Recommended MPI Settings¶
These are our recommendations to tune your application environment and achieve the best possible performance with Bodo.
These recommendations are only applicable when you are running your workload on a cluster. You do not need to do any of this on your laptop.
Intel-MPI library is the preferred distribution for message passing interface (MPI) specification.
Note that Bodo automatically installs
mpich. Hence, after installing
Intel-MPI, remove [mpich] using this command:
Intel-MPI provides different tuning collective algorithms.
Based on our internal benchmarking, we recommend setting these environment variables as follows:
MPI Process Placement¶
Bodo assigns chunks of data and computation to MPI processes, also called ranks. For example, for a dataframe with a billion rows on a 1000-core cluster, the first one million rows are assigned to rank 0, the second one million rows to rank 1, and so on. MPI placement indicates how these ranks are assigned to physical cores across the cluster, and can significantly impact performance depending on hardware configuration and application behavior. We recommend trying block mapping and round-robin mapping options below for your application to achieve the best performance.
In block mapping, cores of each node in the
hostfile are filled with
ranks before moving on to the next node. For example, for a cluster with
50-core nodes, the first 50 ranks will be on node 0, the second 50 ranks
on node 1 and so on. This mapping has the advantage of fast
communication between neighboring ranks on the same node.
We provide instructions on setting block placement for MPICH and Intel MPI below. The following assumes the hostfile only contains a list of hosts (e.g. it does not specify number of processes per host) and the number of cores on each host is the same.
Block Mapping with MPICH and Intel MPI:
Nis the number of MPI processes,
hostfilecontains the list of hosts, and
Pthe number of processes (cores) per node.
In round-robin mapping, MPI assigns one rank per node in hostfile and starts over when it reaches end of the host list. For example, for a cluster with 50-core nodes, rank 0 is assigned to node 0, rank 1 is assigned to node 1 and so on. Rank 50 is assigned to node 0, 51 to node 1, and so on. This mapping has the advantage of avoiding communication hotspots in the network and tends to make large shuffles faster.
We provide instructions on setting round-robin placement for MPICH and Intel MPI below. The following assumes the hostfile only contains a list of hosts (e.g. it does not specify number of processes per host) and the number of cores on each host is the same.
Round-Robin with MPICH:
More information on controlling process placement with Intel MPI can be found here.
See how to use the Hydra Process Manager for MPICH here.
Recommended AWS Network Interface¶
Elastic Fabric Adapter (EFA) is a network interface for Amazon EC2 instances that has shown better inter-node communications at scale on AWS.
To enable EFA with Intel-MPI on your cluster, follow instructions here.
Some points to note in addition to the referenced instructions:
All instances must be in the same subnet. For more information, see the "EFA Limitations" section here.
All instances must be part of a security group that allows all inbound and outbound traffic to and from the security group itself. Follow these instructions to set up the security group correctly.
For use with Intel-MPI, a minimal installation of the EFA drivers is sufficient and recommended:
Depending on where the drivers were downloaded from, you might need to include a
We recommend the following versions for the EFA installer and Intel-MPI:
Other version combinations are not guaranteed to work as they have not been tested.
For EFA installer versions >= 1.12.0, enabling fork is required by
setting environment variable
To confirm correct settings are enabled, run following
libfabric provider is
environment variables are set as shown below:
 MPI startup(): Intel(R) MPI Library, Version 2021.3.1 Build 20210719 (id: 48425b416)  MPI startup(): Copyright (C) 2003-2021 Intel Corporation. All rights reserved.  MPI startup(): library kind: release  MPI startup(): libfabric version: 1.13.0rc1-impi  MPI startup(): libfabric provider: efa ...  MPI startup(): I_MPI_ADJUST_ALLREDUCE=4  MPI startup(): I_MPI_ADJUST_REDUCE=3  MPI startup(): I_MPI_DEBUG=5