HPC storage vendor DataDirect Networks (DDN) will soon offer integrated clustered file system support in its Storage Fusion Architecture (SFA) product line. The idea is to drastically reduce the amount of storage switches and file system servers, and thus the cost and complexity of supercomputer-sized file storage. The initial product to provide this capability is the SFA10000E, which is being marketed as an "Open Storage Appliance"  by DDN.

To recap: the SFA10000, which the company unveiled at the International Supercomputing Conference in June, represented a design shift toward using commodity server-type hardware in its high-end storage platform. According to the company, embedding file system support in the SFA enclosure is just the first step toward the development of a unified multi-platform architecture.

The SFA10000 uses Intel Nehalem CPUs, DDR3 memory, and PCIe 2.0 as its foundational hardware. The design is based on a controller couplet pair that can support SATA, SAS, as well as SSD devices. At max capacity (using SATA drives) a single controller pair can drive 2.4 petabytes of storage. Data throughput per enclosure is 10 GB/second.

But because the SFA10000 is using a big chunk of DRAM for I/O caching along with the latest generation Intel CPU platform, it also is able to deliver lots of IOPS -- up to 1 million to internal cache storage and up to 300 thousand IOPS to external disk. In a system with a full complement of SSDs, DDN expects that number to approach 400 to 500 IOPS, although as yet no such system has been built.

"The Nehalem with the newer QuickPath technology to connect the processors together, and with the embedded memory controller is very, very good at moving lots of small bits of data, and hence is very good at IOPS," explains Josh Goldenhar, DDN's director of product management.

A high IOPS capability has become increasingly important, since modern multicore servers -- which, coincidentally are also based on x86 CPUs -- are sending multiple simultaneous I/O requests to the storage system. The result is that even sequential I/O ends up looking like random I/O at the storage end, thus the need for high IOPS on top of high throughput.

Since DDN's new storage architecture encompasses what is essentially a server platform, the company can now use it to bring applications inside. The low-hanging fruit is to add clustered file storage applications, specifically, the options to include a Lustre-based or IBM's GPFS-based file system on top of the platform's block storage. This will be implemented via DDN's ExaScaler platform (for Lustre) and its GridScaler platform (for GPFS). The integrated parallel file system/block storage platform will be sold as the SFA10000E.

According to Goldenhar, the big win here is being able to consolidate the file system server clients, network switches, and storage arrays into a single platform. "The goal for the Storage Fusion Architecture, from the beginning, has been to be able to collapse multiple layers of infrastructure into the storage itself," he says.

In addition, since the network layer, along with the ensuing protocol translations and data copying, has been eliminated, performance, especially latency, stands to be much improved. With the file system and block storage controller sharing DRAM, DDN has been able to map file reads and writes directly to the same memory used by the DDN RAID stack. On each controller in the enclosure, one of the quad-core Nehalem CPUs runs the DDN RAID stack plus manages the SAS HBAs, while the other CPU is available as an application processor, in this case, to run Lustre or GPFS processes. According to Goldenhar this is done via a virtualization scheme, such that Lustre, GPFS, and the network drivers can run unmodified. But, he says, this is implemented in such a way as not to introduce any significant overhead.

The fact that the storage filers are embedded in the platform should reduce datacenter operational costs significantly, given that a significant chunk of server and network complexity has been squeezed out. That should translate directly into lower requirements for power, cooling, and floor space, as well as reduced management. "We think you'll be able to build the next generation of multi-petaflop computers using far fewer components," says Goldenhar. Acquisition cost should be somewhat lower as well and be competitively priced, he adds, although at this point DDN has not offered a price list. The SFA10000E products will be generally available in the first quarter of 2010.

In the second half of 2010, DDN is planning to do a lot more with flash memory technology. In the current SFA offering, flash, in the form of SSDs, are only supported as plug-in replacements for spinning disks. But the packaging around the flash chips that turns a memory product into a storage drive product undercuts some of the potential performance of the technology. Goldenhar says that they intend to put flash memory "a lot closer to the controller," although he wouldn't divulge if they're looking at a PCIe flash design, flash DIMMs, or some as yet undefined solution.

In the longer term, the company is intending to open up the SFA architecture in a much more generalized way, but specifically for end-user applications and storage server virtualization. The rationale is the same as for the embedded filer solution: to enable tighter integration between applications and storage. For example, if a particular storage system is being used for checkpointing, a user might want to take advantage of the idle processors after a checkpoint completes to perform data reduction or to determine if the solution is converging. At this point, DDN is looking to support Linux, Windows, and perhaps even OpenSolaris applications, and is planning to include this support toward the end of 2010.