SGI boasts of Altix UV installs
Prefabs Hadoop clusters to chew big data
Supercomputer maker Silicon Graphics says it has shipped more than 500 of its Altix UV line of machines in the past year, perhaps a larger number than you thought was possible and possibly indicative of the benefit of selling a Xeon-based system over one based on Itanium processors.
The Altix UV family is made up of three different distinct lines of servers, and the 500-plus installations cited above includes boxes in all three families, Jill Matzke, director of shared system marketing, tells El Reg. So don't get the wrong idea and start thinking that SGI has sold more than 500 of the high-end Altix UV 1000 machines.
UV is short for "UltraViolet", which was the code-name SGI used as it was developing the NUMAlink 5 interconnect and the system boards based on Intel's "Nehalem-EX" Xeon 7500 processors and related "Boxboro" 7500 chipset. The line includes the entry Altix UV 10 machines, announced in April 2010. It's basically SGI's riff on a standard 4U, four-socket Xeon 7500 server using the plain-vanilla Boxboro chipset and "Millbrook" memory buffers that all the other server makers employ. The Altix UV 100 and 1000 machines are based on a two-socket blade server and make use of the NUMAlink 5 interconnect.
The Altix UV 100s use the interconnect to make a 2D torus that links 48 blades – that's 96 processors and either 768 cores using the Xeon 7500s or 960 processors using the new "Westmere-EX" Xeon E7 processors released in April – into a shared memory system that spans up to 6TB of memory.
The big bad UltraViolet machine is the Altix UV 1000, of course, which lashes 128 of the SGI blades together using an 8x8 (paired node) 2D torus interconnect, yielding a 128-socket shared memory system that supported up to 2,048 cores and 16TB of global shared memory using last year's Xeon 7500s. The NUMAlink 4 interconnect using in the Itanium 9100-based Altix 4700 petered out at 512 sockets (1,024 cores) and 2TB of shared memory, by comparison. SGI could have used the quad-core "Tukwila" Itaniums to make a refreshed Altix 4800 if it wanted to, but decided instead to focus on Xeon processors and expand from its own variant of SUSE Linux to a machine that can run standard SUSE Linux Enterprise Server or Red Hat Enterprise Linux as well as Microsoft's Windows Server 2008 R2.
SGI has been making a lot of noise lately about how Windows Server 2008 can run on its Altix UV 100 and 1000 machines, and in fact, the UltraViolet hardware scales far beyond the limits of that Windows OS at this point. The Windows kernel sees the NUMAlink 5 interconnect as it would any other NUMA chipset grafted on top of an Intel chipset and runs any Windows application that a normal SMP server can. (That doesn't necessarily mean that it runs well, mind you.)
Windows Server 2008 R2 currently tops out at 256 cores (or threads if you have HyperThreading on your Xeon or Itanium processors) and 2TB of main memory. You can put eight complete, maxxed out Windows Server images on a single top-end Altix UV 1000 machine and run them all at the same time on the box. (No one is suggesting anyone should do this.)
While SGI is bragging a bit about its Altix UV installations, the company isn't providing any kind detailed breakdown by product number, blades shipped, revenues, or anything that would help us quantify how well the server line is doing. But Matzke was able to provide some color.
SGI's Altix UV super rack
She says that life sciences and defense contractors and integrators that build systems for the US government's military and intelligence community have been big growers in the past year for the Altix UV line. As for the product mix, Matzke says "it's a healthy mix of all three" and that although Windows support is in the early days on the Altix line, which only came out in February and was ramped up in March, SGI is "starting to see some uptake with Windows".
She would not hazard a guess at how far Windows could pump up sales at SGI over the long haul. But her boss, SGI president and CEO Mark Barrenechea – once the man in charge of application development at Oracle – gave some guidance about how the Altix UV line and support for both Linux and Windows had expanded the company's addressable market significantly.
Back in February, when speaking with Wall Street analysts, Barrenechea said that traditional high-end HPC workloads represented about a $1bn potential market, and that chasing RISC and Itanium shops running Unix workloads on large shared memory SMP or NUMA systems represented another $1bn or so in total addressable market now that SGI could peddle Xeon-based, large memory Windows and Linux boxes. Another $1bn in potential new market comes from running "big data" analytics and data warehouses on Windows or Linux atop the Altix UV line.
The good news for SGI is that it has broadened its customer base since its acquisition by Rackable Systems (which took its name) and the introduction of the Altix UV line. "We're selling to a higher number of customers with more modest systems than we did with our Itanium-based machines," explains Matzke. "Our installed base has grown pretty dramatically."
A larger installed base means that there is a great probability down the road that SGI will be able to sell much larger systems as workloads grow – and SGI is counting on this as well as word of mouth as companies deploy fat Linux and Windows systems on its shared memory architecture to help it grow.
Not all big data needs big memory
Not all big data jobs require a big shared memory system like the Altix UV 1000s. Some big data munchers, like Hadoop and other Map/Reduce variants, do better on clusters of machines. To that end, SGI has started preconfiguring Hadoop clusters based on its Rackable and Altix ICE clusters, which are made of standard x64 rack servers.
SGI is not religious about any particular processor, and SGI is building Hadoop clusters based on two-socket motherboards using either Intel Xeon or Advanced Micro Devices Opteron processors – whatever customers will cut the check for. SGI doesn't want to give out all of the feeds and speeds of these Hadoop setups, but gave El Reg some idea of what is in the racks and boxes.
The Hadoop compute nodes, sometimes called workers, come with 12 or 18 memory slots and use 4GB or 8GB memory sticks; they are equipped with six to twelve SATA disks, ranging in capacity of 500GB to 1TB, and are available with Gigabit Ethernet, 10 Gigabit Ethernet, or InfiniBand network ports.
The Hadoop master nodes, which have namenode, task tracker, and job tracker server programs, and are a bit fatter. They come from SGI in two- or four-socket variants, with each node coming with 18 or 24 memory slots. The chassis supports the same six to twelve drives, although with either fast and skinny SAS or slow and fat SATA options.
The Hadoop clusters are not just about server hardware. "Anyone can do Hadoop on their own since it is open source," explains Bill Mannel, vice president of product marketing at SGI. "What we have done is create extra value in terms of packaging, density, and management."
At least one SGI Hadoop customer is deploying them in ICE Cube containerized data centers. SGI is considering extending its Management Center cluster management tools so they can see into Hadoop clusters and visually display what is going on with the data chunking and chewing inside the cluster.
Most of the Hadoop installations that SGI has been involved with have customers plunking the open source Hadoop atop of Red Hat Enterprise Linux, but the company is happy to support SUSE Linux or CentOS if customers want that, or any number of commercial extensions to Hadoop.
Pricing information for the Hadoop prefabs is not being released by SGI. ®