Google Compute Engine vs. Amazon EC2: Video Transcoding
blog.zencoder.com
blog.zencoder.com
This isn't the performance we expect here and we will be looking in to see what is going on.
I suspect they will do a decent job of supporting GCE, either through a premium offering themselves, or through third party developers -- same thing AWS has done. I had some meetings with AWS application/security platform consultants recently, and they do a really good job of it, you just have to pay for it.
This isn't a good argument for consumer products (if a random grandmother gets locked out of her gmail, and doesn't know anyone, she may be doomed), but I suspect any developer building products on GCE either knows someone or can pay for support, or can bitch in a high visibility forum and get help from third parties or by Google.
Network obviously isn't; the numbers here include about a dozen test runs. We should make that more clear. Even a dozen isn't enough to be a scientific test, so hopefully we (or someone else) will do more benchmarking in the future.
To be clear, for every machine type, we are offering a HT per virtual CPU. So that means that an n1-standard-8 instance gets 4 physical cores and 8 hyperthreads. (We are also offering 3.75GB of RAM and ~440G of ephemeral disk per vCPU).
So I would say:
n1-standard-1: 1/2 physical core
n1-standard-2: 1 physical core
...
n1-standard-8: 4 physical cores
This stuff is very confusing and I'd love to find the right way to communicate it clearly.
In the ideal world for a set of cores C if I run a piece of independent, memory bound, code on all of them I get O(C) scale. Generally that doesn't hold in shared L2/L3 cores or where HyperThreads are counted as cores.
We can define a scaling factor Cf such that 0 < Cf < 1.0 independent code scales at O(Cf).
In one school you could price out at cores such that 1/Cf physical cores were assigned per Core purchased (rounded up to the nearest unit of computation) but that doesn't begin to get into the question of shared memory vs shared network vs shared disk.
I think the best you can do for now is call it a core if Cf is > .95.
The biggest blocking factor is amount of code that needs to be ported to make it work. High end SFX companies spend virtually all their man hours implementing new stuff. They don't have the time to go back and reimplement everything, and their customers aren't so cost sensitive that they demand it.