December 14, 2007
Last month DataDirect Networks announced its 8th generation high performance storage offering, the S2A9900. The new system will double the 3 GB/second performance of the current S2A9550, reaching 6 GB per second per S2A9900 storage appliance. The new technology continues in the tradition of previous S2A systems in providing enterprise-class performance and reliability with inexpensive SATA disk hardware.
Like its predecessors, the S2A9900 architecture operates as a self-healing storage array that operates in real time. Unlike solutions with file system-based error management, the DataDirect philosophy is to let the I/O system localize error management at the level of the drive. Using RAID technologies, data integrity is managed below the application layer, which enables error detection/correction to be done on the fly and transparently to the user's software.
"We developed a product that always presents the same consistent operation in reads and writes regardless of error conditions or any type of failure below the control system," said Dave Fellinger, DataDirect's chief technology officer. "That's what sets us apart."
Using one terabyte disks, a single S2A9900 appliance can drive up to 1.2 petabytes of storage, which can all fit in two racks (20 4U drive enclosures plus the 4U storage appliance). Besides doubling the storage bandwidth per appliance, DataDirect has also doubled the bandwidth available to a host compute node by adding support for 8 Gbps Fibre Channel and 20 Gbps (DDR) InfiniBand. This allows for a more consolidated switching fabric, reducing the number of adapter cards.
Besides performance, the biggest change to in the new technology is the ability to mix serial attached SCSI (SAS) drives and SATA drives behind the same S2A appliance. SAS disks spin twice as fast as SATA, but are lower capacity. The idea is to be able to mix enterprise-class, low latency SAS storage with less expensive, high capacity SATA storage.
In general, SATA drives make sense where price and capacity are the most important criteria, as they are in most HPC applications. A few HPC customers may choose the SAS drive to overcome some level of paranoia about using SATA, but DataDirect's S2A technology was precisely designed to overcome reliability issues with the lower cost hardware. SAS makes most sense in areas outside of HPC, where non-sequential metadata access in the norm, such as in online video serving. In this case, the real-time nature of the user experience makes low latency SAS drives a natural fit. In some cases, HPC customers may prefer a certain amount of SAS capacity where non-sequential data access requires a level of latency that can't be met with SATA hardware.
Prior to the S2A9900 generation, the backend technology was based on the Fibre Channel protocol, which supported SATA through a bridge chip that made the drives appear to be Fibre Channel devices. DataDirect felt the architecture would be better served with a SAS backend, since both SAS and SATA are based on a compatible serial, point-to-point interface. This makes it much easier to deal with SATA commands over SAS and allows both architectures to be easily accommodated in the same storage system, as has been done for the S2A9900.
Although each appliance can read and write 6 GB/second to the host, a lot more bandwidth is available internally thanks to more powerful processing hardware, a shift to the faster PCI express bus, and the use of the SATA-friendly backend. Each appliance can achieve 24 GB/second, compared to 4 GB/second with the current S2A9550. This will allow expanded capabilities to be developed via software, making features like mirroring and other data migration strategies possible. Since there's so much bandwidth headroom available, these operations can be performed concurrently with application I/O, without compromising host-side performance.
Another new feature is a "sleep mode" that allows idle drives to spin down when not in use. The idle period is configurable so that users can select an appropriate amount of time for their own setup. With the growing focus on datacenter energy efficiency, there could be a lot of interest in this feature, especially for the kinds of applications that write data once, but only read that data at long intervals. DataDirect plans to introduce this capability in the S2A9550 and carry it into the S2A9900.
The S2A9900 also has the ability to cycle power on badly behaving drives, in attempt to get them back to a good state. According to Fellinger, occasionally the microcode in an otherwise sound drive loses its way, and the only way to reset it is with a power cycle. The idea is to avoid trashing good -- if somewhat flaky -- drives in order to minimize hardware replacements and data rebuilds. In installations that deal with classified data, a disk drive replacement is extremely inconvenient and is best avoided whenever possible. Taking a drive down for a minute or so while the power is cycled is not a problem for S2A technology, since the RAID implementation allows the storage system to operate at full speed with a missing drive. When the drive comes back online, the system restores the missing writes transparently.
"The whole point here is uptime without human intervention," says Fellinger. "If you're running an HPC center with tens of thousands of drives, that becomes a very, very important attribute."
Although DataDirect won't ship the S2A9900 until Q1 2008, the company has already announced that Argonne National Laboratory has purchased an 8 PB storage setup for their new Blue Gene/P system scheduled to be installed next year. The storage will be driven by 17 S2A9900 appliances and contain over 8000 1 TB SATA drives. Aggregated storage bandwidth is expected to reach 100 GB/second.
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?