Virsto CEO Mark Davis said his company's Virsto One software can tackle common problems associated with enterprise data storage management in a server virtualization environment. Virsto is addressing the Windows Server 2008 R2 version of Hyper-V first, but Davis said there will also be a version supporting VMware Inc.'s ESX Server.
Virsto One is deployed within the hypervisor on the physical server, inside what Microsoft calls the parent partition. It intercepts I/O coming from guest machines and creates a sequential log file as a database would, so that writes and reads from disk are sequential in all cases. "Because of this log we can also optimize the layout of data when we destage it to disk," Davis said.
This approach, Virsto claims, can mitigate the "I/O blender" problem that comes when the hypervisor itself feeds multiple I/O streams from multiple guests to the underlying disk subsystem, causing contention and performance problems.
The log also makes "unlimited high-performance thin snapshots and clones" of virtual machines (VMs) possible, according to Davis, which is another selling point for Virsto One. "In a virtual machine environment, new virtual machines are deployed starting with a clone of an existing machine or template," he said. "Almost all the data is duplicated, causing virtual server storage sprawl."
Based on the idea that "the best data deduplication is not to create a duplicate in the first place," Virsto One uses headers and markers inserted into the sequential log to create space-efficient clones similar to NetApp Inc.'s space efficient snapshots. The similarities of the I/O intercept log to NetApp's Write Anywhere File Layout (WAFL) are also not lost on Davis, but he said the difference is the "anywhere" part – Virsto writes sequential streams of data rather than spreading blocks throughout the system, which WAFL does.
"Whatever the number of snapshots NetApp tells you they can do, add a couple of zeroes to it," Davis said. "There are already data centers with virtual machines numbered in the five figures today – six figures and above is not unimaginable."
Davis said Virsto One can do all this because of secret sauce in its software that allows for management of the virtual machine cluster with minimal communication among nodes, which Davis said is the biggest impediment to large-scale clusters today. "Our nodes talk once every few hours rather than hours or minutes," he said. "We're also mapping what the guest sees to optimize what look like multiple disks, while [storage] blocks are shared by multiple systems."
So how does this not recreate the disk "hot-spot" issue that was the problem in the first place? "The scalable, high-performance mapping technique we use is part of our secret sauce," he said. "We have to update maps very, very frequently and maintain the semantics of different virtual machines in a consistent state."
Improved storage allocation
While the idea of putting a startup's software in the data path to intercept virtual machine I/Os might be daunting for conservative IT managers, at least one beta tester said his deployment of Virsto One has improved his Hyper-V storage environment. Roger Johnson, technical lead, enterprise systems group at Charlottesville, Va.-based consumer electronics retailer Crutchfield Corp., said his company has been running Hyper-V in production since October 2008 because the software is cheaper than licensing VMware. However, because of the "I/O blender" problem identified by Virsto Software's Davis, storage in the Hyper-V environment was not efficiently allocated.
"I had been running one virtual machine per [disk] spindle," Johnson said. "But I wanted a ratio of two to three virtual machines per disk spindle to maintain my cost savings with Hyper-V." With Virsto One installed, Johnson said he's able to attach 50 virtual machines spread across two physical servers, attached to 9 TB of disk. "In testing right now we're gaining up to 30% more I/O over native Microsoft virtual hard drives and a 50% gain in snapshot performance," he said. Because of the snapshot performance boost, Johnson said he's also been able to deploy new disk-based backup for his Hyper-V environment.
Hyper-V first, and then VMware
Virsto Software's Davis said the company chose to focus on Hyper-V first for two reasons. As a newer product, Hyper-V doesn't yet have the equivalent of VMware's Virtual Machine File System (VMFS) clustered file system to specifically address storage I/O in an environment of clustered virtual machines. Hyper-V was also easier to integrate into than VMFS.
That's not to say VMware data storage management is perfect – friction between storage vendors and VMware over previous versions of VMFS has been well documented. Some of these issues have been addressed with new vStorage APIs released to storage partners with vSphere 4, but concerns about storage I/O remain, especially at scale
Davis said there's no timeframe yet for a version that supports VMware. "It's not going to be in the next three months, but I'm hoping for sooner rather than later," he said.
There are other well-documented problems with virtual machines and storage management that Virsto Software hasn't addressed yet, including the efficient proliferation of patches and other updates through linked virtual machine or virtual desktop clones, as well as how to address quality of service within the sequential log file. Davis said Virsto engineers are discussing both items as part of the product roadmap.
QoS is one item on Crutchfield Corp.'s Johnson's wish list for Virsto One. "Part of my beta feedback for them is that if they do start to see an I/O squeeze, I want to make sure certain [VMs] get priority," he said. Another item Johnson said he has asked for is integration between Virsto and Microsoft's self-service management console in addition to Virsto's current integration with the main Microsoft Management Console (MMC).
Virsto One will become generally available at the end of February. The firm's Davis said it will be priced at $1,250 per two-socket physical server and $2,500 per four-socket server.