blog.dt.org

“Is colocation cheaper than using a cloud computing service to run the same workload?”

This quote comes from an analysis of the costs of cloud-based computing vs. traditional colocation as a function of the work load and duty cycle. This type of analysis is increasingly germane for companies that are looking to make a transition to cloud-based service providers in the hope that it will allow them to lower their overall IT costs. The results, while not surprising, do raise some interesting points. First, here’s the link:

https://vijaygill.wordpress.com/2010/08/09/cloud-economics/

The crux of the argument here is the concept of the duty cycle. The duty cycle for a deployed application is the percentage of available hardware resources that are being consumed at any given moment in time. Intuitively, a higher duty cycle corresponds to more efficient and cost-effective use of the underlying hardware. One of the fundamental promises of cloud computing is that it will allow you to run applications in a way that will produce a much higher duty cycle through elasticity, i.e. idle resources can be released without impacting the ability to scale back up in the future.

The analysis includes a spreadsheet with some hard data for one specific style of application, and the result is that an equivalent workload for this style of application would cost $118,248 at Amazon and $70,079 in a colocation facility, with the implication that a higher duty cycle can be achieved via colocation. However, this result is not as clear cut as it might seem, owing to the fact that the “application style” is an extremely subjective and important attribute for any given application. In my experience, it is rare to find an application that can be characterized in this way; instead, most applications that I have run across are inherently custom in some important way. I would think that this analysis would need to be performed on a case-by-case basis for any application that is considering a move to the cloud, and the specific result of that analysis would apply only to that application.

A subtle conclusion of this type of analysis is that duty cycle optimization for a given application should be a key criterion for cost reduction. And, somewhat conversely, if an application already has a high duty cycle then the opportunities for cost reduction through cloud-based resources will be limited at best. Or, more simply: if you already run highly-utilized servers then you might do better with collocation.

Hope this helps.

Amazon EC2 I/O Performance: Local Ephemeral Disks vs. RAID 0 Striped EBS Volumes

I recently ran into an issue with I/O bandwidth in EC2 that produced some unexpected results. As part of my day job, I built a system to run clustered deployments of enterprise software within AWS. The enterprise software I chose for the prototype is, as it turns out, very sensitive to I/O bandwidth, and my first attempt at a clustered deployment using m1.large instances with only ephemeral disks did not provide enough I/O performance for the application to operate. After a few internet searches, I was convinced that I needed to use EBS volumes, probably in a RAID 0 striped volume configuration, to effectively increase the I/O performance through parallelism of the network volumes. We also got some guidance from Amazon to this effect, so it appeared to be, by far, our best bet to solve the problem.

So I converted my prototype to use RAID 0 4X-striped EBS volumes. And the I/O performance was different, but still not enough for the application. Uh-oh.

To better understand the problem, we decided that the next step would be to quantify the I/O performance of the different volume options, in the hope that we could ascertain how to tweak the configuration to get enough performance. A colleague of mine ran a variety of tests with iometer and produced the following data that captures read and write performance for six different volume configurations: a single EBS volume, a 4X EBS striped volume, a 6X EBS striped volume, an m1.large local ephemeral drive, an m1.xlarge local ephemeral drive, and “equivalent hardware” which, roughly speaking, corresponds to a server box that should be similar to a large EC2 instance in terms of performance and which is not heavily optimized for I/O. Some of these choices deserve some explanation. We looked at both m1.large and m1.xlarge instances because we were told that the m1.xlarge instances had a larger slice of I/O bandwidth from the underlying hardware, and we wanted to see how much larger that slice actually is, given the cost difference. The EBS-based volumes ran on m2.large instances, as we were told that these instances had a larger slice of real I/O bandwidth as well as a larger network bandwidth slice that would better support the EBS volumes.

So here is the hard data:

A few things jumped out at us when we saw these numbers:

• The performance of the 6X EBS striped volume was significantly worse than the 4X EBS striped volume. Going from 4X to 6X must saturate something that effectively causes contention and reduced performance. Some information online suggests that EBS network performance can be quite variable, but we got consistent results over several runs.
   
• The EBS-based volumes all had better read performance than the local ephemeral disk-based volumes in both IOPS and Mbps. The local ephemeral disk-based volumes had much better write performance in IOPS, and similar write performance in Mbps.
   
• Neither the EBS-based volumes nor the local ephemeral disk-based volumes could stand up to a real piece of dedicated hardware. Which is reasonable, given what EC2 is. But which also implies that it is not a good idea to run I/O intensive operations in EC2. And hardware optimized for I/O would probably blow away the performance of our unoptimized “equivalent hardware” box.

Hopefully this data can point you in the right direction if you are dealing with I/O issues in EC2. We’re hoping for a little magic from Amazon on this issue later this year (and I can’t say anything more about that). In the interim, I/O issues will continue to require some trial and error to find a configuration that best matches your particular I/O profile.

Hope this helps.

My experimental search engine is now available in 10 languages…

About 18 months ago, I created an experimental search engine to play with some new ways to extract and distill information from the web and present it in a more topic-focused way. The site is called doryoku (which is Japanese — 努力 — meaning supreme effort). It has been doing well, and my experiments and enhancements continue. My latest addition involves automatic translation: the site is now available in 10 languages: English, Japanese, Chinese, Korean, French, German, Italian, Spanish, Russian, and Greek. The translation capabilities are courtesy of the Google Language API.

More new features to come. Stay tuned!

Tags: machine learning, search, translation