MPI programming

Message passing programming

Basic operations:

Pairwise messaging: send/receive
Collective messaging: broadcast, scatter/gather
Collective computation: parallel prefix (sum, max)
Barriers (no need for locks!)
Environmental inquiries (who am I? do I have mail?)

(Much of what follows is adapted from Bill Gropp.)

The Message Passing Interface (or MPI) is a big interface with a number of different types of operations. Today, we'll talk about five main ones. First, there's pairwise messaging: point-to-point data sends and receives. Then there's collective messaging operations that involve several senders and receivers simultaneously. There is also some support for simple types of collective computation operations, like computing a sum with contributions from all processors (called a reduction operation) or computing a list of partial sums (called a scan or parallel prefix operation). Then there are synchronization mechanisms like barriers, though their role in MPI programming is a little subtler than in shared memory programming - and if you have taken a course where you've seen synchronization with locks, you'll be happy to know that there is no need for locks here. And, last but not least, there are environmental inquiries that tell us things about the run-time setup, like how many processors are involved in a given MPI run, or the process number (or rank, in MPI lingo) of a given processor.

This slide deck has evolved some over time, but it was originally modeled pretty closely on a similar presentation given by Bill Gropp of UIUC.

MPI

Message Passing Interface
An interface spec — many implementations
(OpenMPI, MPICH, MVAPICH, Intel, ...)
Single Program Multiple Data (SPMD) style
Bindings to C, C++, Fortran

MPI is short for Message Passing Interface. The "interface" part is important. There are many different implementations out there, with OpenMPI, MPICH, and the Intel MPI being maybe the most popular. These different implementations are not guaranteed to do things the same way behind the scenes, nor to inter-operate with each other. They just implement the same programming interface.

There are several versions of the MPI programming interface, including in C, C++, and Fortran. Other languages (like Python) often have unofficial MPI interfaces, too. We can usually use MPI across language boundaries, even if we can't mix MPI implementations. For example, it's completely legitimate for a C routine to send MPI messages to a Fortran routine, though it certainly doesn't happen all that often, partly because of the nature of MPI codes. MPI programs are SPMD: Single Program, Multiple Data. It's possible that different MPI ranks can do very different things during program execution, but they all start from the same main routine.

MPI

Version 1.0 in 1994, 2.2 in 2009, 3.0 in 2012
MPI 3 goodies:
- Nonblocking collectives
- Neighborhood collectives
- RMA and one-sided comm
Will stick to MPI-2 today

Hello world

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

int main(int argc, char** argv) {
    int rank, size;
    MPI_Init(&argc, &argv);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Comm_size(MPI_COMM_WORLD, &size);
    printf("Hello from %d of %d\n", rank, size);
    MPI_Finalize();
    return 0;
}

Let's start by walking through a "hello world" example in MPI. This example, or something very similar, is in the demos/graphite-hello directory.

We start by including the MPI header file at the very beginning. The MPI header is where we have the declarations for the MPI functions. This code uses only four of those functions, but they are calls that almost every MPI program makes.

The first call is to MPI_Init, and this call always has to come first. Notice that we pass in pointers to the argc and argv inputs to main, which are used to pass in command-line parameters. That's because some MPI implementations can use argc and argv to pass in data about the MPI setup when the program gets started. The MPI_Init is supposed to take any of that extra stuff out, so you should ideally call it before you do any argument processing (for example).

The last call is to MPI_Finalize. This always has to come at the end of your MPI programs, after you've finished any communication.

The two calls in between are not required in the same way that you require the MPI_Init and MPI_Finalize calls, but they show up in most MPI codes nonetheless. MPI indexes processes by "ranks," and so MPI_Comm_rank is used to get the rank of the current process (zero-indexed) and MPI_Comm_size gets the total number of processes.

The first argument to both of these functions, MPI_COMM_WORLD, is the name of the default communicator. A communicator in MPI is a collection of processes that can communicate with each other for a different context. It's not uncommon for small MPI codes to just use MPI_COMM_WORLD for communication, and we'll use that as our default communicator argument today.

