In the world of cloud computing every action has a cost. Every HTTP request fires off a chain of actions, each uniquely measured on a variety of billable meters. Gone are the days of idle or unused resources on our local servers. Cloud computing charges by the sip (when sips are available) aligning business goals of resource efficiency and its cost. The cloud computing world shares many similarities with the plug-in and go world of electricity, including the need to run green for the sake of resources and cost savings. What can the world of green energy teach us about the future of cloud computing? How can we measure computing resources in the cloud for efficiency, replacement costs, and cost savings? I shared a few ideas on green clouds at last week’s Ignite at ETech.
Marginal resources at a marginal cost
At the most basic level cloud computing is a marginal measure of resource consumption across processor time, memory use, disk use, and bandwidth consumption. Cloud consumption meters are much more precise, measuring every process and its dependent system APIs. Wasteful programs, processes, or external libraries carry a real and measurable cost. Upgrading to the latest version of an application or library, complete with bug fixes and inefficiencies can save real money each month. We might even choose between similar software packages based on their measured efficiency.
Google App Engine measures each HTTP method, differentiates between database selects and updates, and even measures the system cost of each e-mail message. Below is a highly simplified break-down of App Engine’s consumption meters detailing available resources in a day. App Engine users may purchase additional access to most APIs.
| Meter | Allowed |
|---|---|
| Request Handling | |
| CPU Time | 6:30 |
| HTTP requests | 1,333,328 |
| Outgoing bandwidth | 1 GB |
| Incoming bandwidth | 1 GB |
| HTTPS requests | 1,333,328 |
| Secure outgoing bandwidth | 1 GB |
| Secure incoming bandwidth | 1 GB |
| DB | |
| Database calls | 10,368,000 |
| Database CPU time | 62:07 |
| Database size | 1 GB |
| Inserts or Updates | 12 GB |
| Selects | 116 GB |
| Memcache | |
| Calls | 8,640,000 |
add, set | 10 GB |
fetch | 50 GB |
| Within Google | 5,000 |
| Outside Google | 2,000 |
| Size of message body | 61.4 MB |
| Number of attachments | 2,000 |
| Size of attachments | 102 MB |
| External requests | |
| HTTP requests | 657,084 |
POST, PUT | 4 GB |
GET | 4 GB |
| Dynamic images | |
| Transformations | 2,592,000 |
| Source images | 1 GB |
| Manipulations | 5 GB |
| Deployments | 250 |
App Engine is extremely precise in its application meters and I only included a small sampling! A reduction in resource consumption carries real cost savings and opens up additional headroom for other processes within your application.
Writing efficient code for the cloud
The specialized cloud stack and its meters exposes code inefficiencies that may have gone undetected in a standard hosting environment. Programmers who learn the inner workings of their virtual machine under an environment of constraints will ultimately write better code in any system. Java developers with experience on mobile have operated under device constraints that serves them well in an environment of assumed excess. A cloud programmer taught to tune his code for the interpreter and its inner workings will similarly benefit inside and outside the cloud.
Coding inside a constrained environment such as App Engine has changed how I write web applications. I was able to look through Guido van Rossum‘s code, notice different styles and techniques, and inquire about his coding style. It turns out my code was wasteful in ways I had not considered, yet by observing the virtual machine’s architect and tuner I learned how to provide the right processing hints and optimizations to speed up code and reduce its resource consumption. Training engineers for cloud computing isn’t a specialized investment competing with the cost of machines; it’s a long-term investment in style and proficiency that should pay off in a measurable way under a metered runtime environment.
Measuring efficiency of packaged applications
What if installable web applications carried their own cloud efficiency ratings similar to Energy Star ratings placed on home appliances? We will have Energy Star ratings for web servers and datacenters beginning in May to create a direct comparison of energy usage across server vendors. Similar measurements can be applied to installable software packages or library dependencies in the cloud.
The mockup above shows a possible cloud rating for WordPress inside a PHP cloud instance. A potential WordPress customer could compare efficiency ratings and total cost of operation of WordPress’s code base over the course of a month or a year. He might then compare WordPress against products of similar functionality but varying operational costs such as Drupal or Serendipity. Measuring inefficiencies could motivate software vendors to reduce waste in their products to speed up execution times and save customers real money. Directly measuring resource consumption in this way motivates change up and down the value chain.
We are already somewhat familiar with energy ratings in our daily lives. We evaluate a new refrigerator or washing machine based on its initial price as well as its total cost to operate and repair. Applications bundled into a machine instance for easy deploys could be similarly measured in a direct to cloud and managed cloud provisioning structure.
Rolling back the meter
Google, Microsoft, and Amazon operate cloud computing farms and advertising platforms to pay back the cost of operation. I expect AdSense, AdCenter, and Amazon Associates programs will offer discount premiums for customers generating ad revenue from the same company powering their web presence. Google and Microsoft could also make calls to their own APIs free or cheap to cloud customers since such requests would not need to touch the public Internet.
Consumption dashboards
Cloud dashboards of the future might provide insights into our software in the same ways smart meters and home monitoring solutions hope to measure our electricity use. Electricity monitoring startups are trying to raise awareness around resource consumption, highlight wasteful outliers, and ultimately affect change. Cloud computing companies such as Google can apply lessons learned from funding smarter power meters to cloud computing dashboards of the future.
Summary
In the world of cloud computing every action has a direct and measurable cost. Companies can calculate the savings or a business decision such as increasing cache times from 5 minutes to 10 minutes or new code decisions such as dependent library updates. The meters of the cloud will make us much more aware of our server computing consumption and provide new motivations for change.
Cloud computing changes the infrastructure we use to power our applications. It also changes how we program by introducing constraints, optimization rewards, and systems designed for parallelization and scale. Some of our fear over change and lock-in is a lack of familiarity in operating at scale across distributed nodes. Programming against cloud computing systems retrains software engineers for a world of symmetric multiprocessing and better prepares us for our a future of multiple computing processes in our racks or in the cloud.