This lesson is still being designed and assembled (Pre-Alpha version)

Running MPI parallel jobs using Singularity containers

Overview

Teaching: 30 min
Exercises: 20 min
Questions
  • How do I set up and run an MPI job from a Singularity container?

Objectives
  • Learn how MPI applications within Singularity containers can be run on HPC platforms

  • Understand the challenges and related performance implications when running MPI jobs via Singularity

Running MPI parallel codes with Singularity containers

MPI overview

MPI - Message Passing Interface - is a widely used standard for parallel programming. It is used for exchanging messages/data between processes in a parallel application. If you’ve been involved in developing or working with computational science software, you may already be familiar with MPI and running MPI applications.

When working with an MPI code on a large-scale cluster, a common approach is to compile the code yourself, within your own user directory on the cluster platform, building against the supported MPI implementation on the cluster. Alternatively, if the code is widely used on the cluster, the platform administrators may build and package the application as a module so that it is easily accessible by all users of the cluster.

An important aspect of MPI is the support for PMI - Parallel Process-Management Interface which allows MPI programs to be run with many different process managers that actually run the MPI code. This reduces the overhard of having to match MPI libraries at runtime with what the code was built with at buildtime.

MPI codes with Singularity containers

We’ve already seen that building Singularity containers can be impractical without root access. Since we’re highly unlikely to have root access on a large institutional, regional or national cluster, building a container directly on the target platform is not normally an option.

If our target platform uses OpenMPI, one of the two widely used source MPI implementations, we can build/install a compatible OpenMPI version on our local build platform, or directly within the image as part of the image build process. We can then build our code that requires MPI, either interactively in an image sandbox or via a definition file.

If the target platform uses a version of MPI based on MPICH, the other widely used open source MPI implementation, there is ABI compatibility between MPICH and several other MPI implementations. In this case, you can build MPICH and your code on a local platform, within an image sandbox or as part of the image build process via a definition file, and you should be able to successfully run containers based on this image on your target cluster platform.

As described in Singularity’s MPI documentation, support for both OpenMPI and MPICH is provided. Instructions are given for building the relevant MPI version from source via a definition file and we’ll see this used in an example below.

While building a container on a local system that is intended for use on a remote HPC platform does provide some level of portability, if you’re after the best possible performance, it can present some issues. The version of MPI in the container will need to be built and configured to support the hardware on your target platform if the best possible performance is to be achieved. Where a platform has specialist hardware with proprietary drivers, building on a different platform with different hardware present means that building with the right driver support for optimal performance is not likely to be possible. This is especially true if the version of MPI available is different (but compatible). Singularity’s MPI documentation highlights two different models for working with MPI codes. The hybrid model that we’ll be looking at here involves using the MPI executable from the MPI installation on the host system to launch singularity and run the application within the container. The application in the container is linked against and uses the MPI installation within the container which, in turn, communicates with the MPI daemon process running on the host system.

In this section we will be using the srun command as an alternate launcher to take advantage of PMI.

Building and running a Singularity image for an MPI code

Building and testing an image

This example makes the assumption that you’ll be building a container image on a local platform and then deploying it to a cluster with a different but compatible MPI implementation (using PMI). See Singularity and MPI applications in the Singularity documentation for further information on how this works.

We’ll build an image from a definition file. Containers based on this image will be able to run MPI benchmarks using the OSU Micro-Benchmarks software.

In this example, the target platform is a remote HPC cluster that uses Intel MPI and Slurm. The container can be built via methods used in the previous episode of the Singularity material.

Begin by creating a directory, within that directory save the following definition file content to a .def file, e.g. osu_benchmarks.def:

Bootstrap: docker
From: ubuntu:20.04

%environment
    export SINGULARITY_MPICH_DIR=/usr

%post
    apt-get -y update && DEBIAN_FRONTEND=noninteractive apt-get -y install build-essential libfabric-dev libibverbs-dev gfortran wget autoconf
    cd /root
    wget http://www.mpich.org/static/downloads/3.3.2/mpich-3.3.2.tar.gz
    tar zxvf mpich-3.3.2.tar.gz && cd mpich-3.3.2
    echo "Configuring and building MPICH..."
    ./configure --prefix=/usr --with-device=ch3:nemesis:ofi && make -j2 && make install
    cd /root
    wget https://mvapich.cse.ohio-state.edu/download/mvapich/osu-micro-benchmarks-5.7.tar.gz
    tar zxvf osu-micro-benchmarks-5.7.tar.gz
    cd osu-micro-benchmarks-5.7/
    echo "Configuring and building OSU Micro-Benchmarks..."
    autoreconf -vif
    ./configure --prefix=/usr/local/osu CC=/usr/bin/mpicc CXX=/usr/bin/mpicxx
    make -j2 && make install

