By Steve Fisher, Editor In Chief -- SGI announced early this morning that it has augmented its already substantial product portfolio. These augmentations came in the form of two products: the Total Performance 900 (TP900) storage system, and the SGI Origin 300 server. At a press briefing, one could easily sense that there was a noticeable level of excitement in the air and that the SGI people’s spirits were high. The TP900 is the newest member of SGI’s Total Performance family of storage arrays and is targeted at the media, manufacturing, sciences and defense/intelligence markets for file serving, digital content management and media streaming solutions. It is being touted as an “entry-level disk array providing maximum storage density with the versatility of SCSI at an attractive price.” Features of the TP900 include: -- Support for eight high-capacity, low-profile 73GB SCSI drives for up to 584GB of storage in a 2U-high (3.5 inch) configuration -- powerful capacity in a compact form factor; -- Latest Ultra160 SCSI technology with an internal peak throughput capability of 320MB per second -- robust performance for bandwidth-centric applications -- A disk array design that provides SCSI versatility and the ability to add dense storage capacity by combining two TP900 units for up to 1.1TB of data in a single industry-standard rack; -- Ultra160 SCSI 10,000 RPM for deep-archive and 15,000 RPM disk drives for fast data retrieval configurations; and -- A flexible storage solution as either a standalone device or as a disk expansion module for the new SGI™ Origin™ 300 server. The TP900 provides a very high-density device in a very small package. It is based on the Ultra160 SCSI protocol, so obviously one has 160MB per second on a single channel. When comparing that to the 100MB/s you get out of current fibre channel offerings, the TP900 is about 60% faster. SGI also offers an enhanced bandwidth feature, which allows the user to have two channels on a single enclosure, thus doubling your rate to 320 MB/s. “We have a lot of flexibility in configurations. We are targeting the SGI platform market as opposed to trying to sell it out into the general public to anybody who wants to attach a SCSI device to any sort of a server, so that’s our focus area and we do offer it as a standalone storage product that you can attach to an ONYX 3000 Series server, Octane, Origin 3000 or for O2’s as well. And it also comes as a module in the Origin 300 server,” said Sylvia Crain, SGI Marketing Manager for Complex Data Management. One of the areas SGI is targeting is the media streaming area. The TP900 is particularly well suited to things like Webcasting, learning on demand, corporate, educational, or entertainment uses. “So, let’s imagine a sample configuration,” Crain said. “Say we’ve got 6U of space, so we take the Origin 300, which is 2U, and then we’ve got two TP900 enclosures that are 2U apiece. We’ll say the Origin 300 has two CPU’s and 512MB of memory, a fairly small server. And then we’ve got one TP900 that’s fully configured with the 15K RPM drive and the bandwidth enhancement, and then the other enclosure is half populated with the bandwidth enhancement, so obviously we’re trying to focus on performance here and bandwidth as opposed to maximizing the storage capability. “So in the example here we can manage with this configuration up to 500 simultaneous streams of video which gives you about 280 hours of video in the 6U package. And if you do the calculations, say in the case of a movie is about two hours in length, that provides about 640 two-hour movies that you can store on a system of this size.” Crain stated. “We’re still working on the final benchmark numbers for this but we do believe that the TP900 in this configuration with the Origin 300 will provide a significant improvement in the price for stream performance in this market space.” Data and information management are of course incredibly important to enterprises everywhere. In these troubled economic times a small, powerful, relatively inexpensive, pretty darn high performance storage offering may be just what the doctor ordered for companies struggling with data management issues on top of financial issues. Whether the TP900 will succeed or not is of course unclear. Personally I think it will do well, but who knows? Times are tough and these are strange days. At the very least though, SGI has exhibited marketing savvy absent from many high-end vendors lately. The Origin 300 If I worked in marketing at SGI, this would be my tagline for the Origin 300: “For those who want the most power they can get in the least amount of space for least amount of money.” I think it’s quite appropriate these days. The system seems to have in mind those who need to run higher performance apps, simulations, what have you, then your average company but who also do not require something like an Origin 3000 or a Cray SX-6. Mid-Range HPC? SGI’s Origin 300 server is able to deliver substantial performance in a compact, relatively inexpensive, rack-mounted system. The base server module holds two or four MIPS® 64-bit processors, up to 4GB of memory and two disk drives in only 3.5 inches of rack space. Optional components add I/O bandwidth, or additional system bandwidth for scalability up to 32 processors in a single, shared-memory system. Two 32-processor SGI Origin 300 systems fit into