Solving DCOPF Optimization Problem on AWS EC2 Instances

This series of articles explore solutions to the DCOPF (Direct Current Optimal Power Flow) optimization problem using various hardware configurations. We will focus on solving the approximate Linear version of the problem using solvers the open-source GLPK and the commercial FICO.

The first article will focus on CPU configurations. We will utilize AWS EC2 instances with different configurations, primarily varying the number of cores and the amount of RAM allocated to the instances.

Amazon Web Services (AWS) offers a variety of Elastic Compute Cloud (EC2) instances to suit different workload requirements. Some of the common instance types include:

• General Purpose Instances: These instances offer a balance of compute, memory, and networking resources, making them suitable for a wide range of applications, such as web servers and code repositories. Examples include T2 and M5 instances.
• Compute Optimized Instances: These instances are designed for compute-intensive tasks that require high performance processors. They are well-suited for applications such as batch processing and media transcoding. Examples include C4 and C5 instances.
• Memory Optimized Instances: These instances are built for workloads that require large amounts of memory, such as in-memory databases and real-time big data analytics. Examples include R4 and R5 instances.

Experiment Setup and Results

Our research began using T2 instances, and we solved the configurations below for both GLPK and FICO. We observed that while doubling the number of cores or RAM initially decreased processing time, this improvement ceased beyond 4 cores and 16GB RAM. Therefore, we found that 4 cores and 16GB RAM was the optimal configuration for these problems.

To determine if the problem is compute or resource intensive, we will test two different EC2 instances from group C (compute-intensive) and group R (resource-intensive) and note the observations.

The results indicate that C instances solved the problem faster than T and R instances with the same configuration, which suggests that the problem is compute intensive. 

To further verify this, we ran the process on a T2 instance with 4 cores, but limited the RAM to 8GB to confirm that memory was not a factor. The results are below.

The data above indicates that the issue does not get resolved more quickly simply by increasing RAM. This suggests that the problem is not memory intensive.

Conclusion

The problem is computationally intensive, not memory intensive. We found that the problem scales up to a limit with the addition of computing resources.

Note for Second Post

In the second post of this series, we will explore whether solving the DCOPF optimization problem runs faster on GPU-enabled EC2 instances. We will compare the performance of GPU instances with CPU-only instances and discuss the factors that affect the performance gains if any.

‍