%runscript
    echo "Rank ${PMI_RANK} - About to run: /usr/local/osu/libexec/osu-micro-benchmarks/mpi/$*"
    exec /usr/local/osu/libexec/osu-micro-benchmarks/mpi/$*

A quick overview of what the above definition file is doing:

Note that base path of the the executable to run is hardcoded in the run script so the command line parameter to provide when running a container based on this image is relative to this base path, for example, startup/osu_hello, collective/osu_allgather, pt2pt/osu_latency, one-sided/osu_put_latency.

Build and test the OSU Micro-Benchmarks image

Using the above definition file, build a Singularity image named osu_benchmarks.sif.

Once you have built the image, use it to run the osu_hello benchmark that is found in the startup benchmark folder.

Solution

You should be able to build an image from the definition file as follows:

$ singularity build --remote osu_benchmarks.sif osu_benchmarks.def

If instead, you’re running the Singularity Docker container directly from the command line to undertake your build, you’ll need to provide the full path to the .def file at which it appears within the container - for example, if you’ve bind mounted the directory containing the file to /home/singularity within the container, the full path to the .def file will be /home/singularity/osu_benchmarks.def._

Assuming the image builds successfully, you can then try running the container locally and also transfer the SIF file to a cluster platform that you have access to (that has Singularity installed) and run it there.

Let’s begin with a single-process run of osu_hello on the system to ensure that we can run the container as expected:

$ srun -n 1 -p c_compute_mdi1 --reservation=training -A scw1148 --pty bash
$ srun singularity run osu_benchmarks.sif startup/osu_hello

You should see output similar to the following:

Rank  - About to run: /usr/local/osu/libexec/osu-micro-benchmarks/mpi/startup/osu_hello
# OSU MPI Hello World Test v5.7
This is a test with 1 processes

If run outside of a cluster, note that no rank number is shown since we didn’t run the container via mpirun and so the ${PMI_RANK} environment variable that we’d normally have set in an MPICH run process is not set.

Running Singularity containers via MPI

Assuming the above tests worked, we can now try undertaking a parallel run of one of the OSU benchmarking tools within our container image.

This is where things get interesting and we’ll begin by looking at how Singularity containers are run within an MPI environment.

If you’re familiar with running MPI codes, you’ll know that you use mpirun, mpiexec or a similar MPI executable to start your application. This executable may be run directly on the local system or cluster platform that you’re using, or you may need to run it through a job script submitted to a job scheduler. Your MPI-based application code, which will be linked against the MPI libraries, will make MPI API calls into these MPI libraries which in turn talk to the MPI daemon process running on the host system. This daemon process handles the communication between MPI processes, including talking to the daemons on other nodes to exchange information between processes running on different machines, as necessary.

When running code within a Singularity container, we don’t use the MPI executables stored within the container (i.e. we DO NOT run singularity exec mpirun -np <numprocs> /path/to/my/executable). Instead we use the MPI installation on the host system or the process manager in a scheduler such as Slurm that supports PMI, to run Singularity and start an instance of our executable from within a container for each MPI process. Without Singularity support in an MPI implementation, this results in starting a separate Singularity container instance within each process. This can present some overhead if a large number of processes are being run on a host. Where Singularity support is built into an MPI implementation this can address this potential issue and reduce the overhead of running code from within a container as part of an MPI job.

Ultimately, this means that our running MPI code is linking to the MPI libraries from the MPI install within our container and these are, in turn, communicating with the MPI daemon on the host system which is part of the host system’s MPI installation. These two installations of MPI may be different but as long as there is ABI compatibility between the version of MPI installed in your container image and the version on the host system, your job should run successfully. If running with srun this is solved by PMI support in-built into the scheduler.

We can now try running a 2-process MPI run of a point to point benchmark osu_latency. If your local system has both MPI and Singularity installed and has multiple cores, you can run this test on that system. Alternatively you can run on a cluster. Note that you may need to submit this command via a job submission script submitted to a job scheduler if you’re running on a cluster (and use the scheduler’s PMI support). If you’re attending a taught version of this course, some information will be provided below in relation to the cluster that you’ve been provided with access to.

