Amazon Makes It Almost Impossible To Calculate Their "Virtual CPU" Equivalent (informationweek.com)

dkatana writes: AWS started out defining its virtual CPUs as being composed of EC2 compute units, or ECUs, which it defined as an equivalent to a physical Xeon processor. However, a virtual CPU now looks suspiciously variable... A virtual CPU is whatever Amazon wants to offer in an instance series. The user has no firm measure to go by. From the article: [B]y doing a little math, you could actually compare what you were getting in virtual CPUs in EC2 versus Azure. Also by doing a little math, you knew how to compare one Amazon instance to another based on the ECU count in each virtual CPU. Microsoft didn't look too bad in the comparison. That is one of the casualties of the nomenclature change. I have searched for updated information on how a virtual CPU is measured and found nothing comparable to the definition of the 2012 ECU measure. I have questioned Amazon representatives three times between Oct. 27 and Dec. 21, and don't have much of an answer."

Comments

Re:Quit whining

[pijokela]

That is not really possible. Amazon is getting to a monopoly position if you want to use all the modern Cloud stuff. Sure, you can get boxes from many different providers, but AWS has a ton of other services that you cannot buy from others and even more importantly, all the 3rd party Cloud services are running on AWS so they are faster if you are on AWS too.

I would not be surprised if in 10 years Amazon would be a verb for running server software like google is now for search.

Mostly irrelevant to most people

[AmiMoJo]

Most people don't care about the exact performance, so it's not with spending the money and effort to precisely define or guarantee it.

Amazon is generic and cheap. Microsoft has really good integration with visual studio and .NET. Those are the factors people choose by.

Re: Mostly irrelevant to most people

[TheReaperD]

After rereading my post, I realized I needed to post a clarification for the easily offended: "your alcoholic uncle" is used as a generic example and not referring to the AC's uncle.

Re: Mostly irrelevant to most people

[TheReaperD]

The uncle clarification was to make sure that you and others knew I wasn't just trying to be a troll. When I get marked as troll, I want to earn it. As far as the counters, it's not that their random, it's that they are optimized to give inflated tallies for benchmarks. So, in some areas, they're going to vastly overstate their performance (and thus, the "fish story" analogy) but, when you actually call on that level of performance, it isn't there. Which areas that do this will vary by manufacturer and model. The companies that try to offer serious and fair benchmarks have had a nightmare time trying to resolve this. They often have to reverse-engineer where each processor is lying, figure out an estimate of percentage it is lying on that particular test and modify the final results by that percentage. Of course, the company that made the chip will say the adjustment is unfair, then they rebuttal, etc. If that alone wasn't bad enough, chips (though this is more common on graphics chips) that are labeled as mid-range are, in fact, higher-end chips that have been stepped down by firmware, that remove the firmware imposed speed limiter for benchmarks only. These are just the examples I know of; I have no doubt there are more.

Re: Mostly irrelevant to most people

[TheReaperD]

Unrelated side comment: Where the hell are all my line breaks going?

Re:Quit whining

[vikingpower]

Alternative: build your own private cloud out of the smallest servers you can find that still suit your need. I did so: a private cloud on HP Microserver (gen 8). The things consume almost no power when idle. Taken together, they provide quite the computing power ( 64 cores, Xeon E3 ) and quite the storage (32 TB). Cost me around € 450 in electric power per year. Am not dependent on Azure or Amazon. Use no bandwidth when doing cloudy things.

what a surprise

[climb_no_fear]

that the "Cloud" is nebulous....

Re: Quit whining

[DThorne]

I'm not sure I'd call it whining - it's a significant investment for many to set up cloud computing the way they need it so I think it's a fair demand that if your gas station insists on measuring it's product in frackles that it give you a fair conversion rate for litres to frackles. My needs would be something predictable like rendering 3d images, so the first thing I'd be doing is measuring render times for the same image there and here. That's easy and relatively cheap to do, and the numbers scale directly, but I can see this being a serious issue for other uses.

If you need well defined performance...

[drolli]

then buy dedicated instances.

I like micro instances/instances which do not occupy full physical processors at Amazon because of availability and price for low-impact/bandwidth applications. For all other use lambda or dedicated instances.

Virtual CPUs are anyway difficult to asses - to me it may be very relevant to have the 1st level cache of the core which i run on undisturbed by other applications (since changing the cache hits is a big deal for specific numerical problems), and for you 20% more share of the CPU may be important.

1% of computation time not spend in my task on a physical processor can do as much damage as 50% change in speed.

A small side remark: the price for the different VCPUs also varies.

Re: Quit whining

[Anonymous Coward]

Microsoft is a joke and since Satya insist on moving everything to India, treating people like shit, and firing everyone after a few years for "IP" reasons, it's a trainwreck of a company that only still survives because of its initial brand-name and market penetration.

Re:Quit whining

[ShanghaiBill]

Alternative: build your own private cloud

The whole point of doing computing in the cloud is to handle variable demand. I spend 98% of my time writing and debugging code, and only 2% running the final model. But when I run it, I want to do it at scale. On AWS I can rent a dozen K80 GPUs for a few hours a week. There is no way it would make sense to own them, even if I could afford to do that.

Re: Quit whining

[NormalVisual]

And with all the crap they're pulling with the telemetry and whatnot in Windows 10, they seem hell-bent on losing that as well.

Microsoft makes it almost impossible to be license

[mrmaster]

Big deal. Try asking azure support about licensing and they will flat out say I don't know. Assuming you bugged them at least a half a dozen times about the same questions beforehand.

Re: Totally true

[mSparks43]

when I did some testing a while back. $20 a month would get you processing of a crappie mobile you could buy for $50. Amazon has never had good cpu.

so it's probably more that they don't want to say.

anyone who's benchmarked them against alternatives has nothing good to say.

http://openmymind.net/Why-I-Di...

Re: Totally true

[lgw]

That's a good link. My real complaint about AWS is what's mentioned at the end:

It also feels like a lot of services are stuck at version 1.0, lacking that polish and continual improvement

This is what annoys me. SQS is a good 1.0 version of message queues, but the features are the just above the minimum you could possibly call a message queue. DynamoDB is a good 1.0 version of a NoSQL DB, just above the minimum you could possibly call a NoSQL DB.

These services are years old, but look like what most software does at version 1.1 or so: minimal features, no glaring bugs, but nothing great either.

I can't say anything about EC2 CPU performance, as I've never benchmarked that myself, but aren't they just (mostly older) Xeons? The way the T2 instance is described makes it sound like the cores are oversubscribed, but I haven't heard that about the other instances.

In any case, unless you're CPU-bound, it doesn't matter. I'm more concerned about price for the memory I need. For compute-intensive jobs, EC2 Spot is cheap if you're really fault tolerant.