Bjorn Andersson: Sun does not have an "official" definition. The traditional definition is that HPC is when you turn a CPU bound problem into an I/O bound problem. You apply so much compute power that instead of the CPU being the bottleneck, it's now instead getting the data in or out of the system that is the bottleneck. Another definition could be that HPC is generally used for problems that can benefit by more compute power right away for creating better quality information or getting to the answer quicker. Where do you see the market for commercial HPC headed?
HPC started in the research space, but one of the key trends is more commercial companies are using it as a competitive weapon. Sun is interested in HPC because it's a big and growing market. From Sun's perspective, it's time to invest in it.
We're seeing a mix. HPC is coming from the world of grow-your-own. We see that still in the research area. But as we see commercial companies coming to HPC, they value the packaging much more. Getting something factory integrated and guaranteed to work is high value. Performance is still the capital P in HPC, but with commercial companies, they're starting to value the softer side of HPC, as well. I haven't heard much about SunGrid lately. Is Sun still behind utility computing?
Yes. You need to be able to offer the whole spectrum. Where is Sun's play in HPC hardware?
From the core technology, from the CPU level, we're selling a lot of Opteron-based systems. It tends to be the best solution for most HPC apps today. And there are other applications that work well with Intel [Inc.]. Going forward, you will see more Sparc-based systems. But it really comes down to what applications you want to run. What is holding HPC back from greater commercial adoption?
Part of it for many commercial companies is that it's a new thing. It's been popping up on the radar, people more or less forced into it by their competitors using it.
Also, HPC is almost a mission critical investment. Companies are for a vendor they can trust. What is happening to make applications more available for HPC? Two things are happening in this area. One is that the applications themselves are being designed for more horizontal scaling. That takes time. For example, I've seen it in oil and gas. A few years ago, all of their applications were built to run on an SMP system, now they are being...
rewritten to take advantage of clusters. I've seen this in other verticals, as well.
On the systems side, the pendulum swung from giant machines to one socket and two socket systems -- now it's coming back a little bit. We're seeing more customers asking for systems with a few more sockets. Not quite large-scale SMP systems. We're seeing the sweet spot being larger three-to-four socket systems with dual-core eight, 16, 32-way machines. It's almost like SMP. We're going to start talking about diagonal scaling rather than horizontal or vertical. Do commercial shops have the expertise to run these systems?
I think it depends on the customer. In some cases they're coming into this fresh. But we're also seeing a lot of customers with technical computing in the form of workstations. We have seen some customers that don't have all of the expertise in-house, and that's why we've picked up on the focus on factory integrated systems to remove that obstacle. What is Solaris' play in HPC?
Coming from the educational research area, choice and open source have been really important. Linux has been a core part of HPC for many years now. Since we've open-sourced Solaris there has been a lot more interest in it as the key OS -- shops that often already have experience with Solaris on the commercial side of the enterprise. Now there is an opportunity to run Solaris on the HPC side. For Sun as a company, it's up to us to provide choice. What are the applications for HPC at this point that haven't been exploited?
Certain verticals stick out. For example, financial services. It's not so much on the transaction side, with the banking. It's more on the back end with risk analysis and investment advice.
Also, data warehousing where you have a lot of sales data to analyze, is another area. Wal-Mart is very much about analysis, making certain decisions on stocking stores. FedEx and UPS could be solving the problems in the routing of packages.
Those are very tricky problems that HPC could help solve.