Extreme Cloud Computing, CERN Style
Mar 17, 2010 6:00 AM PT
What are the likely directions for cloud computing? Based on the exploration of expected cloud benefits at a cutting-edge global IT organization, the future looks extremely productive.
In this podcast we focus on the thinking on how cloud computing -- both the private and public varieties -- might be used at CERN, the European Organization for Nuclear Research in Geneva.
CERN has long been an influential bellwether on how extreme IT problems can be solved. Indeed, the World Wide Web owes a lot of its usefulness to early work done at CERN. Now the focus is on cloud computing. How real is it, and how might an organization like CERN approach cloud?
In many ways, CERN is quite possibly the New York of cloud computing. If cloud can make it there, it can probably make it anywhere. That's because CERN deals with fantastically large data sets, massive throughput requirements, a global workforce, finite budgets, and an emphasis on standards and openness.
So please join us, as we track the evolution of high-performance computing (HPC) from clusters to grid to cloud models through the eyes of CERN, and with analysis and perspective from IDC, as well as technical thought leadership from Platform Computing.
Join me in welcoming our panel today: Tony Cass, group leader for fabric infrastructure and operations at CERN; Steve Conway, vice president in the high-performance computing group at IDC; and Randy Clark, chief marketing officer at Platform Computing. The discussion is moderated by BriefingsDirect's Dana Gardner, principal analyst at Interarbor Solutions.
Listen to the podcast (34:41 minutes).
Here are some excerpts:
Steve Conway: Private cloud computing is already here, and quite a few companies are exploring it. We already have some early adopters. CERN is one of them. Public clouds are coming. We see a lot of activity there, but it's a little bit further out on the horizon than private or enterprise cloud computing.
Just to give you an example, we at IDC just did a piece of research for one of the major oil and gas companies, and they're actively looking at moving part of their workload out to cloud computing in the next 6-12 months. So this is really coming up quickly.
CERN is clearly serious about it in their environment. As I said, we're also starting to see activity pick up with cloud computing in the private sector with adoption starting somewhere between six months from now and, for some, more like 12-24 months out.
Randy Clark: At Platform Computing we have formally interviewed over 200 customers out of our installed base of 2,000. A significant portion -- I wouldn't put an exact number on that, but it's higher than we initially anticipated -- are looking at private-cloud computing and considering how they can leverage external resources such as Amazon, Rackspace and others. So, it's easily one-third and possibly more [evaluating cloud].
Tony Cass: At CERN we're a laboratory that exists to enable, initially Europe's and now the world's, physicists to study fundamental questions. Where does mass come from? Why don't we see anti-matter in large quantities? What's the missing mass in the universe? They're really fundamental questions about where we are and what the universe is.
We do that by operating an accelerator, the Large Hadron Collider, which collides protons thousands of times a second. These collisions take place in certain areas around the accelerator, where huge detectors analyze the collisions and take something like a digital photograph of the collision to understand what's happening. These detectors generate huge amounts of data, which have to be stored and processed at CERN and the collaborating institutes around the world.
We have something like 100,000 processors around the world, 50 petabytes of disk, and over 60 petabytes of tape. The tape is in just a small number of the centers, not all of the hundred centers that we have. We call it "computing at the terra-scale," that's terra with two Rs. We've developed a worldwide computing grid to coordinate all the resources that we have with the jobs of the many physicists that are working on these detectors.
If you look at the past, in the 1990s, we had people collaborating, but there was no central management. Everybody was based at different institutes and people had to submit the workloads, the analysis, or the Monte Carlo simulations of the experiments they needed.
We realized in 2000-2001 that this wasn't going to work and also that the scale of resources that we needed was so vast that it couldn't all be installed at CERN. It had to be shared between CERN, a small number of very reliable centers we call the Tier One centers and then 100 or so Tier Two centers at the universities. We were developing this thinking around the same time as the grid model was becoming popular. So this is what we've done.
[Our grid] pushes the envelope in terms of the scale to make sure that it works for the users. We connect the sites. We run tens of thousands of jobs a day across this, and gradually we've run through a number of exercises to distribute the data at gigabytes a second and tens of thousands of jobs a day.
We've progressively deployed grid technology, not developed it. We've looked at things that are going on elsewhere and made them work in our environment.
