October 05, 2009, 1:38 PM — Virtualization is moving further into the datacenter all the time, with even critical large-scale applications like Exchange and databases now being virtualized. Whether you're using virtualization to make large applications more manageable or to consolidate many small applications, a server with lots of RAM, lots of processor cores, and lots of I/O is a good thing. And so is a SAN packed with features that ease the management of storage for virtual machines.
From a storage viewpoint, each virtual machine uses a file to simulate a physical hard disk. This file, a VMDK (Virtual Machine Disk Format) under VMware or VHD (Virtual Hard Disk) under Microsoft, can be located on a server's internal drive or on a SAN. There are several advantages to putting the file on a SAN: The file can be duplicated using the storage's snapshot function, the file can be moved easily from one hardware server to another for scalability or fault tolerance, and the storage itself can be more easily made fault tolerant.
[ See the Test Center reviews: VMware vSphere 4: The once and future virtualization king | Microsoft's Hyper-V R2 is hot on VMware's heels. ]
I tested five midrange SAN systems that deliver the goods for virtual environments: the Compellent Storage Center 4.0, the Dell EqualLogic PS4000, the HP StorageWorks 2000sa G2 Modular Smart Array, the Pillar Axiom 600 from Pillar Data Systems, and a build-it-yourself pairing of the Promise vTrak E610f hardware and DataCore's SANmelody 3.0 storage software. This review focuses on these systems' virtualization-friendly features. For a close look at the full range of their capabilities, see Network World's companion review.
The prices (as tested) of these systems range widely, from less than $10,000 for the Promise Technologies and DataCore combination to $130,000 for the Pillar Data system. All the manufacturers have models ranging from inexpensive starters to very high-performance datacenter-ready systems. As you can tell from the range in price, the models I tested don't necessarily compete with each other.
Virtual Machine magic Using a snapshot to duplicate a VMDK or VHD file bypasses the difficulty of backing up a virtual disk while the VM is running. It also makes it easy to clone a working VM, convert the snapshot to a new volume, and mount the new volume as a new VM. The process of converting snapshots to a new volume and mounting the new volume varies widely among the SANs tested, not so much in the procedure involved, but in the time required to unify all the snapshots into a new version of the volume. In testing, I created a 100GB volume, put several VMDK files on it, made changes to the VMs over several days, then cloned the volume and mounted the clone as a new volume. Depending on the storage system, the time required to create the new volume ranged from a few seconds to almost 20 minutes.
My five SANs also differed in their connectivity options, a factor that made an apples-to-apples comparison of I/O performance infeasible. For example, the HP is Fibre Channel only, the Dell is Ethernet (iSCSI) only, and the other three systems support both Fibre Channel and Ethernet. My testing therefore focused on features and functionality rather than IOps, which can vary widely even across SAN models from the same manufacturer.
Fibre Channel versus iSCSI is one of the major decisions you'll need to make when configuring a SAN. In the virtualization space, Fibre Channel has an edge if you want to put a lot of VMs on a single server. This is because network port count and throughput tend to be bottlenecks anyway; if you're using Ethernet ports for both storage and network, it's that much harder to get a lot of VMs running on the same server. With multiple VMs contending for a single iSCSI connection, performance can really suffer.
The way around this is to use multiple gigabit connections or one 10-gigabit Ethernet connection, but 10Gbps connections and switch ports aren't any cheaper than Fibre Channel, and you can stuff only so many four-port Ethernet cards into a server. The exception is when performance isn't as much a consideration as cost -- iSCSI can be less expensive.
