This post is contributed by Alex Kimber, Global Solutions Architect

No one asks for their High Performance Computing (HPC) jobs to take longer, cost more, or have more variability in the time to get results. Fortunately, you can combine Amazon EC2 and Amazon EC2 Auto Scaling to make the delivery of batch workloads fast, cost-efficient, and reliable. Spot Instances offer spare AWS compute power at a considerable discount. Customers such as Yelp, NASA, and FINRA use them to reduce costs and get results faster.

This post outlines an approach that combines On-Demand Instances and Spot Instances to balance a predictable delivery of HPC results with an opportunistic approach to cost optimization.

 

Prerequisites

This approach will be demonstrated via a simple batch-processing environment with the following components:

  • A producer Python script to generate batches of tasks to process. You can develop this script in the AWS Cloud9 development environment. This solution also uses the environment to run the script and generate tasks.
  • An Amazon SQS queue to manage the tasks.
  • A consumer Python script to take incomplete tasks from the queue, simulate work, and then remove them from the queue after they’re complete.
  • Amazon EC2 Auto Scaling groups to model scenarios.
  • Amazon CloudWatch alarms to trigger the Auto Scaling groups and detect whether the queue is empty. The EC2 instances run the consumer script in a loop on startup.

 

Testing On-Demand Instances

In this scenario, an HPC batch of 6,000 tasks must complete within five hours. Each task takes eight minutes to complete on a single vCPU.

A simple approach to meeting the target is to provision 160 vCPUs using 20 c5.2xlarge On-Demand Instances. Each of the instances should complete 60 tasks per hour, completing the batch in approximately five hours. This approach provides an adequate level of predictability. You can test this approach with a simple Auto Scaling group configuration, set to create 20 c5.2xlarge instances if the queue has any pending visible messages. As expected, the batch takes approximately five hours, as shown in the following screenshot.

In the Ireland Region, using 20 c5.2xlarge instances for five hours results in a cost of $0.384 per hour for each instance.  The batch total is $38.40.

 

Testing On-Demand and Spot Instances

The alternative approach to the scenario also provisions sufficient capacity for On-Demand Instances to meet the target time, in this case 20 instances. This approach gives confidence that you can meet the batch target of five hours regardless of what other capacity you add.

You can then configure the Auto Scaling group to also add a number of Spot Instances. These instances are more numerous, with the aim of delivering the results at a lower cost and also allowing the batch to complete much earlier than it would otherwise. When the queue is empty it automatically terminates all of the instances involved to prevent further charges. This example configures the Auto Scaling group to have 80 instances in total, with 20 On-Demand Instances and 60 Spot Instances. Selecting multiple different instance types is a good strategy to help secure Spot capacity by diversification.

Spot Instances occasionally experience interruptions when AWS must reclaim the capacity with a two-minute warning. You can handle this occurrence gracefully by configuring your batch processor code to react to the interruption, such as checkpointing progress to some data store. This example has the SQS visibility timeout on the SQS queue set to nine minutes, so SQS re-queues any task that doesn’t complete in that time.

To test the impact of the new configuration another 6000 tasks are submitted into the SQS queue. The Auto Scaling group quickly provisions 20 On-Demand and 60 Spot Instances.

The instances then quickly set to work on the queue.

The batch completes in approximately 30 minutes, which is a significant improvement. This result is due to the additional Spot Instance capacity, which gave a total of 2,140 vCPUs.

The batch used the following instances for 30 minutes.

 

Instance Type Provisioning Host Count Hourly Instance Cost Total 30-minute batch cost
c5.18xlarge Spot 15  $     1.2367  $     9.2753
c5.2xlarge Spot 22  $     0.1547  $     1.7017
c5.4xlarge Spot 12  $     0.2772  $     1.6632
c5.9xlarge Spot 11  $     0.6239  $     3.4315
c5.2xlarge On-Demand 13  $     0.3840  $     2.4960
c5.4xlarge On-Demand 3  $     0.7680  $     1.1520
c5.9xlarge On-Demand 4  $     1.7280  $     3.4560

The total cost is $23.18, which is approximately 60 percent of the On-Demand cost and allows you to compute the batch 10 times faster. This example also shows no interruptions to the Spot Instances.

 

Summary

This post demonstrated that by combining On-Demand and Spot Instances you can improve the performance of a loosely coupled HPC batch workload without compromising on the predictability of runtime. This approach balances reliability with improved performance while reducing costs. The use of Auto Scaling groups and CloudWatch alarms makes the solution largely automated, responding to demand and provisioning and removing capacity as required.

from AWS Compute Blog