[dijit]

Even minescule differences in the hypervisor makes a big difference, also things like over subscription which (used) to be common in AWS meaning limited or varying memory bandwidth.

Hell, Brendan Gregg, father of observability, noticed a 10x performance difference of machines of the same type inside AWS and that’s before talking about different providers.

GCP is also quite transparent with their CPU capabilities. But I would really hesitate to say 1x2.2GHz vCPU in GCP is equal to 1x2.2GHz vCPU in AWS

[belter]

I am familiar with Brendan Gregg work, but do you have a reference for the point you mentioned of 10x performance difference for machines of the same type?

[dijit]

it was actually 5x, my bad, but here's the references:

Original info being mentioned as a rumour (2016): https://www.youtube.com/watch?v=pYbgcDfM2Ts&t=1631s

Slides for confirmation by Greeg (2017): https://www.brendangregg.com/Slides/AWSreInvent2017_performa...

Talk (2017): https://www.youtube.com/watch?v=89fYOo1V2pA

More discussion: https://www.reddit.com/r/aws/comments/547xbx/netflix_found_5...

Original tweet has been deleted, but I have a screenshot here slide 9: https://sh.drk.sc/~dijit/devfest2019-msv.pdf

[belter]

Thanks! Looking at pag 9 of the presentation it seems the comment would talk about Performance differences, at the same price, not of the same type.

While looking at the StrangeLoop presentation it would be 5x difference for the same type.

I am skeptical. Need to dig deeper here.

Edit: Ok. I was already familiar with the Brendan Gregg presentation you shared. I have just carefully reviewed what you kindly shared. Particularly the Reddit discussion and I think this come up out of a misunderstanding.

Brendan Gregg does not mention it on his presentation. I would be shocked that he would omit mention of such a possibility. Many users in the Reddit discussion tried to investigate such a scenarios and could not observe it, other than the +/- variation expected with instances like T2/T3.

So I think somebody heard variance ( possible with the T2/T3) -> then also variance on performance for same price ( but possibly different types) -> Then variance with workloads and over time of a deployment. And that might explain it.

But none of the resources helps point to actual factual statement. I am not saying it's not possible. Only that personally I have not seen it, and I also have not seen other than tweets and rumors. ( That could be true but not with the evidence seen so far.)