4 ms·
As you might imagine, during this exercise, we've had a lot of conversations about the points you raise. We have our own opinions, but we ultimately removed mo
by bhauer 14y ago
As you might imagine, during this exercise, we've had a lot of conversations about the points you raise. We have our own opinions, but we ultimately removed most of that content from the blog post because we didn't want it to be too editorial. We will be posting follow ups with some of our opinions.
Some things are really difficult to answer in a vacuum. If you already have a competent devops staff, hosting your own hardware is probably beneficial. The increased performance per "server" is substantial. But no devops staff? Then it's either very risky or cost-prohibitive to own hardware.
- ochoseis 14y agoI too was shocked at the deltas between Amazon and dedicated hardware. I think AWS runs on Xen, so I wonder if you could strike a more optimal performance/flexibility balance using lighter virtualization (LXC or OpenVZ) with either in-house hardware or another VPS provider.
- spartango 14y agoWhile it is true that EC2 runs on a customized version of Xen, it's very unlikely that latency is being introduced by the hypervisor itself. With paravirtualized kernel and drivers, Xen introduces negligible overhead, and thus LXC would likely be no better. The reason that the virtualized setup performs more poorly than the dedicated setup is that you are fighting for CPU time with other AWS customers, so those other customers are introducing latency into your application. Any shared/virtualized host will have this problem. Interestingly, where Netflix really wants to squeeze CPU performance out of EC2 instances, they allocate the largest instance type so that they know that there's nobody else on the underlying machine.
- e12e 14y agoRegarding performance of Amazon, this post from 2009 is (still) very interesting: http://uggedal.com/journal/vps-performance-comparison/ http://uggedal.com/journal/vps-performance-comparison/ Amazon performance is surprisingly low, but also surprisingly consistent. For some use cases, consistency (knowing what you get you for what you pay) might be worth a considerable hit in performance.