Launch Meep Simulations using AWS following this Step-by-Step Guide
1. Obtain an Amazon Web Services (AWS)
account. A credit card is required. Renting multi-core virtual machines is
billed by the second. You pay for only the hardware you use. There are no fees for software. Simpetus provides full technical support and training for $100/month with no contracts and cancel anytime. AWS provides
instructions for launching a virtual machine using Windows, macOS, or Linux.
2. Login to the AWS console. First, choose your
region. Second, create a
key pair for the Elastic Compute Cloud (EC2) and download the private key (.pem file). To use a different region, you must create a different set of keys.
3. Subscribe to the free
Simpetus Electromagnetic Simulation Platform from the AWS Marketplace. This is an Amazon Machine Image (AMI) which includes the latest, pre-installed versions of Meep and MPB on Ubuntu. AWS provides
free use of a t2.micro instance for 1 year as well as
credits for academic research. These video tutorials provide a demonstration of launching a virtual instance using or . Connect to the instance with IP address XX.XXX.XXX.XX via SSH from your local shell terminal. As a requirement for SSH, make sure that the private key has the correct file permissions:
$ chmod 400 private_key.pem
$ ssh -i private_key.pem ubuntu@XX.XXX.XXX.XX
After logging into your virtual instance, you can launch Meep simulations using Python 3 from the default Bash terminal via:
$ mpirun -np 2 python3 myscript.py |tee output.out
Note: the Bash script which is launched automatically at start
/home/ubuntu/.bashrc sets the
PYTHONPATH environment variable. This is necessary to use the Meep module which is installed in the system directory. For simulations involving the Scheme interface, the command is:
$ mpirun -np 2 meep myscript.ctl |tee output.out
Meep's linear-scaling performance on EC2 for a multi-node virtual cluster set up using
AWS ParallelCluster is shown below. The benchmarking test involves time stepping the fields in 3d. The simulation time decreases
linearly with the number of virtual cores. For sufficiently large problems, internode communication is not a bottleneck.
For more details regarding running Meep simulations using cloud-based MPI clusters, see
https://arxiv.org/abs/2003.04287.