Alestic.com - A Personal AWS Blog by Eric Hammond

multiply the speed of compute-intensive Lambda functions without (much) increase in cost

Given:

• AWS Lambda duration charges are proportional to the requested memory.

• The CPU power, network, and disk are proportional to the requested memory.

One could conclude that the charges are proportional to the CPU power available to the Lambda function. If the function completion time is inversely proportional to the CPU power allocated (not entirely true), then the cost remains roughly fixed as you dial up power to make it faster.

If your Lambda function is primarily CPU bound and takes at least several hundred ms to execute, then you may find that you can simply allocate more CPU by allocating more memory, and get the same functionality completed in a shorter time period for about the same cost.

I was chatting with Kevin Boyd (aka Beryllium) on the ##aws Freenode IRC channel about the challenge of invalidating a large number of CloudFront objects (35,000) due to a problem where the cached copies of the objects were out of date and the system had not been designed with versioning in the object path or name.

In addition to the work to perform all of these invalidations (in batches of up to 1,000 in each request with at most 3 request outstanding) there is also the issue of cost. The first thousand CloudFront invalidations are free in a month, but the remainder of the invalidations in this case would cost $170 (at $0.005 for each object).

It occurred to me that one could take advantage of the on-demand nature of AWS by using the following approach:

how I was surprised by a large AWS charge and how to calculate the break-even point

Glacier Archival of S3 Objects

Amazon recently introduced a fantastic new feature where S3 objects can be automatically migrated over to Glacier storage based on the S3 bucket, the key prefix, and the number of days after object creation.

This makes it trivially easy to drop files in S3, have fast access to them for a while, then have them automatically saved to long-term storage where they can’t be accessed as quickly, but where the storage charges are around a tenth of the price.

…or so I thought.

You may be able to save on future EC2 expenses by selling an unused Reserved Instance for less than its true value or even $0.01, provided it is in the “Heavy Utilization” class.

In the description of the Heavy Utilization Reserved Instance, is this statement:

you pay […] a significantly lower hourly usage fee, and you’re charged that lower hourly rate for every hour in the Reserved Instance term you purchase [emphasis added]

What may not be clear to the casual reader is the fact that when you purchase a Heavy Utilization Reserved Instance, you commit not only to paying the one-time up front cost, but you are also committing to paying the hourly charge for every hour of every month, even if you are not running a matching instance!

The Light Utilization and Medium Utilization descriptions state:

Scott Moser and I were wondering if stopping and starting an EBS boot instance on EC2 would begin a new hour’s worth of charges or if AWS would not increase your costs if the stop/start were done a few minutes apart in the same hour.

For some reason, I had assumed that it would start a new hour of fees, possibly because of my experience with the somewhat unrelated terminating old instances and starting new instances. However, we decided it would be easy to test, so here are the results.

I tested with an Ubuntu 10.10 Maverick 32-bit server EBS boot AMI on the m1.small instance type in the ap-southeast-1 (Singapore) region. The AMI should have no effect on charges, so these results should apply to any OS you run on EC2.

I used an AWS account that did not have any EC2 instance fees in the Singapore region that month (Scott’s idea) so that this activity would be easy to see as the only charges on that account.