Heads up: 3 cloud performance gotchas

The common cloud performance mistakes include:

• Porting code without performing platform localization modifications
• Not considering I/O tuning
• Not considering network latency
  

Porting code without performing platform localization modifications is an error made by those who think they can take C++ code from on-premise platforms and move it to the cloud without a hitch. They can’t.

The fact is you need to localize and optimize any code as you move from platform to platform. Moving to a cloud, either IaaS or PaaS, is no exception. The confusion may stem from cloud computing providers who brag about A-to-A portability, which many do provide. However, that won’t get you A-to-A performance characteristics unless you do additional work.

As for not considering I/O tuning, much to the previous point, you need to optimize the I/O subsystems by tweaking the tunables. Keep in mind this is different than elasticity, where cloud platforms can autoprovision servers as you saturate processors. By contrast, the I/O issue is about access to the native I/O system in the most efficient way. Some cloud providers provide access to tunables; some don’t.

It’s a similar issue with network latency. The Internet doesn’t always provide consistent performance, so you need to consider that latency in the overall performance model for your cloud environment. Do not move to a cloud if it’s going to be a significant issue. 

Also, don’t forget the network latency that may occur as systems communicate within the cloud environment. I find this is often overlooked and actually very difficult to monitor, much less understand, given that you don’t have access to the physical systems. Work with your provider on this one.