Hacker News Reader

Comment by trotsky | Hacker News Reader

I'm not super up on os level virtualization providers, but I'm curious why exposing 8 cores is an advantage. Based on the ram sizing it seems like they expect to run 100+ containers per processor - so surely they're restricting cpu time per process (instance) somehow. If you use a bunch of cpu time won't you just get starved out when you hit an invisible cpu wall? If you're only getting 1/10th or less of a cpu core wouldn't it make more sense to pin each instance to just a couple of cores minimizing context switches?

View on HN

You can use as much spare CPU as is available. Upon contention, the CPU is (in theory) dished out fairly. Other schemes are of course possible (like Amazon's EC2 where you get a fixed CPU allotment).

EC2 doesn't actually reserve physical CPU time per vm, do they? I had assumed that the lack of burst just means they are keeping the statistical multiplexing windfall for themselves.

Excluding t1.micro instances you are getting a fixed, dedicated amount of CPU per instance, CPU is not a shared resource on EC2.

Why is it that you get such variable CPU performance on different instances of the same type then?

I was using a whole bunch of c1.medium instances and found the CPU performance varied by a factor of at least 2.

Could that be because of old vs new hardware?

So just to be clear: If I have a reserved instance I'm not using there is a CPU sitting somewhere doing running HLT cycles?

Wow.

If you have a reserved instance that is in the running state but you aren't using it, then yes. If you have a reserved instance that isn't running then that slot is likely available on the spot market.

Vs needing CPU cycles and you having to wait if someone else is using that shared CPU time? There are always tradeoffs.

I didn't mean to be critical, customers certainly get better latancy on a fixed time slot implementation over an interrupt based QoS system. I'm surprised from Amazon's perspective, especially because they are competitive on price.

They're relying on the fact that not everyone maxes out their CPU usage all the time.

Context switching is less of a problem when you consider that the L1 and L2 caches of the CPU(s) are also shared amongst the 100+ VMs running on it.

If you're running something that is CPU intensive, or sensitive to something else taking over the cache, then running in a VM can be painful especially when you're not in control of what else is running on that server.

Reply on news.ycombinator.com