Monday, January 05, 2009

A Technical Look at How Parallel Processing Brings Vast New Capabilities to Large-Scale Data Analysis

Transcript of BriefingsDirect podcast on new technical approaches to managing massive data problems using parallel processing and MapReduce technologies.

Listen to the podcast. Download the podcast. Find it on iTunes/iPod. Learn more. Sponsor: Greenplum.

Dana Gardner: Hi, this is Dana Gardner, principal analyst at Interarbor Solutions, and you're listening to BriefingsDirect. Today we present a sponsored podcast discussion on new data-crunching architectures and approaches, ones designed with petabyte data sizes in their sights.

It's now clear that the Internet-size data gathering, swarms of sensors, and inputs from the mobile device fabric, as well as enterprises piling up ever more kinds of metadata to analyze, have stretched traditional data-management models to the breaking point.

In response, advances in parallel processing, using multi-core chipsets have prompted new software approaches such as MapReduce that can handle these data sets at surprisingly low total cost.

We'll examine the technical underpinnings that support the new demands being placed on, and by, extreme data sets. We'll also uncover the means by which powerful new insights are being derived from massive data compilations in near real time.

Here to provide an in-depth look at parallelism, modern data architectures, MapReduce technologies, and how they are coming together, is Joe Hellerstein, professor of computer science at UC Berkeley. Welcome, Joe.

Joe Hellerstein: Good to be here, Dana.

Gardner: Also Robin Bloor, analyst and partner at Hurwitz & Associates. Thanks for joining, Robin.

Robin Bloor: It's good to be here.

Gardner: We're also joined by Luke Lonergan, CTO and co-founder of Greenplum. Welcome to the show, Luke.

Luke Lonergan: Hi, Dana, glad to be here.

Gardner: The technical response to oceans of data is something that has been building for some time. Multi-core processing has also been something in the works for a number of years. Let's go to Joe Hellerstein first. What's different now? What is in the current confluence of events that is making this a good mixture of parallelism, multi-core, and the need to crunch ever more data?

Hellerstein: It's an interesting question, because it's not necessarily a good thing. It's a thing that's emerged that seems to work. One thing you can look at is data growth. Data growth has been following and exceeding Moore's Law over time. What we've been seeing is that the data sets that people are gathering and storing over time have been doubling at a rate of even faster than every 18 months.

That used to track Moore's Law well enough. Processors would get faster about every 18 months. Disk storage densities would go up about every 18 months. RAM sizes would go up by factor of two about every 18 months.

What's changed in the last few years is that clock speeds on processors have stopped doubling every 18 months. They're growing very slowly, and chip manufacturers like Intel have moved instead to utilizing Moore's Law to put twice as many transistors on a chip every 18 months, but not to make those transistors run your CPU faster.

Instead, what they are doing is putting more processing cores on every chip. You can expect the number of processors on your chip to double every 18 months, but they're not going to get any faster.

So data is growing faster, and we have chips basically standing still, but you're getting more of them. If you want to take advantage of that data, you're going to have to program in parallel to make use of all those processors on the chips. That's the confluence that's happening. It's the slowdown in clock speed growth against the continued growth in data.

Effects on mainstream compute problems

Gardner: Joe, where do you expect that this is going to crop up first? I mentioned a few examples of large data sets from the Internet, such as with Google and what it's doing. We're concerned about the mobile tier and how much data that's providing to the carriers. Is this something that's only going to affect a select few problems in computing, or do you expect this to actually migrate down into what we consider mainstream computing issues?

Hellerstein: We tend to think of Google as having the biggest data sets around. The amazing thing about the Web is the amount of data there that was typed in by people. It's just phenomenal to think about the amount of typing that's gone on to generate that many petabytes of data.

What we're going to see over time is that data production is going to be mechanized and follow Moore's Law as well. We'll have devices generating data. You mentioned sensors. Software logs are big today, and there will be other sources of data ... camera feeds and so on, where automated generation is going to pump out lots of data.