a single, industry-standard rack. The Origin 300 base server module can work completely standalone with nothing else involved, as a standalone entry-level technical server. It has system disks, it has base I/O, it has integrated power. So there’s nothing else required to run it as a technical server. The key to scaling the 300 is the NUMAlink port on the back of the machine. As the user scales it, they can direct-connect a second base Origin 300 module, or if their application is I/O bound rather than compute bound, they can direct connect a PCI module. So, there’s a choice, and that is no doubt something SGI is counting on being appealing to users. “They can go to 8 processors, 8GB, or can go to 2 or 4 processors with two additional or fourteen total PCI slots. There are about seven buses that are all 64 bits, 66Mhz, so that provides very high I/O to CPU ratio,” said Addison Snell, Product Marketing Manager for SGI HPC servers. “If a user wanted to scale beyond that, a NUMAlink module has eight NUMAlink ports so they could hang eight additional modules off of that, either base server modules or PCI modules in just about any ratio.” “By making our modular computing NUMAflex approach into a more compact, affordable form factor…when you combine our Origin 300 with our high end Origin 3000, this is really a very compelling product line that offers the best performance or most powerful computer at any price point,” Snell continued. “And I’m continuing to say this despite IBM’s launch last week claiming something to the effect of the ‘most powerful computer’ overall.” Modular Computing is a term that’s been batted around quite a bit recently. You might even call it a buzzword. This is how it was explained to me… There is no backplane, there’s no central switch. It is simply (as indicated by the example above) modules and cables. Veritable plug and play upgrades for the high-end that eliminate the need for so-called “forklift upgrades.” It seems that if you wanted to line the various modules up on the floor and cable them together, you’d be good to go. Pretty hard to beat that. And as someone once said, “simplicity is elegance.” Target industries for the Origin 300 include: manufacturing, life sciences, government and media. -- Manufacturing: Automakers can use the SGI Origin 300 server to makecritical decisions earlier in the design process, including departmental-level virtual safety testing of models. On LS-DYNA(R) crash simulation software, SGI Origin 300 delivers about 50% more performance than comparably priced HP midrange servers, including the newly announced HP rp8400. Oftentimes in manufacturing, the emphasis is on shortening design cycle times. According to Snell, Origin is already the number one platform for CAE, or Computer Aided Engineering. The reason behind this is reportedly because of the big, shared memory scalability. When you get latent at the design phase and you’ve got a car model or your virtual model with millions of elements and you’re trying to do a crash simulation analysis on that while simultaneously solving all of your structure and weight problems, aerodynamic problems, noise vibration or harshness, they’re really getting into multidiscipline design optimization where they solve all those together, subject to each other’s constraints. No one else has the scalability Origin 3000 offers, again according to Snell, so that’s why SGI dominates in that space. But back up to much earlier in the design phase. Imagine the user still has four years until they plan to release their product. They’ve got some early prototype models that don’t have millions of elements; they’ve got a couple thousand. And if they’re doing crash simulation, they’re not looking at a whole crash, perhaps they’re only looking at the first couple milliseconds, looking for any gross fault points so they can change them early on in the prototype process before they get further along where significant monetary issues may apply. That’s where the Origin 300 is designed to prosper, and unseat HP, who has historically been very strong in this, shall we say, midrange space. -- Life sciences: Biochemists can use the SGI Origin 300 server to search for cures with high-throughput bioinformatics clusters. On BLAST application software, which compares protein sequences against large databases, SGI Origin 300 delivers about 50% more performance than PC running Linux(R) operating system cluster solutions with the same three-year total cost of ownership. In sciences, a big target area for SGI is going to be areas such as high-throughput bioinformatics. This is a hot area these days, and arguably the fastest growing area in what most HPC professionals would call Life Sciences right now. The need in this space would seem to be computational scalability, but because the computation is very parallel, whereas the manufacturing code scale in shared memory mostly, these might be scaled in a cluster implementation. Because of this, it’s been very popular to use Linux clusters to do high performance computing in this space since they’re so cheap per processor. According to Snell however, “If you only look at dollars per CPU, you’re really fooling yourself as to how much you’re paying to get to the same level of performance.” “Here’s where I’m going to stack up a bunch of Origin 300 modules, but rather than NUMAlinking them together, I’m going to