Building, queueing, running

Several steps to actually run

mpicc -o foo.x foo.c   # Compile the program
sbatch foo.sub         # Submit to queue (SLURM)
# mpirun -n 2 ./foo.x  # (in foo.sub) Run on 2 procs

This is all platform-specific.

Writing the MPI program is only the first step. Once we have the code, we have to compile it and run it!

We usually compile with mpicc or mpif90 or some similarly-named command. This builds on your usual GCC or CLang or Intel compiler behind the scenes, but does the work for you of figuring out where the MPI support libraries and header files are.

Once we have compiled the code, we use a command like mpirun or mpiexec - depends on the implementation - to actually launch the MPI job on some number of nodes. The mpirun interface gets a bunch of information both from command line variables (e.g. we use -n 2 in the example above to say that we're running on two processors) and from shell environment variables and from files.

On my laptop, I would just use mpirun directly. But if I want to run on a cluster, like the Graphite cluster or the SDSC Comet system that we will use in the class, I would typically submit a runner job to a batch queue system. The batch queue system is responsible for sharing the resources in a big cluster or supercomputer: the idea is that rather than always running tasks interactively, you write a script that says what you want to do and what resources you will need, and a scheduler figures out when the system will be able to give you those resources, and runs the code for you asynchronously. We've used the syntax here for the SLURM batch system, but there are others, like PBS and SGE. They all use different commands and work in slightly different ways, but they all have the same basic functionality: you tell the scheduler the resources, and it finds time on those resources for you so that you can run your job. The scheduler will also frequently set up the environment variables and files that the MPI implementation needs in order to know where it should run things for you. So there's a lot that goes on behind the scenes here.

Communicators

Processes form groups
Messages sent in contexts
- Separate communication for libraries
Group + context = communicator
Identify process by rank in group
Default is MPI_COMM_WORLD

Sending and receiving

Need to specify:

What’s the data?
- Different machines use different encodings (e.g. endian-ness)
- $\implies$ “bag o’ bytes” model is inadequate
How do we identify processes?
How does receiver identify messages?
What does it mean to “complete” a send/recv?

We use communicators to communicate, naturally enough, but what does communication in MPI look like? We need to specify at least three things: who is sending, who is receiving, and what is in the message. There is actually some subtlety to each of these, starting with the content of the messages. For example, suppose that we wanted to send an integer across the wire. Different systems can encode integers in different ways: for example, most modern Intel machines are "little-endian," meaning that the least significant bytes of an integer appear first in memory; but there used to be common machines that were "big-endian," meaning that the most significant bytes appear first. So if we want to send messages from a big-endian machine to a little-endian machine, we'd better re-order the bytes in our integers first. That means that we need to know not only a region in memory where the message resides, but also a bit about the meaning of the bytes in that region.

By the way, the terms "big-endian" and "little-endian" come from Gulliver's Travels. These were the names of the opposing parties in Lilliput. The Big-Endians broke their eggs at the big end; the Little-Endians, at the other end. Yes, the whole thing is obscure and a little bit silly - but that describes a lot of CS, doesn't it?

MPI datatypes

Message is (address, count, datatype). Allow:

Basic types (MPI_INT, MPI_DOUBLE)
Contiguous arrays
Strided arrays
Indexed arrays
Arbitrary structures

Complex data types may hurt performance?

MPI tags

Use an integer tag to label messages

Help distinguish different message types
Can screen messages with wrong tag
MPI_ANY_TAG is a wildcard

MPI Send/Recv

Basic blocking point-to-point communication:

int
MPI_Send(void *buf, int count,
         MPI_Datatype datatype,
         int dest, int tag, MPI_Comm comm);

int
MPI_Recv(void *buf, int count,
         MPI_Datatype datatype,
         int source, int tag, MPI_Comm comm,
         MPI_Status *status);

All right. Let's introduce the next pair of MPI commands: blocking send and receive.