That data doesn't naturally go to Web search, per se. That's data that manufacturers will have, based on their manufacturing processes. There is security data that people who have large physical plants will have coming from video cameras. All the retail data that we are already capturing with things like Universal Product Code (UPC) and radio-frequency identification (RFID) is going to increase as we get finer-grain monitoring of objects, both in the supply chain and in retail.

We're going to see all kinds of large organizations gathering data from all sorts of automated sources. The only reason not to gather that data is when you run out of affordable processing and storage. Anybody with the budget will have as much data as they can budget for and will try to monetize that. It's going to be pervasive.

Gardner: Robin Bloor, you've been writing about these issues for some time. Now, we have had multi-core silicon, and we've had virtualization for some time, but there seems to be a lag between how the software can take advantage of what's going on on the metal. What's behind this discrepancy, and where do you expect that to go?

Bloor: There are different strands to this, because if we talk about parallelization, then with large database products, to a certain extent, we have already moved into the parallelization.

It's an elastic lag that comes from the fact that, when a chip maker does something new on the chip, unless its just a speed -- which was a great thing about clock speed -- you have to change your operating system to some degree to take advantage of what's new on the chip. Then, you may have to change the compilers and the way you write code in order to take advantage of what's on the chip.

It immediately throws a lag into the progress of software, even if the software can take advantage of it. With multi-core, we don't have specific tools to write parallel software, except in one or two circumstances, where people have gone through the trouble to do that. They are not pervasive.

You don't have operating systems naturally built for sharing the workload of multi-core. We have applications like virtualization, for example, that can take advantage of multi-core to some degree, but even those were not specifically written for multi-core. They were written for single-core processes.

So, you have a whole lag in the works here. That, to a certain extent, makes multi-core compelling for where you have parallel software, because it can attack those problems very, very well and can deliver benefit immediately. But you run into a paradox when Intel comes out with a four-way or an eight-way or a 16-way chip set. Then the question is how are you going to use that?

Multi-core becomes the killer app

Gardner: You've written recently Robin that the killer app, so to speak, for multi-core is data query. Why do you feel that's the case?

Bloor: There are a lot of reasons for it. First of all, it parallelizes extremely well. Basically, you have a commanding node that's looking after a data query. You can divide the data and the resources in such a way that you just basically run everything in parallel.

The other thing that's really neat about this application is it's a complete batch application, in the sense that you just keep pushing the data through an engine that keeps doing the queries. So, you're making pretty effective use of all the processes that are available to you. It's very high usage.

If you run an operating system that's based upon intervals, you're waiting for things to happen. At various times, the operating system is idle. It doesn't seem like they're very long times, but mostly on a PC the operating system is never doing anything. Even when you're running applications on a PC, it's rarely doing very much, even in a single-CPU situation. In a multiple-CPU situation, it's very hard to divide the workload.

So that's the situation. You've got this problem that we have with very large heaps of data. They've been growing roughly at a factor of about 1,000 every six years. It's an awesome growth rate. At the same time, we have the technology where we can take a very good dash at this and use the CPU power we've got very effectively.

Gardner: Luke Lonergan, we now have a data problem, and we have some shifts and changes in the hardware and infrastructure. What now needs to be brought into this to create a solution among these disparate variables?

Longergan: Well, it's interesting. As I listened to Joe and Robin talk about the problem, what comes to mind is a transition in computing that happened in the 1970s and 1980s. What we've done at Greenplum is to make a parallel operating system for data analysis.

If you look back on super computing, there were times when people were tackling larger and larger problems of compute. We had to invent different kinds of computers that could tackle that kind of problem with a greater amount of parallelism than people had seen before -- the Connection Machine with 64,000 processors and others.

What we've done with data analysis is to make what Robin brings forward happen -- have all units available within a group of commodity computers, which is the popular computing platform. It's really required for cost-efficient analysis to bring that to bear automatically on structured query language (SQL) queries and a number of different data-intensive computing problems.

The combination of the software-switch interconnect, which Greenplum built into the Greenplum product, and the underlying use of commodity parallel computers, is brought together in this database system that makes it possible to do SQL query and languages like MapReduce with automatic parallelism. We're already handling problems that involve thousands of individual cores on petabytes of data.

