Advanced Computing in the Age of AI | Tuesday, June 25, 2024

Dell CTO And Fellow Talk Enterprise Tech Convergence 

Dell HQ

Dell is no stranger to the enterprise data centers of the world and it is the dominant supplier of systems to hyperscale data center operators. The company also has built a few very large supercomputer clusters and is the number three supplier of systems in the traditional HPC market, in fact. That gives Dell a view into the convergence of technologies that EnterpriseTech is dedicated to understanding.

To get a sense of what Dell sees going on out there in the upper echelons of enterprise data centers these days, EnterpriseTech sat down with Sam Greenblatt, chief architect and technology officer of the Enterprise Solutions division of Dell, and Jimmy Pike, a technical fellow in that same division, to have a wide-ranging chat.

Internally at Dell, Greenblatt and Pike are known as Click and Clack, in homage to the MIT graduates who hosted Car Talk on National Public Radio for three and a half decades. Just like Tom and Ray Magliozzi, the two Dell techies have fun with what they are doing and are more serious than they might appear.

Timothy Prickett Morgan: We just launched EnterpriseTech to, among other things, focus on the convergence of HPC, hyperscale, and other extreme-scale technologies as they are getting adopted by enterprise customers. This convergence is just getting started, and it is going to radically change the way IT is deployed in enterprise data centers in the years to come.

Sam Greenblatt: I think you are right in your timing, and you are not early. We are at the cusp of this.

Here's one small example. There is not much to say positive about the US government these days, but the President's Council of Advisors on Science and Technology is funding the movement of HPC technology into manufacturing, which is probably the best news for both us and you. We have met with a couple of the major auto companies that are still left in this country, and one of the things that they want to do is use a whole new model of compute.

Jimmy Pike: There are two things at work here, and it goes beyond manufacturers. There is the large data model and the compute model. The one interesting thing that we have seen is, I don't think I would call it a collision, but certainly a paradigm shift going on. If you look at the hyperscale space and the large data models that have to cope with, it is just astounding. When you talk to the HPC guys, they talk about big data. But they hyperscale guys have really big data. We think there will be a merging between their two approaches.

One thing about the enterprise HPC guys, especially when you look at the need to do business intelligence, is that they are all about compute. It is all about compute models and having an enormous amount of computing.

There is going to be a blurring between traditional HPC and what the hyperscale guys are doing, certainly with respect to data. Now, the hyperscale guys, they are not all about floating point operations per second, where HPC applications are. But we still think there will be a lot of borrowing between these two sets of customers.

TPM: So what does that mean? Does it mean that enterprise HPC customers are going to be running simulations with their data sets and at the same time performing analytics against those data sets? And then conversely, does that mean the hyperscale data center operators are going to start writing applications that are very floating point intensive? And therefore they will have to start building things that look more like a modern supercomputer? Let's take natural language processing, which Web application operators like Google and Apple already have and others will no doubt be adding. That is something that can be accelerated with GPUs, and when you build a system to support it, will it look like a hybrid supercomputer?

Sam Greenblatt: That's right. The difference is going to be the workloads they focus on. The underlying systems to do the work are going to start looking remarkably similar.

I will give you a concrete example. I recently met with one of the large auto makers, and the company is putting 34 sensors in the 2014 cars. They expect that to double next year, and they want to transmit the data gathered about the vehicle back to them for analysis in real time. Think about something on the order of 100 million cars on the road from this manufacturer sending back all of that data. This poses all sorts of problems in terms of networking and storage, and this company is not ready to handle the processing of all of this data.

TPM: To bring it all back home for big enterprises, which is what I am all about these days, as they add sensors into their products and want to gather up telemetry from those products, they are going to similarly find they don't have the storage and processing capacity to handle all of that telemetry. Correct?

Sam Greenblatt: The Internet of Things, as people are calling it, is probably the first thing in a long time that I actually believe in. [Laughter] Cloud, I think, has been over-hyped, but this is real.

So here is another example. I spent yesterday with an executive of a big energy company, and this company is really concerned about environmental health and safety. I can't name them, but this company is going to put sensors on everything on their drilling platforms out in the Gulf of Mexico. Since I now live in Texas, I hope no more catastrophes happen. . . .

Anyway, what they are going to do is predictive analysis on the rigs, and that is clearly an enterprise application. Another thing they are going to do tag all of the assets on the platforms, because building these is not cheap and things disappear all the time. They are not only putting RFID tags on equipment, but other kinds of sensors so you can see the movement of these assets.

We think a lot of the compute is moving from the human out into the environment. Not to knock enterprise computing, but everybody can do general ledger, ERP, and all that. Where the competition and competitive advantage is going to come in is how do you use the data in the environment that defines your enterprise.