MPI_Send sends a message, blocking until the message is "in the system" - whatever that might mean. We have to tell the routine where the message is in memory, how many items are in the message, the type of those items, the destination rank, the message tag, and the communicator.

MPI_Recv receives a message. The arguments look a lot like those for the send, but there are a couple differences to know about. First, we don't need to know the exact size of the data to be received in advance; the receive count just tells us the size of the buffer we are providing. Second, the source and tag can be either a processor rank and tag that we are supposed to match, or they can be wildcard values. Finally, we have an additional output argument, called status, appearing at the end of the argument list. We use this to communicate additional information about what we received.

MPI send/recv semantics

Send returns when data gets to system
- ... might not yet arrive at destination!
Recv ignores messages that mismatch source/tag
- MPI_ANY_SOURCE and MPI_ANY_TAG wildcards
Recv status contains more info (tag, source, size)

Ping-pong pseudocode

Process 0:

for i = 1:ntrials
  send b bytes to 1
  recv b bytes from 1
end

Process 1:

for i = 1:ntrials
  recv b bytes from 0
  send b bytes to 0
end

Ping-pong MPI

void ping(char* buf, int n, int ntrials, int p)
{
    for (int i = 0; i < ntrials; ++i) {
        MPI_Send(buf, n, MPI_CHAR, p, 0,
                 MPI_COMM_WORLD);
        MPI_Recv(buf, n, MPI_CHAR, p, 0,
                 MPI_COMM_WORLD, NULL);
    }
}

(Pong is similar)

Ping-pong MPI

for (int sz = 1; sz <= MAX_SZ; sz += 1000) {
    if (rank == 0) {
        double t1 = MPI_Wtime();
        ping(buf, sz, NTRIALS, 1);
        double t2 = MPI_Wtime();
        printf("%d %g\n", sz, t2-t1);
    } else if (rank == 1) {
        pong(buf, sz, NTRIALS, 0);
    }
}

Blocking and buffering

Block until data “in system” — maybe in a buffer?

Blocking and buffering

Alternative: don’t copy, block until done.

Problem 1: Potential deadlock

Both processors wait to send before they receive!
May not happen if lots of buffering on both sides.

Solution 1: Alternating order

Diagram of breaking deadlock scenario by alternate send/recv

Could alternate who sends and who receives.

Solution 2: Combined send/recv

Common operations deserve explicit support!

Combined sendrecv

MPI_Sendrecv(sendbuf, sendcount, sendtype,
             dest, sendtag,
             recvbuf, recvcount, recvtype,
             source, recvtag,
             comm, status);

Blocking operation, combines send and recv to avoid deadlock.

Problem 2: Communication overhead

Partial solution: nonblocking communication

Blocking vs non-blocking

MPI_Send and MPI_Recv are blocking
- Send does not return until data is in system
- Recv does not return until data is ready
- Cons: possible deadlock, time wasted waiting
Why blocking?
- Overwrite buffer during send $\implies$ evil!
- Read buffer before data ready $\implies$ evil!

Blocking vs non-blocking

Alternative: nonblocking communication

Split into distinct initiation/completion phases
Initiate send/recv and promise not to touch buffer
Check later for operation completion

Overlap communication and computation

Diagram of nonblocking send/receive primitive

Nonblocking operations

Initiate message:

MPI_Isend(start, count, datatype, dest
          tag, comm, request);
MPI_Irecv(start, count, datatype, dest
          tag, comm, request);

Wait for message completion:

MPI_Wait(request, status);

Test for message completion:

MPI_Test(request, status);

Multiple outstanding requests

Sometimes useful to have multiple outstanding messages:

MPI_Waitall(count, requests, statuses);
MPI_Waitany(count, requests, index, status);
MPI_Waitsome(count, requests, indices, statuses);

Multiple versions of test as well.

