OSU Micro benchmarks

The OSU Micro Benchmarks (OMB) are a widely used suite of benchmarks for measuring and evaluating the performance of MPI operations for point-to-point, multi-pair, and collective communications. These benchmarks are often used for comparing different MPI implementations and the underlying network interconnect.

We use OMB to show that Sarus is able to provide the same native MPI high performance to containerized applications when using the native MPICH hook. As indicated in the documentation for the hook, the only conditions required are:

  • The MPI installed in the container image must comply to the requirements of the MPICH ABI Compatibility Initiative.
  • The application in the container image must be dynamically linked with the MPI libraries.

Test cases

Latency

The osu_latency benchmark measures the min, max and the average latency of a ping-pong communication between a sender and a receiver where the sender sends a message and waits for the reply from the receiver. The messages are sent repeatedly for a variety of data sizes in order to report the average one-way latency. This test allows us to observe any possible overhead from enabling the MPI support provided by Sarus.

All-to-all

The osu_alltoall benchmark measures the min, max and the average latency of the MPI_Alltoall blocking collective operation across N processes, for various message lengths, over a large number of iterations. In the default version, this benchmark report the average latency for each message length up to 1MB. We run this benchmark from a minimum of 2 nodes up to 128 nodes, increasing the node count in powers of two.

Running the container

We run the container using the Slurm Workload Manager and Sarus.

Latency

sarus pull ethcscs/mvapich:ub1804_cuda92_mpi22_osu
srun -C gpu -N2 -t2 \
 sarus run --mpi ethcscs/mvapich:ub1804_cuda92_mpi22_osu \
 /usr/local/libexec/osu-micro-benchmarks/mpi/pt2pt/osu_latency

A typical output looks like:

# OSU MPI Latency Test v5.3.2
# Size          Latency (us)
0                       1.11
1                       1.11
2                       1.09
4                       1.09
8                       1.09
16                      1.10
32                      1.09
64                      1.10
128                     1.11
256                     1.12
512                     1.15
1024                    1.39
2048                    1.67
4096                    2.27
8192                    4.21
16384                   5.12
32768                   6.73
65536                  10.07
131072                 16.69
262144                 29.96
524288                 56.45
1048576               109.28
2097152               216.29
4194304               431.85

Since the Dockerfiles use the WORKDIR instruction to set a default working directory, we can use that to simplify the terminal command:

srun -C gpu -N2 -t2 \
 sarus run --mpi ethcscs/osu-mb:5.3.2-mpich3.1.4 \
 ./osu_latency

All-to-all

srun -C gpu -N2 -t2 \
 sarus run --mpi ethcscs/osu-mb:5.3.2-mpich3.1.4 \
 ../collective/osu_alltoall

A typical outpout looks like:

# OSU MPI All-to-All Personalized Exchange Latency Test v5.3.2
# Size       Avg Latency(us)
1                       5.46
2                       5.27
4                       5.22
8                       5.21
16                      5.18
32                      5.18
64                      5.17
128                    11.35
256                    11.64
512                    11.72
1024                   12.03
2048                   12.87
4096                   14.52
8192                   15.77
16384                  19.78
32768                  28.89
65536                  49.38
131072                 96.64
262144                183.23
524288                363.35
1048576               733.93

Running the native application

We compile the OSU micro benchmark suite natively using the Cray Programming Environment (PrgEnv-cray) and linking against the optimized Cray MPI (cray-mpich) libraries.

Container images and Dockerfiles

We built the OSU benchmarks on top of several images containing MPI, in order to demonstrate the effectiveness of the MPI hook regardless of the ABI-compatible MPI implementation present in the images:

MPICH

The container image ethcscs/mpich:ub1804_cuda92_mpi314_osu (based on mpich/3.1.4) used for this test case can be pulled from CSCS DockerHub or be rebuilt with this Dockerfile.

MVAPICH

The container image ethcscs/mvapich:ub1804_cuda92_mpi22_osu (based on mvapich/2.2) used for this test case can be pulled from CSCS DockerHub or be rebuilt with this Dockerfile. On the Cray, the supported Cray MPICH ABI is 12.0 (mvapich>2.2 requires ABI/12.1 hence is not currently supported).

OpenMPI

As OpenMPI is not part of the MPICH ABI Compatibility Initiative, sarus run --mpi with OpenMPI is not supported. Documentation can be found on this dedicated page: OpenMPI with SSH launcher.

Intel MPI

Because the Intel MPI license limits general redistribution of the software, we do not share the Docker image ethcscs/intelmpi used for this test case. Provided the Intel installation files (such as archive and license file) are available locally on your computer, you could build your own image with this example Dockerfile.

Required OCI hooks

  • Native MPI hook (MPICH-based)

Benchmarking results

Latency

Consider now the following Figure that compares the average and standard deviation of the osu_latency test results for the four tested configurations. It can be observed that Sarus with the native MPI hook allows containers to transparently access the accelerated networking hardware on Piz Daint and achieve the same performance as the natively built test.

OSU Latency results

Results of the OSU Latency benchmark for the native MPI and three different containers with ABI-compliant MPI libraries. The MPI in the container is replaced at runtime by the native MPICH MPI hook used by Sarus.

All-to-all

We run the osu_alltoall benchmark only for two applications: native and container with MPICH 3.1.4. We collect latency values for 1kB, 32kB, 65kB and 1MB message sizes, computing averages and standard deviation. The results are displayed in the following Figure:

OSU All-to-all results

Results of the OSU All-to-all benchmark for the native MPI and MPICH 3.1.4 container. The MPI in the container is replaced at runtime by the native MPICH MPI hook used by Sarus.

We observe that the results from the container are very close to the native results, for both average values and variability, across the node counts and message sizes. The average value of the native benchmark for 1kB message size at 16 nodes is slightly higher than the one computed for the container benchmark.

It is worthy to note that the results of this benchmark are heavily influenced by the topology of the tested set of nodes, especially regarding their variabiliy. This means that other tests using the same node counts may achieve significantly different results. It also implies that results at different node counts are only indicative and not directly relatable, since we did not allocate the same set of nodes for all node counts.