Cloud Computing

LUNKWILL:  Are you not as we designed you to be, the greatest, most powerful computer in all creation?
DEEP THOUGHT:  I described myself as the second greatest …Deep Thought… and such…
LUNKWILL:  Yes yes but…
DEEP THOUGHT:  …I am.
LUNKWILL:  But, but, but - this is preposterous! Are you not a greater computer than The Milliard Gargantu-Brain at Maximegalon, which can count all the atoms in a star in a millisecond?
DEEP THOUGHT:  The Milliard Gargantu-Brain, a mere abacus. Mention it not.
FOOK:  And are you not a more fiendish disputant than The Great Hyperlobic Omnicognate Neutron-Wrangler? Which can destroy -
DEEP THOUGHT:  The Great Hyperlobic Omnicognate Neutron-Wrangler can talk all four legs off an Arcturan Mega-Donkey but only I can persuade it to go for a walk afterwards. Molest me not, with this, pocket calculator stuff!
LUNKWILL:  Then what's the problem?
DEEP THOUGHT:  I speak of none, but the computer that is to come after me.

(The Hitchhiker's Guide to the Galaxy)

Nick Carr recently commented on IBM's new initiative called Project KittyHawk, which sets out to use their Blue Gene technology.  The project aspires to create a “global-scale shared computer capable of hosting the entire Internet as an application”.

There have been a range of online discussions on the back of the article as, once again, Nick Carr manages to hit more than a couple of raw nerves.

The premise of the article is that IBM Blue Gene technology is creating computers of such power that data centres can offer vast amounts of  computational power that businesses can plug into and use according to need at a particular time.

These supercomputers can emulate many individual smaller servers (virtualisation) so businesses can migrate their IT services to this new model.

Rather than data centres just offering a place to put your own servers, they can start to offer virtual servers or services, enabling new business models to be adopted.

The IBM technology is so fast that Project Kittyhawk can emulate the entire internet. 

In the past, there have been two ways of creating a supercomputer.  Firstly, there is the Blue Gene style approach, which creates a massive computer with thousands (or hundreds of thousands) of CPUs.  The other approach, as adopted by Google, is to take hundreds of thousands of small, low cost, computers and hook them together in a “cluster” in such a way that they all work together as one large computer.

Basically, supercomputers have many processors plugged into a single machine, sharing common memory and I/O, while clusters are made up of many smaller machines, each of which contain a fewer number of processors with each machine having it's own local memory and I/O.

There have always been advocates on both sides of the fence, and Nick Carr's article has done a fine job of stirring them into action again - but this time it has become clear that the concept of “The Cloud” is gaining momentum, a concept whose origins lie in clustering and grid computing.

John Willis seeks to 'demystify' clouds and received some interesting comments.  James Urquhart is an advocate of cloud computing and thinks that, as with any disruptive change, some people are in denial about The Cloud.  He has responded to some criticism of his opinions.  Bob Lewis, one of Urquhart's “deniers” has written a few posts on the subject and offers a space for discussion of Nick Carr's arguments.

In order to discuss some of the issues surrounding The Cloud concept, I think it is important to place it in historical context.  Looking at the Cloud's forerunners, and the problems they encountered, gives us the reference points to guide us through the challenges it needs to overcome before it is adopted.

In the past computers were clustered together to form a single larger computer.  This was a technique common to the industry, and used by many IT departments.  The technique allowed you to configure computers to talk with each other using specially designed protocols to balance the computational load across the machines.  As a user, you didn't care about which CPU ran your program, and the cluster management software ensured that the “best” CPU at that time was used to run the code.

In the early 1990s Ian Foster and Carl Kesselman came up with a new concept of “The Grid”.  The analogy used was of the electricity grid where users could plug into the grid and use a metered utility service. If companies don't have their own powers stations, but rather access a third party electricity supply, why can't the same apply to computing resources?  Plug into a grid of computers and pay for what you use.

Grid computing expands the techniques of clustering where multiple independent clusters act like a grid due to their nature of not being located in a single domain. 

A key to efficient cluster management was engineering where the data was held, known as “data residency”.  The computers in the cluster were usually physically connected to the disks holding the data, meaning that the CPUs could quickly perform I/O to fetch, process and output the data.

One of the hurdles that had to be jumped with the move from clustering to grid was data residency.  Because of the distributed nature of the Grid the computational nodes could be situated anywhere in the world.  It was fine having all that CPU power available, but the data on which the CPU performed its operations could be thousands of miles away, causing a delay (latency) between data fetch and execution.  CPUs need to be fed and watered with different volumes of data depending on the tasks they are processing.  Running a data intensive process with disparate data sources can create a bottleneck in the I/O, causing the CPU to run inefficiently, and affecting economic viability.

Storage management, security provisioning and data movement became the nuts to be cracked in order for grid to succeed.  A toolkit, called Globus, was created to solve these issues, but the infrastructure hardware available still has not progressed to a level where true grid computing can be wholly achieved.

But, more important than these technical limitations, was the lack of business buy in.  The nature of Grid/Cloud computing means a business has to migrate its applications and data to a third party solution.  This creates huge barriers to the uptake.

In 2002 I had many long conversations with the European grid specialist for the leading vendor of grid solutions.  He was tasked with gaining traction for the grid concept with the large financial institutions and, although his company had the computational resource needed to process the transactions from many banks, his company could not convince them to make the change.

