Wednesday, March 27, 2013

parlor games: sizing AWS with "terrible math (tm)"

People have been playing this parlor game for years I wanted to follow up on my last post about potentially 3.7 million hosts to play a game to see how many there actually are. I borrowed from the great work over at, many thanks to @powdahound for a quick flash of inspiration.

Caveat: this has absolutely *nothing* to do with my day job, it is all based on purely fictitious (and dubious) math.  The only parts here that are even slightly correct were lifted from random public sources.  Everything below is a pure construct for fun and pleasure only, do not use these numbers to advise your stock broker, set career goals, or make any decision about anything what so ever!

Enjoy :-)

NameDRAM (GB)STORAGE (GB)BITSAPI NAME/hour (linux)/hour (windows)
M1 Small1.716032/64-bitm1.small$0.06$0.12
M1 Medium3.7541032/64-bitm1.medium$0.12$0.23
High-CPU Medium1.735032/64-bitc1.medium$0.14$0.28
M1 Large7.585064-bitm1.large$0.24$0.46
High-Memory Extra Large17.142064-bitm2.xlarge$0.41$0.57
M1 Extra Large15169064-bitm1.xlarge$0.48$0.92
M3 Extra Large15064-bitm3.xlarge$0.50$0.98
High-CPU Extra Large7169064-bitc1.xlarge$0.58$1.14
High-Memory Double Extra Large34.285064-bitm2.2xlarge$0.82$1.14
M3 Double Extra Large30064-bitm3.2xlarge$1.00$1.96
Cluster Compute Quadruple Extra Large23169064-bitcc1.4xlarge$1.30$1.61
High-Memory Quadruple Extra Large68.4169064-bitm2.4xlarge$1.64$2.28
Cluster GPU Quadruple Extra Large22169064-bitcg1.4xlarge$2.10$2.60
Cluster Compute Eight Extra Large60.5337064-bitcc2.8xlarge$2.40$2.97
High I/O Quadruple Extra Large60.5204864-bithi1.4xlarge$3.10$3.58
High Memory Cluster Eight Extra Large24424064-bitcr1.8xlarge$3.50$3.83
High Storage Eight Extra Large117480064-biths1.8xlarge$4.60$4.93

Makes for a nice picture, M3 double extra has a bit of what I've decided to call a "redmond bump", but you can see the server/complexity price performance model pretty well here. Although I do prefer me some straight lines for financial models... heh

So if we take the following averages over all instances, with application of "bad math", but not yet patent pending "terrible math (tm)", we are working up to a new level of "awesome sums" shortly, pay attention that bit comes next and is loaded with "awesome" ;-):

                        Linux   Windows
Ave. inst. cost         $1.28   $1.65
Median inst. cost       $0.70   $1.14

And think daft thoughts, assuming a 1:1 mapping of all public space based on a total number of hosts based of 3,702,664 from my previous epic one liner... also assuming a 1:1 mapping of all private to public hosts, ignore S3, and generally assume even more than we are stating we are assuming :-)

                                     /hour             /year
Average Linux                   $4,733,239   $41,490,672,122
Average Windows                 $6,092,939   $53,409,548,382

Median Linux                    $2,591,865   $22,719,794,382
Median Windows                  $4,221,037   $37,000,807,994
Warning! fake math, average 
of averages, of averages...     $4,409,770   $38,655,205,720

So in theory, EC2 alone could represent between $22B and $53B a year!  Or $4MM/hour!

Random unconfirmed reports from the internet show AWS as $2B/yr, so factor this against a potential $38B... @ 3,700,000 maximum potential servers (from IP space).... gives (drum roll please!), oh and if you are playing along at home, this is the bit when the math goes from bad to "extremely suspect" and onwards to patent pending "terrible math (tm)"!

so, with logical consequence, # servers in AWS  


he he he, ho, ho ho!

I'm guessing that we should all just quit guessing! Although it is super fun!

The real facts are that no matter how you look at the numbers for public cloud (and in particular AWS), they are all enormous, staggering and vast. Even more so when you look at the sheer range of capability on offer, it is just fascinating. The only numbers in this game that continue to be small (and the only ones you really need to care about!) are the $/instance/hour.

And yet, these numbers continue to become ever, ever smaller each and every day. Plus [even with "terrible math(tm)"] there are clearly a lot of servers out there. Basically, this is awesome great news for everyone needing to run science and computation at scale and for short $. That's the really cool part of this parlor game!

And that, oh yes - that is the bit that I do use in the day job :-)

p.s. comments about issues with "terrible math (tm)" will be met with mild amusement

[any opinions here are all mine, and have absolutely nothing to do with my employer]
(c) 2011 James Cuff