Jimmy Pike: That is the real value. You have to use your data to figure out how you should do things.

TPM: So this telemetry data will be a large driver of compute, storage, and networking capacity at companies of all sizes. And at the largest companies, presumably with large numbers of customers and a diverse product portfolio, this will push them up into the extreme scale of systems for storing and processing all of this telemetry.

By the way, I happen to think that not all computing is just going to be done on clusters, but that in many cases it will be done in-memory on large NUMA systems, which offer some advantages compared to clusters for certain kinds of workloads.

Sam Greenblatt: We agree with you. Looking ahead, computing is going to take twenty different forms, and only one of them is going to be the classic server. It is going to disaggregate.

Jimmy Pike: I think that because of this, you are going to see what are almost appliance solutions, where there is a specific problem and you have a top-to-bottom optimization around that problem.

Everybody seems to think that, in the end, computing goes back into the cloud, but that is not what we are seeing. We are seeing people using computing in new ways, and they are capitalizing on large data stores. It is almost the next generation of the client/server model, although this time it is client and service and the service is provided by a variety of different kinds of platforms. There are local platforms, some mid-stream platforms, and some data center platforms. It is heterogeneous, and a lot of optimizations on each of those pieces is what we are going to see.

TPM: You have had the PowerEdge-C machines out for a few years now, and these are the commercialized versions of some of the custom machines you have built for hyperscale data centers. Have you seen sales for the PowerEdge-C take off among enterprises that are distinct from HPC centers and hyperscale data center operators?

Jimmy Pike: The answer is yes. We have Web tier version of the PowerEdge-C that is between the hyperscale stuff and the traditional enterprise servers, and we have seen terrific growth there. But the thing we have learned is that one size does not fit all any more. The more agile you are, the more the computing has to fit the environment.

Here's an interesting observation that we have made. The enterprise is not moving to hyperscale platforms and the enterprise is not moving to the cloud. And when I say that, people look at me and say, "What are you, nuts?"

It is not that enterprises are moving to a hyperscale model, but rather that they are finding ways of doing things with hyperscale services that displace what they were doing inside of their own enterprises.

That said, companies are concerned about the security and privacy of their data. And we do see plenty of customers doing private clouds and private hyperscale because they don't want to use public clouds or shared solutions. Like I said, we are moving into a more diverse environment.

TPM: Take me out five years from now, then. What does the enterprise data center look like? It seems to me that there will be a lot of mass customization of platforms.

Jimmy Pike: There is going to be great, great chaos. And to quote the founder of this company, from great chaos comes great opportunity. And this is where Dell has always been able to thrive. Where there have been disruptions and paradigm shifts, we have been able to take advantage of them.

Sam Greenblatt: If you walk into a data center five years from now, you won't be able to recognize it. Intel thinks it is going to be system-on-chips doing specialized processing, and we think that is a little bit of an oversimplification. But what we do believe is going to happen is that there will be specialized processors.

Jimmy Pike: And specialized solutions. And by the way, it is hard to look five years into the future because my crystal ball goes dark out about three years. And if it were not my job, it would be dark in six months. [Laughter.]

One of the things that I see has to do with the thirst and quest for exascale machines. A lot of people are saying that it will happen at the end of the decade, but we think it will happen before then. Necessity is the mother of invention. We are positioning ourselves to be right in the middle of that.

TPM: To circle back to the EnterpriseTech focus, the technology used in those exascale machines will either trickle down to large enterprises or be used concurrently by hyperscale data centers that are facing similar computational needs. I know Dell has built Stampede at the Texas Advanced Computing Center using the hyperscale server designs and a mix of Xeon processors and Xeon Phi coprocessors. But what I really want to know is this: When do I see a quarter of a TACC installed at ExxonMobil, Wal-mart, or JPMorgan Chase?

Sam Greenblatt: You will see that in the next several years.

Let's look at banks, which have to look for unusual activities to root out fraud. They are already looking at supercomputing today, if you will, because if you go back and audit these fraud events, there are always telltale signs.

Jimmy Pike: You can't teach a computer intuition. But what you can do is iterate all possibilities and measure against those possibilities in the right ways and with the right combinations. This will take an enormous amount of compute.

Here's a concrete example. I recently too a trip to Ecuador. And when I went to use the ATM, it wouldn't give me any money because I was not in my home country. I would have much preferred if my bank had realized that I probably was Jimmy, as opposed to someone stealing my card, because the same card was used to buy my airplane tickets, food at the airport, and other items along the way on the trip. This would be a better option than me being in the middle of Ecuador, broke.