John Lee, VP of Advanced Technology Solutions at Appro, sees the HPC space as a continuum with customers focused on reliable, cost effective, turnkey systems at one extreme, and engineering or scientific organizations focused solely on performance at the other end. Where does his company fit? According to Lee, closer to the turnkey end.
But turnkey doesn’t mean small, or nontechnical. Appro has been in the news lately with some very high-profile wins: an aggregated 438 teraflop purchase at the three DOE weapons laboratories; a 95 teraflop system for the University of Tsukuba; and, this week, a 38 teraflop machine for the Renault Formula 1 racing team.
Appro’s supercomputing system strategy is centered on its Xtreme-X Supercomputing Series. The Xtreme-X line, launched at SC07, is built on the idea of a Scalable Unit (SU), a concept that other HPC companies have incorporated into their lines. As Lee explains, “The idea of a Scalable Unit is to bring to supercomputing a Lego building block approach to assembling clusters.” As with Legos, SUs are the basic building blocks of larger clusters. A single SU can hold up to 128 sockets of either quad-core Xeon, with the Xtreme-X1 system, or quad-core Opteron, with the Xtreme-X2, announced just last week. Multiple SUs are integrated to form a single system. When shipped, they arrive at the customer site integrated, pre-packaged, burned in, and ready to go.
The DOE tri-labs purchase announced last October took advantage of the SU concept; the initial procurement purchased 21 SUs to be incorporated into eight clusters. The largest of these, the 8 SU system, will have a 162 teraflop peak performance.
Appro says that building its systems out of standard, replicated units ensures it can install and integrate a new system rapidly, drawing upon the experiences of the identical SUs it has previously configured. This has been a real benefit for LLNL, one of Appro’s regular customers. Brent Gorda, a system architect at LLNL, agrees, saying that the last cluster in their Peloton acquisition went from ship to customer handover in just 3 weeks. According to Gorda, Appro’s service and system integration skills are very strong. “Appro is doing a better job than the tier 1 vendors would do for us,” he says.
But as Appro CTO Jim Ballew explained in late 2007, the company’s customers are looking for more than just rapid deployment. As the HPC market matures, customers want reliability, manageability and availability while maintaining good price/performance. This is a challenge that all HPC vendors face. Customers want to drive Cadillacs, but they want to pay for Toyotas. Appro hopes to navigate this cost/feature tension with specific design decisions in its system architecture and enhanced features in the Appro Cluster Engine (ACE), a software stack, currently in beta, that will be deployed in future systems, including Tsukuba and Renault.
As Ballew points out, the nodes in a system are inherently redundant (since there are many of them), but the network isn’t. The Xtreme-X line addresses this problem by incorporating redundancy in both the management network (GigE) and data network (InfiniBand). Software has been added that allows the system to recover when one of a node’s IB links fails with MPI messages already in flight. The system is able to retransmit the lost messages on the surviving link transparently to the application. Also, according to Lee, since the systems are built from more cost-effective 24-port switches rather than 288-port switches, reliability is made more affordable.
Reliability is a big motivator for the dual-rail data (and management) networks, but it isn’t the primary motivator. According to Ballew, they included redundant InfiniBand networks for reliability and bandwidth, but the primary benefit is for better communication latency. Both IB channels can be used for transmit and receive, and ACE manages this by queuing messages to the shortest queue first.
“Although other systems support dual channel InfiniBand,” says Ballew, “this doesn’t normally help short packet performance. Traditional implementations use a primary channel for all communications, only using the second channel to sustain bandwidth on large messages.”
Ballew also identifies disks as a frequent source of failures in large HPC systems, and typical system architectures that feature an operating system disk on each node can lead not only to reliability issues when individual disks fail, but also to management overhead as multiple copies of the operating system have to be maintained throughout the system.
HPC vendors have addressed the management overhead issue through more sophisticated system management tools, but this doesn’t help reliability. Appro’s approach is to pursue a diskless node. Each rack has a server that coordinates with a management server to receive and cache the root file system for nodes in that rack, and the system’s servers are organized hierarchically. The hierarchical organization and caching strategies mean that systems can boot in relatively constant time from 100 to 1,000 nodes.
With this architecture, administrators need maintain only a single copy of the OS for the entire system, and Appro’s software environment supports revision control for OS images to ensure that undesirable changes can be rolled back to a known stable configuration. The management and I/O servers that handle OS image provisioning are also redundant, protecting against failure in these components.
Another benefit that Ballew sees is the fact that the system, while supporting diskless operation, uses a standard distribution of Linux — although Appro does add some special sauce in the communication layers. To handle the potential for MPI messages to arrive out of order when recovering from a lost InfiniBand link, Appro added additional capability between the OFED Verbs layer and MPI. They also added a layer to support mesh (and torus) network architectures with the InfiniBand fabric, rather than simply the traditional fat tree. According to John Lee, this support was added to control the growth of costs in the network when building out very large systems.
ACE is Appro’s opportunity to differentiate itself from the crowd of hardware resellers. These innovations here won’t matter to all of Appro’s customers — the DOE labs, for example, run their own software stack. But the company’s strategy of providing solid delivery and integration with turnkey configuration and added reliability features for those whose want it seems to be a solid one.
Appro seems to have learned from the mistakes of past HPC companies, and is managing its growth carefully. According to Lee, it is building out North American support with its own organization, but taking a more measured approach with international deployments. In the case of the Tsukuba installation in Japan, service is provided through a special relationship with Cray Inc.
Lee indicated that this kind of support arrangement is negotiated where it makes sense. Cray is a partner with a lot of credibility in the HPC market, and this partnership allows Appro to delay building out a global services organization until there is a sustainable business to support it. But what is Cray’s interest in providing services for a competitor? Cray spokesperson Erin McGhee said that while the company’s focus remains on developing supercomputers, it will be increasing technical services in those areas “where it is beneficial and strategic for our customers.”
HPC is a tough business, as evidenced by the recent implosion of Linux Networx, which at one time seemed poised to step into the first tier of HPC vendors. All of the recent froth in the HPC market around the new mid- and low-end customer is attracting a lot of attention and competition from mainstay and startup HPC companies alike.
At the high end, Appro will be competing with established tier 1 vendors that have a long history with a relatively small customer base, although Appro has certainly demonstrated that it can make headway in this market. In the middle and at the bottom, they will compete not only with some of the same tier 1 vendors, but also with existing tier 2 and 3 vendors, as well as with new entrants to the market.
To make matters worse, all these vendors are trying to identify and establish this new market at the same time as they are trying to build a business on it, in an economy that is softening globally. Appro is going to have to work hard for every bit of success that comes its way.