The problem is very much real. As Joe indicated, there are very many people storing and analyzing more data. We're very encouraged that most of our customers are finding new uses for data that are earning them more money. Consequently, the driver to analyze more and more data continues to grow. As our customers get more successful, this use of data is becoming really important.

Gardner: Back to Joe. This seem to be a bright spot in computer science, tackling these issues, particularly in regard to massive data sets, not just relational data, of course, but a multitude of different types of content and data. What's being done at the research level that backs up this direction or supports this new solution direction?

Data-centered approach has huge power

Hellerstein: It's an interesting question, because the research goes back a ways. We talked about how database systems and relational query, like SQL, can parallelize neatly. That comes straight out of the research literature projects, like the Gamma Research Project at Wisconsin in 1980s, and the Bubba Project at MCC. What's happened with that work over time is that it has matured in products like Greenplum, but it's been kind of cornered in the SQL world.

Along came Google and borrowed, reused, and reapplied a lot of that technology to text- and Web-processing with their MapReduce framework. The excitement that comes from such a successful company as Google tackling such a present problem as we have today with the Web, has begun to get the rest of computer science to wake up to the notion that a data-centric approach to parallelism has enormous power.

The traditional approach to parallelism and research in the 1980s was to think about taking algorithms -- typically complicated scientific algorithms that physicists might want to use -- and trying to very cleverly figure out how to run them on lots of cores.

Instead, what you're seeing today is people say, "Wow, well, let's get a lot of data. It's easy to parallelize the data. You break it up into little chunks and you throw it out to different machines. What can we do cleverly in computing with that kind of a framework?" There are a lot of ideas for how to move forward in machine learning and computer vision, and a variety of problems, not just databases now, where you are taking this massively parallel data-flow approach.

Gardner: I've heard this term "shared nothing architecture," and I have to admit I don't know anything about what it means. Robin, do you have a sense of what that means, and how that relates to what we are discussing?

Bloor: Yeah, I do. The first time I ran into this was not in respect to this at all. I did some work for the Hong Kong Jockey Club in the 1990s. What they do is take all the gambling on all the horse racing that goes on in Hong Kong. It's a huge operation, much, much bigger than its name sounds.

In those days, they got, I think, the largest transaction rate in the world, or at least it was amongst the top ten. They were getting 3,000 bets in the last second before a race, and they lose the money from the bet if the bet doesn't go on.

The law in Hong Kong was that the bet has to be registered on disk, before it was actually a real bet. So, if in any way, anything fell over or broke during the minute leading up to a race, a lot of money could be lost.

Basically they had an architecture that was a shared nothing architecture. They had a router in front of an awful lot of servers, which were doing nothing but taking bets and writing them to disk. It was server, after server, after server. If at any point, there was any indication that the volume was going up, they would just add servers, and it would divide the workload into smaller and smaller chunks, so it could do it.

You can think of almost being like a supermarket in the sense of lots and lots of different tools and lots of queues for people, but each tool is a resource on its own, and it shares nothing with anything else. Therefore, no bottlenecks can build up around any particular line.

If you have somebody directing the traffic, you can make sure that the flow goes through. So you go from that, straight into a query on a very large heap of data, if you manage to divide the data up in an efficient way.

A lot of these very big databases consist of nothing more than one big fact table -- a little bit more, but not much more than one big fact table. You split that over 100 machines, and you have a query against a whole fact table. Then, you just actually have 100 queries against 100 different data sets, and you bring the answer back together again.

You can even do fault tolerance in terms of the router for all this. So, with that, you can end up with nothing being shared, and you just have the speed. Basically, any device that's out there is doing a bit of query for you. If you've got 1,000 of them, you go 1,000 times faster. This scales extraordinarily well, because nothing is shared.

Gardner: Luke, tell me how these concepts of being able to scale relate to what the developers need to do. It seems to me that we've got some infrastructure benefits, but if we don't have the connection between how these business analysts and others that are seeking the results can use these infrastructure benefits, we're not capitalizing. What needs to happen now in terms of the logic as that relates to the data?