Other send/recv variants

Other variants of MPI_Send

MPI_Ssend (synchronous) – do not complete until receive has begun
MPI_Bsend (buffered) – user provides buffer (via MPI_Buffer_attach)
MPI_Rsend (ready) – user guarantees receive has already been posted
Can combine modes (e.g. MPI_Issend)

MPI_Recv receives anything.

Another approach

Send/recv is one-to-one communication
An alternative is one-to-many (and vice-versa):
- Broadcast to distribute data from one process
- Reduce to combine data from all processors
- Operations are called by all processes in communicator

Broadcast and reduce

MPI_Bcast(buffer, count, datatype,
          root, comm);
MPI_Reduce(sendbuf, recvbuf, count, datatype,
           op, root, comm);

buffer is copied from root to others
recvbuf receives result only at root
op $\in \{$ MPI_MAX, MPI_SUM, … $\}$

Example: basic Monte Carlo

#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>
int main(int argc, char** argv) {
    int nproc, myid, ntrials = atoi(argv[1]);
    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &nproc);
    MPI_Comm_rank(MPI_COMM_WORLD, &my_id);
    MPI_Bcast(&ntrials, 1, MPI_INT,
              0, MPI_COMM_WORLD);
    run_mc(myid, nproc, ntrials);
    MPI_Finalize();
    return 0;
}

Example: basic Monte Carlo

Let sum[0] = $\sum_i X_i$ and sum[1] = $\sum_i X_i^2$ .

void run_mc(int myid, int nproc, int ntrials) {
    double sums[2] = {0,0};
    double my_sums[2] = {0,0};
    /* ... run ntrials local experiments ... */
    MPI_Reduce(my_sums, sums, 2, MPI_DOUBLE,
               MPI_SUM, 0, MPI_COMM_WORLD);
    if (myid == 0) {
        int N = nproc*ntrials;
        double EX = sums[0]/N;
        double EX2 = sums[1]/N;
        printf("Mean: %g; err: %g\n",
               EX, sqrt((EX*EX-EX2)/N));
    }
}

The idea of Monte Carlo is that we'll run lots of independent computational experiments, and then compute a sample mean and variance over those experiments. To do this, we need to get the sum of experiment outputs and the sum of the squares of the experiment outputs. We'll keep the local contribution to these sums in the my_sums array at each processor. Then, after we have run the trials on each processor, we will reduce from these my_sums array into a sums array that lives at rank 0. Finally, at rank 0 we will use the overall sums to compute a sample mean and standard deviation.

Yes, I know that this is a biased estimator for the variance, since we used an N instead of an N-1. In Monte Carlo computations, if you notice the difference between division by N and division by N-1, you probably should have used more trials.

Collective operations

Involve all processes in communicator
Basic classes:
- Synchronization (e.g. barrier)
- Data movement (e.g. broadcast)
- Computation (e.g. reduce)

Barrier

MPI_Barrier(comm);

Not much more to say. Not needed that often.

Broadcast

Scatter/gather

Allgather

Alltoall

Reduce

Scan

The kitchen sink

In addition to above, have vector variants (v suffix), more All variants (Allreduce), Reduce_scatter, ...
MPI3 adds one-sided communication (put/get)
MPI is not a small library!
But a small number of calls goes a long way
- Init/Finalize
- Get_comm_rank, Get_comm_size
- Send/Recv variants and Wait
- Allreduce, Allgather, Bcast

We have just scratched the surface of what one can do with even MPI-2. In addition to the basic operations sketched so far, there are more All variants (like Allgather), vector variants that work with a chunk of data at a time, combined operations like reduce-scatter (a little like send-receive), and more. It's not a small library!

Fortunately, we can go a long way with a few operations like send/receive pairs and collectives like scatter, gather, broadcast, and reduction.

If I am ambitious, there may be another slide deck on Thursday about the MPI-3 operations. But the real next step for all of you is to try some things out! I will provide some demo codes for this purpose, and we'll talk through some of them in meetings this week. Until then!

CS 5220

Applications of Parallel Computers

MPI Programming