Undertake a parallel run of the osu_latency benchmark (general example)

Move the osu_benchmarks.sif Singularity image onto the cluster (or other suitable) platform where you’re going to undertake your benchmark run.

You should be able to run the benchmark using a command similar to the one shown below. However, if you are running on a cluster, you may need to write and submit a job submission script at this point to initiate running of the benchmark.

$ srun singularity run osu_benchmarks.sif pt2pt/osu_latency

Expected output and discussion

As you can see in the mpirun command shown above, we have called mpirun on the host system and are passing to MPI the singularity executable for which the parameters are the image file and any parameters we want to pass to the image’s run script, in this case the path/name of the benchmark executable to run.

The following shows an example of the output you should expect to see. You should have latency values shown for message sizes up to 4MB.

Rank 1 - About to run: /.../mpi/pt2pt/osu_latency
Rank 0 - About to run: /.../mpi/pt2pt/osu_latency
# OSU MPI Latency Test v5.7
# Size          Latency (us)
0                       0.38
1                       0.34
...

Undertake a parallel run of the osu_latency benchmark (taught course cluster example)

This version of the exercise for undertaking a parallel run of the osu_latency benchmark with your Singularity container that contains an MPI build is specific to this run of the course.

The information provided here is specifically tailored to the HPC platform that you’ve been given access to for this taught version of the course.

Move the osu_benchmarks.sif Singularity image onto the cluster where you’re going to undertake your benchmark run. You should use scp or a similar utility to copy the file.

The platform you’ve been provided with access to uses Slurm schedule jobs to run on the platform. You now need to create a Slurm job submission script to run the benchmark.

Download this template script and edit it to suit your configuration.

Submit the modified job submission script to the Slurm scheduler using the sbatch command.

$ sbatch osu_latency.slurm

Expected output and discussion

As you will have seen in the commands using the provided template job submission script, we have called mpirun on the host system and are passing to MPI the singularity executable for which the parameters are the image file and any parameters we want to pass to the image’s run script. In this case, the parameters are the path/name of the benchmark executable to run.

The following shows an example of the output you should expect to see. You should have latency values shown for message sizes up to 4MB.

INFO:    Convert SIF file to sandbox...
INFO:    Convert SIF file to sandbox...
Rank 1 - About to run: /.../mpi/pt2pt/osu_latency
Rank 0 - About to run: /.../mpi/pt2pt/osu_latency
# OSU MPI Latency Test v5.7
# Size          Latency (us)
0                       1.49
1                       1.50
2                       1.50
...
4194304               915.44
INFO:    Cleaning up image...
INFO:    Cleaning up image...

This has demonstrated that we can successfully run a parallel MPI executable from within a Singularity container. However, in this case, the two processes will almost certainly have run on the same physical node so this is not testing the performance of the interconnects between nodes.

You could now try running a larger-scale test. You can also try running a benchmark that uses multiple processes, for example try collective/osu_gather.

Investigate performance when using a container image built on a local system and run on a cluster

To get an idea of any difference in performance between the code within your Singularity image and the same code built natively on the target HPC platform, try building the OSU benchmarks from source, locally on the cluster. Then try running the same benchmark(s) that you ran via the singularity container. Have a look at the outputs you get when running collective/osu_gather or one of the other collective benchmarks to get an idea of whether there is a performance difference and how significant it is.

Try running with enough processes that the processes are spread across different physical nodes so that you’re making use of the cluster’s network interconnects.

What do you see?

Discussion

You may find that performance is significantly better with the version of the code built directly on the HPC platform. Alternatively, performance may be similar between the two versions.

How big is the performance difference between the two builds of the code?

What might account for any difference in performance between the two builds of the code?

If performance is an issue for you with codes that you’d like to run via Singularity, you are advised to take a look at using the bind model for building/running MPI applications through Singularity.

Key Points

  • Singularity images containing MPI applications can be built on one platform and then run on another (e.g. an HPC cluster) if the two platforms have compatible MPI implementations.

  • When running an MPI application within a Singularity container, use the MPI executable on the host system to launch a Singularity container for each process.

  • Think about parallel application performance requirements and how where you build/run your image may affect that.