Why is the pricing of t2.large similar to m4.large?

Question

They all have 2vcpus and 8G ram. But with t2.large, you can only use 60% of a single vcpu, average 30% for each of the two vcpus. Even considering the "cpu credits", t2.large seems far weak than m4.large.

Answer 1

The main difference between a t2 instance and an m4 instance is the burstable nature of the t2 instance. (See T2 instance type documentation.)

While the machines are of similar specification, a t2.large instance that is bursting has approximately 60% higher performance than an m4.large instance. However, if the instance runs out of credit then it will be limited in performance.

For an explanation, see the T2: From Startups to Enterprise, Performance for a Low Cost presentation from AWS re:Invent 2016. The difference is explained at the 7m:30s point of the presentation:

Answer 2

I researched this myself and everywhere you look everyone is just talking about bursting.

But it's a fact that the hardware specs for e.g. t2.large and m4.large are either exactly or roughly the same. That means even if you burst on a t2 you will get the same performance as on m4.

But AWS, as in the video John Rotenstein linked, claims you still get 60% more performance on t2. This is due to a little difference in how AWS defines a vCPU.

https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instance-optimize-cpu.html

Each vCPU is a thread of a CPU core, except for T2 instances.

So while on m4 a vCPU is a just a thread, on t2 you get a full core.

2 vCPUs on m4 is just 1 core with 2 threads and on t2 it is 2 full cores but without hyperhtreading.

You can see this when reading /proc/cpuinfo. I used this script to make it easier to read:

t2.large:

CPU model: Intel(R) Xeon(R) CPU E5-2676 v3 @ 2.40GHz 1 CPU,
2 physical cores per CPU, total 2 logical CPU units

m4.large:

CPU model: Intel(R) Xeon(R) CPU E5-2686 v4 @ 2.30GHz 1 CPU,
1 physical cores per CPU, total 2 logical CPU units

So depending on your application 2 hyperthreading threads on 1 core might result in the same performance as 2 cores, in other use cases the 2 cores perform up to 60% better.

So this little note, noone seems to mention or know about, in their documentation makes the difference. Hope this clears it up.