Virtualization often focuses on core computing elements like processors and memory, but I/O and disk storage also play important roles in the overall performance and responsiveness of virtual machines. Disk storage and I/O might be even more critical because disk-related functions are so much slower than processing tasks. Virtualization administrators who are eager to enhance virtual machine performance should invest the time needed to optimize disk operations.
Disk options that enhance VM performance
In virtualization, a hypervisor abstracts the workloads from the physical hardware that runs underneath, which allows easy allocation and sharing of computing resources, convenient migration of workloads and other features. Although modern hypervisors and virtualization-compliant processors impose very little overhead, there is a performance penalty introduced by the virtualization layer.
When disk performance is critical to a workload, some administrators may opt to configure the associated
The problem with pass-through mode is that some important virtualization features, such as VM snapshots or clustering support, aren't supported. As a consequence, the VM may actually benefit more from virtualization features than it would from the marginal performance improvement in pass-through mode. Administrators will need to evaluate the needs of each VM and determine the suitability of pass-through mode.
In addition to pass-through mode, hypervisors like Hyper-V can also apply other disk storage options. For example, fixed-size disks allocate all blocks for data and overhead on a fixed-size .VHD file at the time that .VHD file is created. The fixed-size disk cannot change once created. However, dynamically expanding disks creates an initial .VHD file with no blocks, and space provided as data is written to the .VHD -- up to the specified size of the disk. This is similar to the notion of thin provisioning because even though a sizable disk may be created logically, actual disk space is really only used when there is data to write.
A differencing disk is a special type of dynamically expanding disk. The idea is that a parent disk holds a fixed image and a differencing disk is associated to that parent, so any writes that change disk content are written to the differencing disk instead of the .VHD file. Reads are first checked against the differencing disk's .VHD file, and if no changes are present, the parent .VHD file is read. Differencing is a good choice when standardized disk images are needed and rollback capabilities are important, but maintaining parent and child disk configurations can be challenging for administrators.
Allocating the right amount of disk space
There is no single right amount of disk space, because many variables can affect the allocation of computing resources. Ideally, a workload running in a VM should need the same amount of computing resources that it would demand if deployed on a physical server. However, virtualization relies on a software hypervisor, and the added computing needed to operate a hypervisor will add some overhead to most virtualized workloads. As an example, Microsoft suggests that a virtualized workload should receive 105% to 110% of the disk resources needed by the same workload in a physical environment.
Still, it's important to note that this is only a guideline that should be applied loosely, because every application has unique resource requirements, performance needs, user traffic patterns and workload growth expectations. Administrators should consider each of these factors carefully before provisioning disks by testing and benchmarking in a test and development environment before rolling the workload out to production.
In addition, storage can be an expensive commodity, and over-allocating storage can be costly for an enterprise. Administrators can often employ technologies like dynamically expanding disks or other thin provisioning tactics to conserve storage space until it is needed, or use technologies like data deduplication to remove duplicate content and reduce storage demands.
Guidelines for physical disks to enhance VM performance
Storage performance – whether for virtualized or physical applications – always begins with proper design choices at the physical level, and implementing better-performing storage devices will have a positive impact on VM storage performance. The choices typically relate to disk size, speed, spindles and data layout.
For example, select smaller 2.5-inch form factor disks rather than larger 3.5-inch disks. The smaller platters have a correspondingly smaller circumference which can support faster rotational speeds for lower latency as well as faster track seek times. Smaller disks can find data much faster than larger disks. As an added bonus, the smaller platters take less energy to rotate, saving energy costs for data center storage.
Another factor to consider is the composition of disk groups. Fewer disks generally do not perform as well as disk groups because spreading data across multiple disks allows multiple disks to seek at the same time, and this improves performance. Instead of consolidating disks, rely on disk groups like RAID 5 or RAID 6 which support multiple spindles and provide comprehensive data protection within a storage array or server.
If possible, adopt dynamic data layout schemes that will automatically locate the most important or frequently-accessed data at the outermost disk tracks. Remember that the entire disk platter spins at the same speed, so the outer tracks are actually flying under the read-write heads faster than the inner tracks. This gets data onto and off of the platters faster, though the overall disk speed is still limited by the disk cache size.
VMs rely on storage, but limitations and bottlenecks in storage systems can noticeably impair VM performance. Using pass-through disks can offer a marginal improvement, but the loss of virtualization-related functionality is rarely worth it. VMs may require some additional storage compared to physical deployments, but the exact amount is best determined with hands-on testing, combined with established technologies designed to mitigate storage demands.
This was first published in January 2014