Each financial institution needed to know that the grid company understood their business, not just the portfolio of applications they ran and the infrastructure they ran upon.  This was critical to them.  They needed to know that whoever supported their systems knew exactly what the effect of any change could potentially make to their shareholders.

The other bridge that had to be crossed was that of data security and confidentiality.  For many businesses their data is the most sensitive, business critical thing they possess. To hand this over to a third party was simply not going to happen.  Banks were happy to outsource part of their services, but wanted to be in control of the hardware and software - basically using the outsourcer as an agency for staff.

Traditionally, banks do not like to take risks.  In recent years, as the market sector has consolidated and they have had to become more competitive, they have experimented outwith their usual lending practice, only to be bitten by sub-prime lending.  Would they really risk moving to a totally outsourced IT solution under today's technological conditions?

Taking grid further into the service offering, is “The Cloud”.  This takes the concepts of grid computing and wraps it up in a service offered by data centres.  The most high profile of the new “cloud” services is Amazons S3 (Simple Storage Service) third party storage solution.  Amazon's solution provides developers with a web service to store data.  Any amount of data can be read, written or deleted on a pay per use basis.

EMC plans to offer a rival data service.  EMCs solution creates a global network of data centres each with massive storage capabilities.  They take the approach that no-one can afford to place all their data in one place, so data is distributed around the globe.  Their cloud will monitor data usage, and it automatically shunts data around to load-balance data requests and internet traffic, being self tuning to automatically react to surges in demand.

However, the recent problems at Amazon S3, which suffered a “massive” outage at the end of last week, has only served to highlight the risks involved with adopting third party solutions.

So is The Cloud a reality?  In my opinion we're not yet there with the technology nor the economics required to make it all hang together. 

In 2003 the late Jim Gray published a paper on Distributed Computing Economics:

Computing economics are changing.  Today there is rough price parity between (1) one database access, (2) ten bytes of network traffic, (3) 100,000 instructions, (4) 10 bytes of disk storage, and (5) a megabyte of disk bandwidth.  This has implications for how one structures Internet-scale distributed computing:  one puts computing as close to the data as possible in order to avoid expensive network traffic.

The recurrent theme of this analysis is that “On Demand” computing is only economical for very cpu-intensive (100,000 instructions per byte or a cpu-day-per gigabyte of network traffic) applications.  Pre-provisioned computing is likely to be more economical for most applications - especially data-intensive ones.

If telecom prices drop faster than Moore's law, the analysis fails.  If telecom prices drop slower than Moore's law, the analysis becomes stronger.

When Jim published this paper the fastest Supercomputers were operating at a speed of 36 TFLOPS.  A new Blue Gene/Q is planned for 2010-2012 which will operate at 10,000 TFLOPS, out stripping Moore's law by a factor of 10.  Telecom prices have fallen and bandwidth has increased, but more slowly than processing power, leaving the economics worse than in 2003.

I'm sure that advances will appear over the coming years to bring us closer, but at the moment there are too many issues and costs with network traffic and data movements to allow it to happen for all but select processor intensive applications, such as image rendering and finite modelling.

There has been talk of a two tier internet where businesses pay for a particular Quality of Service, and this will almost certainly need to happen for The Cloud to become a reality.  Internet infrastructure will need to be upgraded, newer faster technologies will need to be created to ensure data clouds speak to supercomputer clouds with the efficiency to keep the CPUs working. This will push the telecoms costs higher rather than bringing them in line with Moore's Law, making the economics less viable.

Then comes the problem of selling to the business. Many routine tasks which are not processor intensive and time critical are the most likely candidates to be migrated to cloud computing, yet these are the least economical to be transferred to that architecture.  Recently we've seen the London Stock Exchange fail, undersea data cables cut in the Gulf, espionage in Lithuania and the failure of the most modern and well-known data farm at Amazon.

In such a climate it will require asking the business to take a leap of faith to find solid footing in the cloud for mission critical applications.

And that is never a good way to sell to the business.

If you enjoyed this post you might like these similar posts about SOA, Governance, CMDB, EA and ERP. 

© Paul Wallis 2008 All rights reserved

Comments

[James Urquhart]

Paul,

Excellent post. I've been trying to comment for a while, but haven't been able to get comment submission to work until now.

Let me just say that, contrary to how your description of my position may sound to others, I am not blindly "pro-cloud". In fact, I firmly recommend that existing enterprise data centers and applications think hard before going "outside" to a commercial capacity-on-demand provider. In most cases, it would actually be better for such enterprises to convert their own infrastructure to a utility computing model first, while the necessary technologies and businesses mature.

I also define the cloud broadly, to include SaaS, PaaS (e.g.force.com) and HaaS (e.g. Amazon, Mosso, etc.). SaaS is in clearly in play today, HaaS is being experimented with, but PaaS may be the most interesting facet of the cloud in the long term.

That being said, you have provided very valuable information in this post, and I for one very much appreciate the work put onto it. It is very true that bandwidth is something to be nervous about (especially when Amazon charges as much as it does for bandwidth), and I have had some interesting discussions (such as the one you reference) about how data integration will happen over the cloud. Finally, cloud-lockin is indeed something to be concerned about; as in, what happens if my first choice provider sucks? Can I move my applications, data, etc. to someone else cheaply enough that it doesn't put me out of business? Simon Wardley has a good post on that today.