The net effects on users

Longergan: It's a good question, because, in the end, it's about users being able to gain access to all that power. What really turned the corner for general data analysis using SQL is the ability for a user to not to have to worry about what kind of table structure they have. They can have lots of small tables joining to lots of big tables, and big tables joining to each other.

These are things they do to make the business map better to the data analysis they're doing. That throws a monkey wrench in the beautiful picture of just subdividing the data and then running individual queries.

What the developer needs is an engine that doesn't care how the data is distributed, per se, just being able to use all of that parallelism on the problems of interest. The core problem we've solved is the ability for our engine to redistribute the data and the computation on the fly, as these queries and analysis are being performed.

It's the combination, as Robin put it earlier, of a compiler technology, which is our parallelizing optimizer, and a software interconnect, which we call a soft switch technology. The combination of those two things enables a developer of business logic and business analysis to not to have to worry about what is underneath them.

The physical model of how the database is distributed in a shared nothing architecture in a Greenplum system is not visible to the developer. That is where the SQL-focused data analytics realm has gone by necessity. It really has made it possible to continue to grow the amount of data, and continue to be able to run SQL analysis against that data. It's the ability to express arbitrarily constructed business rules against a large-scale data store.

Gardner: We did one of these podcasts not too long ago with Tim O'Reilly. He mentioned that he'd heard from Joe Hellerstein that every freshman now at UC Berkeley studying computer science is being taught Hadoop, which is related on an open-source development level and community to MapReduce. SQL is now an elective for seniors.

It seems that maybe we've crossed a threshold here in terms of how people are preparing themselves for this new era. Joe, how does that relate to how this new logic and ability to derive queries from these larger data sets is unfolding?

Hellerstein: What you're seeing there is three things happening at once. The first is that we have a real desire on the educational side to teach the next generation of programmers something about parallelism. It's really sticking your head in the sand to teach programming the way we have always taught it and not address the fact that every efficient program over the next ... forever is going to have to be a parallel program. That's the first issue.

The second issue is what's the simplest thing you can teach to computer science students to give them a tangible feeling for parallelism, to actually get them running code on lots of machines and get it going? The answer is data parallelism -- not a complicated scientific algorithm that's been carefully untangled, but simple data parallelism in a language that doesn't really require them to learn any new conceptual ideas that they wouldn't have learned in a high school AP course where they learned say Python or Java.

When you look at those requirements, you come up with the Google MapReduce model as instantiated in the open-source code of Hadoop. They can write simple straight-line programs that are procedural. They look just like "For" loops and "If-Then" statements. The students can see how that spreads out over a lot of data on a lot of machines. It's a very approachable way to get students thinking about parallelism.

The third piece of this, which you can't discount, is the fact that Google is very interested in making sure that they have a pipeline of programmers coming in. They very aggressively have been providing useful pedagogical tools, curriculum, and software projects, to universities to ramp this up.

So it's a win-win for the students, for the university, and frankly for Google, Yahoo, and IBM, who have been pushing this stuff. It's an interesting thing, an academic-industrial collaboration for education.

At the business level

Gardner: Let's bring this from a slightly abstract level down to a business level. We seem to be focusing more on purpose-built databases, appliances, packaging these things a little differently than we had in a distributed environment. Luke, what's going on in terms of how we package some of these technologies, so that businesses can start using them, perhaps at a crawl, walk, run type of a ramp up?

Longergan: Businesses have invested a tremendous amount of their time over the last 15 to 25 years in SQL, and some of the more traditional kinds of business analysis that pay off very well are ensconced in that programming model. So, packaging a system that can do transactional, mixed workloads with large amounts of concurrency, with applications that use the SQL paradigm, is very important.

Second, the ability to leverage the trends in microprocessors and inexpensive servers, and combine those with this kind of software model that scales and takes advantage of very high degree of parallelism, requires a certain amount of integration expertise.

Packaging this together as software plus hardware, making that available as a reference architecture for customers, has been very important and has been very successful in our accounts at New York Stock Exchange, Fox, MySpace, and many others.

