Marvell Tackles The I/O Problem

Marvell Technology Group--primarily known for its storage, communications and consumer silicon solutions--is unveiling a virtual storage accelerator that addresses the I/O choke-point that has been plaguing the IT industry for years. The company says that its DragonFly VSA board, which plugs directly into all commercially available servers, will deliver 10x server I/O performance; reduce network-attached storage/storage-area network/direct-attached storage (NAS/SAN/DAS) costs by more than 50 percent; and cut power and space costs.

"For the first time in many years, the industry has started to really do something about fixing the I/O problem we've been suffering from for decades," says analyst Arun Taneja. "We've seen the performance gains in processors from Intel ... we've gone through the roof in improvement in CPU, but on the disk side ... it's so pathetic".

While Marvell isn't the only vendor addressing the I/O bottleneck, it is somewhat unique in that it is doing so from the server rather than the storage side. "It's absolutely a paradigm-shifting thing that they've done," says Teneja. "I wouldn't be surprised if they weren't underestimating the gain you can get from this."

As an industry-standard PCIe adapter, the DragonFly VSA offers a turn-key appliance-on-a-card that fits seamlessly into all commercially available rack-mount servers and operates system-independently.

Its distributed host I/O cache employs a two-level cache for very low-latency and high-burst throughput. The two levels comprise a Level 1 non-volatile DRAM cache and a Level 2 cache created from off-the-shelf commodity solid-state drives (SSDs). Marvell says that the cache is storage protocol-agnostic and can be configured to serve as a data cache for DAS, SAN or NAS arrays.Expected to ship in limited quantities in the third quarter, the solution is built around the company's HyperScale technology, and its multiprotocol storage caching supports Network File System (NFS), iSCSI, Fibre Channel, Fibre Channel over Ethernet and SCSI DAS.

Taneja sees VMware shops jumping at DragonFly. "All the low-hanging fruit, the less-than-mission-critical applications, have all been virtualized. What's left to do is all of the mission-critical applications. The No. 1 reason why they haven't been doing this is I/O."

He says that IT pros have been telling him for the last two years that they love VMware and average 30 percent virtualization of their applications, but would virtualize their mission-critical apps "over my dead body. This kind of product is going to make those customers, after they test it out and satisfy themselves it's rock-solid, take that 30 percent and double it to 60 percent." And while this isn't just a VMware problem, Taneja says, the other hypervisor vendors--Citrix and Microsoft--haven't gone far enough to run into this I/O problem.

The storage vendors aren't going to like DragonFly and similar offerings because they will impact the typical I/O problem response--just throw more storage at it, says Taneja. On the other hand, they're used to seeing these types of disruptions so they will adjust, he says. He thinks the server vendors will be happy because the card is going into their slot and it makes the server more powerful.

See more on this topic by subscribing to Network Computing Pro Reports Research: Storage & File Virtualization (subscription required).