In-Depth

VMware Meets Storage, Part 2: New Game or Same Old Story?

Do VMware engineers have a clue about storage?

When listening to CIOs and IT managers explain their VMware strategy and what it is expected to provide in terms of consolidation, flexibility and reduced energy costs, it is easy to understand the appeal of this technology. However, hearing server administrators work through the complexities of configuring storage for use by a virtualized server environment, or reading the numerous blogs about the performance issues associated with the variety of storage connectivity approaches, you might conclude that the VMware engineers didn’t have a clue about storage.

That isn’t quite true, according to Jim Price, CEO of Fairway Consulting Group in Sunrise, FL. He has been working with VMware in its many iterations for over a decade. He notes that there were tremendous differences between versions of VMware, and that 2.x is almost a completely different product from 3.x. Much of the confusion about virtualized storage and servers, especially in the blogosphere, he argues, is that "no one is keeping guidance up to date."

Yes, he confirms, there are several ways to connect storage to VMware virtual machines (VMs): creating "datastores" is one, which Price calls a marketing term for a VMware File System (VMFS) partition in a storage repository. Another approach is to map VMs to raw LUNs exposed by the storage. A third way is to connect to a Network File System (NFS) share.

Based on his company’s extensive experience deploying and configuring VMware version 3 infrastructure on behalf of clients, Price says the datastore technique has enormous value to offer because "VMware ESX server provides simple management and does all of the heavy lifting for you."

He quickly qualifies this remark: "Don’t get me wrong. You still have to configure your host bus adapters, you still have to create zones in your SANs, and so forth. When you failover to another VM, you still need to rescan your storage to pick up changes. But, that has less to do with VMware than with Fibre Channel itself."

Contrary to some vendor remarks, notably those of Network Appliance in their white papers on configuring VMware with NetApp filers, "you can put virtually everything on a datastore. The performance of applications running on a VM connected to a datastore is very comparable to the performance of apps running in a raw LUN mapping environment," Price says.

Raw LUN mapping is typically used with Oracle, SQL Server, Exchange Server, and other transaction-oriented environments, Price notes. "This has nothing to do with performance and throughput, but with the need for channel diversity and control—to control I/O pathing through the Fibre Channel fabric."

The difference is linked to queuing of commands and data being exchanged between the application and the storage, Price says. "VMware uses no queues, or rather provides direct queuing. This works for a lot of applications, but not necessarily for transaction-oriented applications like databases or e-mail."

He emphasizes that VMware itself changes nothing about storage. Applications, whether running in VMs or in physical servers, determine the kind of storage connectivity, performance, and throughput that you need. The message he thinks needs to be expressed clearly is that the only significant change that VMware might force in the current sever-storage paradigm has to do with the number of VMs on a particular hardware host. Having 50 VMs on a particular host connecting to the same datastore will have an impact on application performance, just as sharing storage partition among 50 physical servers might reduce overall performance.

"The irony," Price says, "is that 98 percent of consumers still have no concept of the differences in the ways that Windows consumes resources versus the ways that UNIX uses resources. Windows is a shared-nothing operating system, while UNIX has been designed to share a lot of resources. At the center of VMware 3 is a multihost concurrent shared-use model." Configuring storage to support VMware must take into account not only the number of VMs sharing the same datastore but also the operating systems running in the VMware virtual machine and what they require.

Again, Price is quick to point out, this supposed constraint of the datastore method is improperly construed as a VMware constraint in blogs and storage vendor marketing literature, particularly from Network Appliance. The allocation of storage to a VM is fundamentally no different than the way storage must be allocated in a physical server environment: application and operating systems must be considered.

"We don’t sell Network Appliance," Price clarifies, stating that the vendor uses obfuscation and scare tactics to sell its wares. His company does sell EMC, Hitachi Data Systems, IBM, and Hewlett Packard. On a technology level, he does not recommend using NFS-mounted storage with VMware regardless of the vendor of the NAS product.

"VMware has changed its partitioning model to optimize it for machine-level performance. You have dedicated LUNs which, in turn, map to granular LUNs and you have a native file system in VMFS. Using NFS creates challenges. Someone else’s file system is being used on the storage device and you have no implicit knowledge of the underlying configuration or layout of the data."

"VMware," concludes Price, "masks the complexity of the server and changes nothing about storage. Storage itself hasn’t changed. VMware charges $3000 for a class that conveys this pearl of wisdom. In the final analysis, however, the storage you use needs to be appropriate to the application you are running."

What about all of the marketing information from storage vendors extolling their special VMware support features? Price says that marketing materials don’t tell the truth.

To him, the important discriminators are the same in VMware environments as in physical hosts, with the possible exception of server booting. In a physical server environment, storage is exposed as raw LUNs, and the server can boot from any kind of storage over any kind of connection except for software iSCSI and NFS. "You can boot from hardware iSCSI, but nobody does it that way: software iSCSI rules," Price notes, adding, "In a VM environment, you can boot from any datastore."

"One difference worth mentioning in a VM environment," Price adds, "is that you can backup datastores [coherently] only from Fibre Channel or iSCSI storage. NFS-mounted storage doesn’t provide visibility into the underlying physical distribution of data or file system layout. So, you are blind and you need to pay the penalty of NFS overhead."

Price’s observations make enormous sense and create questions about the "VMware compatibility features" that are being touted by just about every storage vendor today. How a vendor is making storage simpler to allocate to servers, whether VM or physical machines, is what really matters.

Storage virtualization vendors like DataCore Software in Ft. Lauderdale, FL have been quick to seize upon this fact. By virtualizing the storage presented in a Fibre Channel fabric or iSCSI SAN, the company can legitimately claim to ease the mapping of storage to servers.

Arguably, LeftHand Networks is also simplifying the server-storage connection. As I’ve discussed in this column before, LeftHand blends a unique iSCSI storage clustering play with a Microsoft Multi-Path I/O driver plug in that lets storage clustering live closer to the operating system kernel. Announced at VMworld last month, LeftHand has now nuzzled its clustered storage handler up closer to the VMware ESX server, as well, demonstrating the value of separating value-add storage functionality from proprietary hardware.

What remains to be seen is whether a true virtualization standard will appear, either in the server or the storage world, that will reduce the complexity of the mapping of resources between the two domains. Otherwise, one can expect that, as VMware’s current lock on server virtualization mindshare erodes—which it most certainly will as competitors ranging from Sun, Microsoft, and Citrix Systems to less well-known brands appear in the market – we will no doubt be dealing with more storage vendor hype about their affinity with the latest v-ware craze

Your views are welcome: jtoigo@toigopartners.com.

About the Author

Jon William Toigo is chairman of The Data Management Institute, the CEO of data management consulting and research firm Toigo Partners International, as well as a contributing editor to Enterprise Systems and its Storage Strategies columnist. Mr. Toigo is the author of 14 books, including Disaster Recovery Planning, 3rd Edition, and The Holy Grail of Network Storage Management, both from Prentice Hall.

Must Read Articles