Finally, as Joe and you were hinting at, there are changes in the programming paradigm. In being able to crawl, walk, and then run, you have to support the legacy, but then give people a way to get to the future. The MapReduce paradigm is very interesting, because it bridges the gap between traditional data-intensive programming with SQL and the procedural world of unstructured text analysis.

This set of technologies, put together into a single operating system-like formulation and package, has been our approach, and it's been very popular.

Gardner: Robin Bloor, this whole notion of legacy integration is pretty important. A lot of enterprises don't have the luxury of starting out "green field," don't have the luxury of hiring the best and brightest new computer scientists, and working on architecture from a pure requirements-based perspective. They have to deal with what they have in place. Increasingly, they want to relate more of what they have in place into an analytic engine of some kind.

What's being done from your perspective vis-à-vis parallelization and things like MapReduce that allow for backward compatibility, as well as setting yourself up to be positioned to expand and to take advantage of some of these advancements?

Bloor: The problem you have with what is fondly called legacy by everybody is that it really is impossible. The kind of things that were done in the past, very strongly bound the software to the data, to the environment it ran in. Therefore, unhooking that, other than starting again from scratch, is a very difficult thing to do.

Certainly, a lot of work is going on in this area. One thing that you can do is to create something -- I don't know if there is an official title to it, but everybody seems to use the word data fabric. The idea being that you actually just siphon data off from all of the data pools that you have throughout an organization, and use the newer technology in one way or another to apply to the whole data resource, as it exists.

This isn't a trivial thing to do, by the way. There are a lot of things involved, but it's certainly a direction in which things are actually going to move. It's possibly not as well acknowledged as it should be, but most of the things that we call data warehouses out there, the implementations have been done in the area of business intelligence (BI), actually don't run very well.

You have situations where people post queries, and it may take hours to answer a query. Because it takes hours to answer a query, and you have a whole scheme, a reason why you are actually mining the data for something, if every step takes a couple of hours, it's very difficult to carry out an analysis like that in a particularly effective way.

A 100-to-1 value improvement

If you take something like the Greenplum technology, and you point to the same problem, even though you are not dealing with petabytes of data, you can still have this parallel effect. You can get answers back that used to take 100 minutes, and you will get 100 to 1 out of this. You may get more, but you will certainly get 100 to 1 out of this, and it changes the way that you do the job that you have.

One thing that's kind of invisible is that there is a lot of data out there that's not being analyzed fast enough to be analyzed effectively. That's something that I think parallelism is going to address.

The other thing where it is going to play a part is that organizations are going to build data fabrics. In one way or another, they will siphon the data off and just handle it in a parallel manner. There is a lot you can do with that, basically.

Gardner: Joe Hellerstein, is there more being brought to this from the data architecture perspective, jibing the old with the new, and then providing yet better performance when it comes to these massive analytic chores?

Hellerstein: What I'm excited about, and I see this at Greenplum -- there's another company called Aster Data that's doing this, and I wouldn't be surprised if we see more of this in the market over time -- is the combination of SQL and MapReduce in a unified way in programming environments. This is short-term step, but it's a very pragmatic one that can help with people's ability to get their hands on data in an organization.

The idea is that, first of all, you want to have the same access to all your data via either an SQL interface or a MapReduce programming interface. When I say all the data, I mean the stuff you used to get with SQL, the database data, and the stuff you might currently be getting with MapReduce, which might be text files or log files in a distributed-file system. You ought to be able to access those with whatever language suits you, mix and match.

So, you can take your raw log files, which are raw text, and use SQL to join those against a customer table. Or, if you're a MapReduce programmer who does analytics and doesn't know SQL, say you're a statistician, you can write a MapReduce program that does some fancy statistical analysis. You can point it at text fields in a database full of user comments, or at purchase records that you used to have to dump out of the database into text formats to get your hands on. So, part of this is getting more access to more people who have programing paradigms at their fingertips.

