Exploiting Hardware Heterogeneity within the Same Instance Type of Amazon EC2

(tl;dw) This talk discusses how to get the most cost effective servers from EC2 taking into account that EC2 instances are heterogeneous (that is to say they don't run the same hardware so it is almost certain that they will perform differently).

Each instance has three-four subgroups of different hardware types depending on when the configuration was installed. (2006/07/10 etc).

Also, each zone has different configurations, especially as new zones are being installed and newer hardware is being added to start with.

Variations of performance can approach 60% between the same instance types (processor,memory,disk,network).

Cost savings can be achieved by seeking for better performing instances, check it it performs well, if it does, keep it, if it does not, drop the machine and request a new instance.

Instances using E5645 are 1.6 times as effecient as the E5507, however, if you are requesting m1.large servers, you only have a 5% probability of getting a E5645 server if you rented in 2011, whereas now you have a 42% probability.

You can also wiew the slides here: https://www.usenix.org/sites/default/files/conference/protec...

We got bit by this recently when we were running code on some m1.large machines and incorrectly assumed that all m1.larges supported SSE4.2 (which this paper doesn't mention). The only way to guarantee that you get a specific processor type is to use cluster compute instance or play roulette while you are booting instances, neither of which are ideal (which this talk presumably says).

I'm not sure what the ideal way to resolve this outside of further subdividing instance types or rebalancing CPUs so that more advanced CPUs are used at the high end in a consistent manner or if users should be able to request a specific CPU type. (but hopefully jeffbarr will read this and put it on a todolist somewhere)

There's a fairly well hidden PDF link at the bottom which has the details

https://www.usenix.org/system/files/conference/hotcloud12/ho...

I liked William Vambenepe's comment on that -- "Are you the dumb money in the Cloud?" http://stage.vambenepe.com/archives/2061

A similar paper was recently presented at SOCC: http://tinyurl.com/8hgvaq9

This one focuses more on a model for how to take advantage of said heterogeneity, and has a few examples where they actually implement the model on EC2 and see some performance improvements.

How is this usable information, though? Do the authors recommend playing roulette to figure out which of your instances are new? Or even if they're AZ based, everyone has differently coded AZs.