Ethernet them together because I don’t need the overhead of the shared memory system. And I can get this as low as about $40,000-50,000 US list per 4 processor node, including a fair amount of memory. And this is very high density,” Snell said. “Just going like this I can get up to 76 processors in a rack because I don’t need any power bays or additional modules of any sort, I’m just cramming these in.” Snell continued, “So, because I was talking about Linux clusters and how they’re the ultimate dominant, cheap (as far as dollars per processor), that’s the comparison we did here. We took again a typical bioinformatics code, in this case a variety of BLAST, which does, you know, protein lookups against DNA databases, that sort of thing. And again set a performance level and took two systems that have the same level of performance. Either of these configurations do 105-107 searches per day, which meets this requirement, right? What you’ll see is that we offer a lot fewer processors to hit that same level of performance. “Their processors are so much cheaper that even though they’re deploying more of them, they have about half the hardware cost going in. But there are a lot more aspects to cost than that, especially if you look at the fact that, that solution is going to take up over five times as much space…it will draw more than five times as much power. It’s more than five times as hard to administer and I have to install the software on 32 separate nodes as opposed to 3. So it’s a lot easier total cost of ownership, in fact, it’s about half the total cost of ownership to go with an Origin 300 compared to a Dell Power Edge system. And again, we have all of the modularity benefits that you don’t get if you’re just using PC’s. No modular upgrades, you have to start over.” -- Government: The defense industry can use the SGI Origin 300 server for high-bandwidth applications such as satellite imaging and data processing for LIDAR (light detection and ranging) and SAR (synthetic aperture radar). A four-processor SGI Origin 300 server offers more than double the I/O bandwidth and scalability of a four-processor Compaq(R) ES40. Specific benchmark results are classified. “In government and defense, there are a lot of applications that are more I/O bound than compute bound. In particular, things like satellite imaging, radar processing, LIDAR processing, spectral image processing, etc. These are all areas where I/O throughput is the key measurement of performance, not computational scalability,” Snell added. “And there are also arms of government subcontractors who are as interested as possible in compact, rugged form factors because they don’t have compute rooms, they have things like submarines and helicopters and vans. There you’re really talking space constraint, you’re talking about whatever your compute room is on a submarine. There’s not a lot of space. Being able to take a supercomputer in small modules that fit down a submarine hatch: Big benefit. Then you just cable them together.” -- Media: Digital content providers can use the SGI Origin 300 server and SGI Total Performance 900 storage system to offer video streaming capabilities on demand over networks. The new solution offers the greatest stream density in the industry, with up to 3,600 hours of video-on-demand-the equivalent of 1,800 two-hour movies-available from a single full-size rack. As mentioned above when discussing the TP900, media streaming is a pronounced area of interested to SGI, particularly video or learning on demand. “…this combination of Origin 300 and TP900 is going to break a threshold people have been talking about in density for a while, where we’re going to have over a hundred video hours per rack unit, streamable,” Snell noted. “Think of that as nearly 4000 hours per rack, or 200 two-hour movies, streaming out of a single rack for about $30,000 per Origin 300 plus TP900 combination. Very low price per stream, unmatched stream and storage density.” I mentioned very early on in this article that the people of SGI seemed in high spirits. Perhaps they deserve to be. The new server and storage system seem solid and well suited to the changing needs of the HPC space, even if the products could be considered as towards its low-end. In any case, I found the following statement encouraging, primarily for SGI and its shareholders, but also for the industry as a whole. “The need for supercomputing power is not going anywhere, and in fact if you compare where you are going to cut back your investments if you are going to spend less on tech. If you look at a large automaker, they’re a lot more likely to maybe delay the purchase of a new payroll system than they are likely to jeopardize the design of their 2004 SUV, Snell commented. “And those areas of core investment are ongoing and in every one of our industries, SGI technology is being used in whatever the core area of research and development is, whether it’s designing a safer minivan or finding a new source of energy or searching the human genome for a cure to cancer. These are the core areas of research and development that people are going to be very hesitant to cut back their investments in.” ---------- Supercomputing Online would like to thank both Sylvia Crain and Addison Snell for their time and insights. It would also like to thank SGI’s Greg Slabodkin for his assistance. ----------