Another piece of this is that some things are easier to do in MapReduce, and some things are easier to do in SQL, even when you know both. Good programmers have a lot of tools in their tool belt. They like to be able to use whatever tool is appropriate for the task. Having both of these things interleaved is really quite helpful.

Gardner: Luke, to what degree are they interleaved now, and to what degree can we expect to see more?

Longergan: It's been very gratifying that just making some of those pragmatic capabilities available and helping customers to use them has so far yielded some pretty impressive results. We have customers who have solved core business problems, in ways they couldn't have before, by unifying the unstructured text-file data sources with the data that was previously locked up inside the database.

As Joe points out, it's a good programmer who knows how to use all of the various tools that they have at their disposal. Being able to pull one that's just right for the task off the shelf is a great thing to do. With the Greenplum system we've made this available as a simple extension and just another language that one can use with the same parallel data engine, and that's been very successful so far.

Impact on cloud computing

Gardner: Let's look at how this impacts one of the hot topics of the day, and that's cloud computing, the idea that sourcing of resources can come from a variety of organizations. You're not just going to get applications as a service or even Web services, but increasingly infrastructure functionality as a service.

Does this parallelization, some of these new approaches to programming, and the ability to scale have an impact on how well organizations can start taking advantage of what's loosely defined as cloud computing? Let's start with you Joe.

Hellerstein: I'm not quite sure how this is going to play out. There are a couple of questions about how an individual organization's data will end up in the cloud. Inevitably it will, but in the short-term, people like to keep their data close, particularly database data that's traditionally been in the warehouses, very carefully managed. Those resources are very carefully protected by people in the organization.

It's going to be some time until we really see everybody's data warehouses up in the cloud. That said, as services move into the cloud, the data that those services spit out and generate, their log files, as well as the data that they're actually managing, are going to be up in the cloud anyway.

So, there is this question of, how long will it be until you really get big volumes of data in the cloud. The answer is that certainly new applications will be up there. We may start to see old data getting uploaded in the cloud as well.

There's another class of data that's already becoming available in the cloud. There is this recent announcement from Amazon that they are going to make some large data sets available on their platform for public access. I think we'll see more of this, of data as a utility that's provided by third parties, by governments, by corporations, by whomever has data that they want to share.

We'll start to see big data sets up there that don't necessarily belong to anyone, and they are going to be big. In that environment, you can imagine big data analytics will have to run in the cloud, because that's where the data will be.

One of the fun things about the cloud that's really exciting is the elasticity of the resources. You don't buy yourself a data center full of machines, but you rent as many machines as you need for a task.

If you have a task that's going to look at a lot of data, you would rent a lot of machines for a few hours, and then you would shrink your pool. What this is going to allow people to do is that even small organizations may, for a short period of time, look at an enormous amount of data, which perhaps doesn't originate in their own data production environment, but is something that they want to utilize for their purposes.

There is going to be a democratization of the ability to take advantage of information, and it comes from this ability to share these resources that compute, as well as the actual content to share them in a temporary way.

Gardner: Let's go to Robin on that. It seems that there is a huge potential payoff if, as Joe mentioned, you can gather data from a variety of sources, perhaps not in your own applications, not your own infrastructure and/or legacy, but go out and rent or borrow some data, but then do some very interesting things with it. That requires joins, that requires us to relate data from one cloud to another or to suck it into one cloud, do some wonderful magic-dust pixie sprinkling on it, and then move along.

How do you view this problem of managing boundaries of clouds, given that there is such a potential, if we could do it well, with data?

Looking at networks

Bloor: There would have to be, because you are looking at a technical problem, and you really are going to have to have specific interfaces for doing that, especially if you are joining data across clouds. Let's drop the word "cloud" and just think large network, because everything that is representative of the cloud ultimately comes down to being somewhat of a larger network.

When you've got something very large, like what Google and Amazon have, then you have this incredible flexibility of resources. You can push resources in or redeploy these resources very, very effectively. But you're not going to be able to do joins across data heaps in one cloud and another cloud, and in perhaps a particular network without there being interfaces that allow you to do that, and without query agents sitting in those particular clouds that are going to go off and do the work. You're going to care very much as to how fast they do that work as well.

