got %d\n", recv);
}

带标签的消息：为 MPI_Recv 匹配标签和源进程。

3. 集合通信

int local = rank + 1;
int global_sum;

MPI_Allreduce(&local, &global_sum, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);

// Broadcast
if (rank == 0) data = 100;
MPI_Bcast(&data, 1, MPI_INT, 0, MPI_COMM_WORLD);

// Scatter/Gather
MPI_Scatter(sendbuf, sendcount, MPI_INT, recvbuf, recvcount, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Gather(sendbuf, sendcount, MPI_INT, recvbuf, recvcount, MPI_INT, 0, MPI_COMM_WORLD);

集合通信	用途
MPI_Bcast	一对多广播
MPI_Scatter	分发数据块
MPI_Gather	收集数据块
MPI_Allreduce	归约并广播结果
MPI_Barrier	同步
MPI_Alltoall	全互换

4. 非阻塞通信

MPI_Request req;
MPI_Isend(buf, count, MPI_INT, dest, tag, MPI_COMM_WORLD, &req);
// overlap computation here
do_local_work();
MPI_Wait(&req, MPI_STATUS_IGNORE);

// Multiple requests
MPI_Request reqs[2];
MPI_Irecv(buf0, n, MPI_INT, 0, 0, comm, &reqs[0]);
MPI_Irecv(buf1, n, MPI_INT, 1, 0, comm, &reqs[1]);
MPI_Waitall(2, reqs, MPI_STATUSES_IGNORE);

通过将通信与计算重叠来隐藏延迟。

5. 子通信器

int color = rank / 4;  // groups of 4
MPI_Comm subcomm;
MPI_Comm_split(MPI_COMM_WORLD, color, rank, &subcomm);

int subrank, subsize;
MPI_Comm_rank(subcomm, &subrank);
MPI_Comm_size(subcomm, &subsize);

MPI_Comm_free(&subcomm);

6. MPI + OpenMP 混合模式

#pragma omp parallel
{
    int tid = omp_get_thread_num();
    // thread-local work on rank's data partition
}
MPI_Barrier(MPI_COMM_WORLD);
MPI_Allreduce(...);

export OMP_NUM_THREADS=4
mpirun -np 8 --bind-to core ./hybrid_app
# 8 ranks × 4 threads = 32 cores

使用 --map-by ppr:2:socket 将各个 rank 绑定到插槽。

7. 使用 hostfile 启动

# hostfile:
# node0 slots=4
# node1 slots=4

mpirun -np 8 --hostfile hosts.txt ./app

# Slurm integration
srun -n 64 ./app
# or
mpirun -np $SLURM_NTASKS ./app

# Debug: tag output by rank
mpirun -np 4 --tag-output ./app

# Sequential debug (one rank at a time)
mpirun -np 4 -gdb ./app

8. MPI-IO

#include <mpi.h>

MPI_File fh;
MPI_File_open(MPI_COMM_WORLD, "output.dat",
    MPI_MODE_CREATE | MPI_MODE_WRONLY, MPI_INFO_NULL, &fh);

MPI_Offset offset = rank * chunk_size;
MPI_File_write_at(fh, offset, buf, count, MPI_DOUBLE, MPI_STATUS_IGNORE);

MPI_File_close(&fh);

使用集合 I/O 提升性能：

MPI_File_write_at_all(fh, offset, buf, count, MPI_DOUBLE, MPI_STATUS_IGNORE);

9. 性能问题

Common bottlenecks
├── Load imbalance → dynamic scheduling (OpenMP) or redistribute MPI chunks
├── Serialization at rank 0 → tree-based reduce, parallel I/O
├── Excessive sync → replace Barrier with point-to-point where possible
├── Small messages → aggregate; use MPI_Pack or larger blocks
└── Alltoall on large process counts → consider MPI neighborhood collectives

# MPI profiling
mpiP  # lightweight profiler
# or IPM, TAU MPI wrappers

常见问题

症状	原因	解决方法
在 MPI_Recv 处挂起	标签或源进程不匹配	检查 Send/Recv 配对；调试时使用 MPI_ANY_TAG
死锁	循环等待	调整通信顺序；使用非阻塞操作
Allreduce 结果错误	数据类型或计数错误	确认使用 MPI_INT 还是 MPI_DOUBLE
扩展性差	rank 0 成为瓶颈	分散 I/O 和聚合操作
MPI_ERR_TRUNCATE	接收缓冲区太小	确保发送和接收计数匹配
混合模式超额订阅	线程数 × rank 数过多	OMP_NUM_THREADS = cores/ranks

相关技能

• skills/hpc/openmp — MPI rank 内的线程级并行
• skills/hpc/rdma-verbs — MPI 下的低延迟互连
• skills/allocators/numa-programming — 将 rank 绑定到 NUMA 节点
• skills/profilers/linux-perf — 分析 MPI rank 的热点
• skills/debuggers/gdb — 调试单个 MPI 进程
• skills/compilers/gcc — MPI 编译器封装器标志