The grid solves the problem in which we have data distributed around the world and it will send jobs to the data. But there are two issues around that. One is that if the grid sends my job to site A, it does so because it thinks that a batch slot will become available at site A first. But maybe a grid slot becomes available at site B and my job is site A. Somebody else who comes along later actually gets to run their job first.
Today, the experiment team submits a skeleton job to all of the sites in order to detect which site becomes available first. Then they pull down my job to this site. You have lots of schedulers involved in this -- in the experiment, the grid, and the site -- and we're looking at simplifying that.
We're now looking at virtualizing the batch workers and dynamically reconfiguring them to meet the changing workload. This is essentially what Amazon does with EC2. When they don't need the resources, they reconfigure them and sell the cycles to other people. This is how we want to work in virtualization and cloud with the grid, which knows where the data is. ...
We're definitely concentrating for the moment on how we exploit effective resources here. The wider benefits we'll have to discuss with our community.
Conway: CERN's scientists have earned multiple Nobel prizes over the years for their work in particle physics. CERN is where Tim Berners-Lee and his colleagues invented the World Wide Web in the 1980s.
More generally, CERN is a recognized world leader in technology innovation. What's been driving this, as Tony said, are the massive volumes of data that CERN generates along with the need to make the data available to scientists, not only across Europe, but across the world.
For example, CERN has two major particle detectors. They're called "CMS" and "ATLAS." ATLAS alone generates a petabyte of data per second, when it's running. Not all that data needs to be distributed, but it gives you an idea of the scale or the challenge that CERN is working with.
In the case of CERN's and Platform's collaboration, the idea is not just to distribute the data but also the applications and the capability to run the scientific problem.
CERN is definitely a leader there, and cloud computing is really confined today to early adopters like CERN. Right now, cloud computing services constitute about $16 billion as a market.
That's just about four percent of mainstream IT spending. By 2012, which is not so far away, we project that spending for cloud computing is going to grow nearly threefold to about $42 billion. That would make it about 9 percent of IT spending. So, we predict it's going to move along pretty quickly. ...
[Being able to manage workloads in a dynamic environment] is the single biggest challenge we see for not only cloud computing, but it has affected the whole idea of managing these increasingly complex environments -- first clusters, then grids, and now clouds. Software has been at the center of that.
That's one of the reasons we're here today with Platform and CERN, because that's been Platform's business from the beginning, creating software to manage clusters, then grids, and now clouds, first for very demanding, HPC sites like CERN and, more recently, also for enterprise clients.
Clark: Historically, clusters and grids have been relatively static, and the workloads have been managed across those. Now, with cloud, we have the ability to have a dynamic set of resources.
The trick is to marry and manage the workloads and the resources in conjunction with each other. Last year, we announced our cloud products -- Platform LSF and Platform ISF Adaptive Cluster -- to address that challenge and to help this evolution.
[Cloud adoption] is being driven by the top of the organization. Tony and Steve laid it out well. They look at the public/private cloud economically, and say, "Architecturally, what does this mean for our business?" Without any particular application in mind they're asking how to evolve to this new model. So, we're seeing it very horizontally in both enterprise and HPC applications.
What Platform sees is the interaction of distributed computing and new technologies like virtualization requiring management. What I mean by that is the ability, in a large farm or shared environment, to share resources and then make those resources dynamic. It's the ability to add virtualization into those on the resource side, and then, on the server side, to make it Internet accessible, have a service catalog, and move from providing IT support to truly IT as a competitive service.
The state of the art is that you can get the best of Amazon, ease of use, cost, accessibility with the enterprise configuration, scale, and dependability of the enterprise grid environment.
There isn't one particular technology or implementation that I would point to, to say "That is state of the art," but if you look across the installations we see in our installed base, you can see best practices in different dimensions with each of those customers.
Conway: People who have already stepped through the earlier stages of this evolution, who have gone from clusters to grid computing, are now for the most part contemplating the next move to cloud computing. It's an evolutionary move. It could have some revolutionary implications, but, from a technological standpoint, sometimes evolutionary is much safer and better than revolutionary.
Dana Gardner is president and principal analyst at Interarbor Solutions, which tracks trends, delivers forecasts and interprets the competitive landscape of enterprise applications and software infrastructure markets for clients. He also produces BriefingsDirect sponsored podcasts. Follow Dana Gardner on Twitter. Disclosure: Platform Computing sponsored this podcast.