This is going to be a job for big engines like Greenplum, rather than your average relational database, because your average relational database is going to be very slow.

Also, you have to master the join. In other words, the result has to arrive somewhere, and be brought together. There are a number of technical issues that are going to have to be addressed, if we're going to do this effectively, but I don't see anything that stops it being done. We have the fast networks to enable this. So, I think it can be done.

Gardner: Luke, last word goes to you. I don't expect you to pre-announce necessarily, but how do you, from Greenplum's perspective, address this need for joining, but recognizing it's a difficult technical problem?

Longergan: Well, the cloud really manifests itself as a few different things to us. When Joe was talking about how people are going to be putting, and are already putting, a lot of services up in the cloud that are generating a lot of new data, then it requires that the kinds of data analysis, as Robin was hinting at, scale to meet that demand.

We already have the engine that implements those kinds of join in between networks abilities. So we are cloud capable. The real action is going to be when people start to do business that counts on public clouds to function properly, and are generating enormous amounts of very valuable data that requires the kind of parallel compute that we provide.

Joining inside clouds, using cloud resources to do the kind of data analysis work, this is all happening as we speak, and this is another aspect of what's forcing the change from an earlier paradigm of database to the modern massively parallel one.

Gardner: I just want to wrap up quickly now. Thank you. Joe Hellerstein, you mentioned earlier on Moore's Law and how it stalled a bit on the silicon. Are we going to see a similar track, however, to what we did with processing over the last 15 years -- a rapid decrease in the total cost associated with these tasks? Even if we don't necessarily see the same effect in terms of the computing, are we going to be able to do what we've been describing here today at an accelerating decreased total cost?

Hellerstein: Absolutely. The only barrier to this is the productivity of programmers.

Just think about storage. I have a terabyte disk in my basement that holds videos, and it costs $100 or so dollars at Amazon.com. Ten years ago a terabyte was referred to by the experts in the field as a "terror byte." That's how worried people were about data volumes like that.

We'll see that again. Disk densities show no signs of slowing down. So, data is going to be essentially no cost. The data-gathering infrastructure is also going to be mechanized. We're going through what I call the industrial revolution of data production. We're just going to build machines to generate data, because we think we can get value out of that data, and we can store it essentially for free.

The compute cost of multi-core with parallelism is going to continue Moore's Law. It's just going to continue it in a parallel programming environment. If we can get all those cores looking at all that data, it won't cost much to do that, and the cost of that will continue to shrink by half.

The only real barrier to the process is to make those systems easy to program and manageable. Cloud helps somewhat with manageability, and programming environments like SQL and MapReduce are well-suited to parallelism. We're going to just see an enormous use of data analysis over time. It's just going to grow, because it gets cheaper and cheaper and bigger and bigger.

Gardner: Well, great, that's very exciting. We've been discussing advances in parallel processing using multi-core chipsets and how that's prompted new software approaches such as MapReduce that can handle these large data sets, as we have just pointed out, at surprisingly low total cost.

I want to thank our panel for today. We have been joined by Joe Hellerstein, professor of computer science at UC Berkeley, and I should point out also an adviser at Greenplum. Thank you for joining, Joe.

Hellerstein: It was a pleasure.

Gardner: Robin Bloor, analyst and partner at Hurwitz & Associates. I appreciate your input, Robin.

Bloor: Yeah, it was fun.

Gardner: Luke Lonergan, CTO and co-founder at Greenplum. Thank you, sir.

Longergan: Thanks, Dana.

Gardner: This is Dana Gardner, principal analyst at Interarbor Solutions, you have been listening to a sponsored podcast from BriefingsDirect. Thanks and come back next time.

Listen to the podcast. Download the podcast. Find it on iTunes/iPod. Learn more. Sponsor: Greenplum.

Transcript of BriefingsDirect podcast on new technical approaches to managing massive data problems using parallel processing and MapReduce technologies. Copyright Interarbor Solutions, LLC, 2005-2009. All rights reserved.