Marvell DragonFly VSA

PCIe board for industry standard system that can accelerate storage I/O performance ten times.

I've been hearing quite a bit from suppliers offering storage acceleration through the use of high performance cache memory. Marvell's DragonFly VSA (Virtual Storage Accelerator) is clearly another entry into this competitive market. Marvell has developed a card that plugs into the PCIe slot that can be found in industry standard systems. Once installed, storage I/O traffic can be greatly accelerated.

Here's what Marvell has to say about DragonFly VSA

With the proliferation of mobile devices—smartphones, tablets, netbooks—consumers are reaping the benefits of an "always on" lifestyle. To meet these growing demands and cut infrastructure costs, public and enterprise clouds are increasingly using virtual machines (VMs) to consolidate applications onto fewer servers. But in solving this challenge, virtualization has inadvertently created another: NAS/SAN arrays cannot keep up with the flood of random write and read I/O requests, resulting in a storage server bottleneck.

Marvell tackles this problem head-on with a revolutionary turnkey solution: the DragonFly Virtual Storage Accelerator (VSA). Designed to cost-effectively break the I/O barrier in public and enterprise cloud computing environments, Marvell DragonFly is a cloud computing I/O acceleration adapter that enables incrementally scalable virtualization while significantly reducing storage costs.

Key Benefits of DragonFly:
  • Yields 10x higher virtual machine I/O per second while lowering storage disk and CPU overhead by more than 50%
  • Enables automatic write/read caching in application servers on its way to and from higher latency, higher capacity shared storage.
  • Designed as a turnkey "appliance-on-a-card" that fits seamlessly into all commercially available rack-mount servers and is operating-system independent.
  • Enables a cost-effective distributed cache that incrementally scales IOPS and NVRAM/SSD capacity as your application servers grow over time.
  • Marvell HyperScale Technology
    At the heart of DragonFly's innovative architecture is HyperScale™ embedded cache technology, combining sophisticated software and hardware-assist engines running inside purpose-built SoCs (system-on-chips).

Snapshot analysis

In the recent past, I have spoken with a number of companies that are trying to find ways to improve database or storage performance. In almost every case, the supplier has developed a way to:
  • use system memory to create a distributed cache for data, this is sometimes refered to as memory virtualization
  • use onboard memory on a system device (this is where Marvell's DragonFly VSA fits)
  • use non-traditional memory devices, such as solid state disks (SSDs) or dynamic random access memory (DRAM), as a way to accelerate storage I/O. Where and how this device is used depends upon the suplier

Marvell's idea appears interesting and could certainly improve the overall storage performance of workload that executes on a single system. The approach appears to have two shortcomings:

  • If an organization's workloads involve multiple systems adding, deleting and updating data, a Marvell device would be needed for each and every system.
  • A driver needs to be installed for the device to work properly.

If one considers how many servers organizations have deployed and how many different operating systems, operating system versions, and patch levels for these operating systems, the Marvell approach could require quite a bit of installation work.

I'm also concerned about hardware and software support once the Marvell device has been installed. Although Marvell claims to support every major operating system and hypervisor, I can't help but wonder that if a problem arose, would the discovery of the root cause require the de-installation of Marvell's hardware and software? Would suppliers of operating systems, database management engines and the like support the combination of a system, their software and the Marvell device/driver? The answers to these questions aren't clear.

Other approaches, ones relying on shared storage appliances or stoage devices based upon SSD or DRAM that attach directly to an organization's storage servers seem to be much simpler. These approaches don't require changes to system hardware or software to work.

In the end, each organization is going to have to weigh the benefits of Marvell's DragonFly or any other approach to storage acceleration versus the problems that approach could cause.

Marvell's approach appears clever and the performance it offers may be worth the challenges.