When organizations virtualize, they typically overrun the I/O capabilities of the underlying storage infrastructure and aren't able to scale the virtual infrastructure as far as they would like. Instead of asking "why" and getting to the root of the problem of performance bottlenecks, they typically run blindly into an expensive rip-and-replace of the SAN architecture to create more I/O overhead and try to "flash" their way out of performance issues. More recently, administrators have begun to discover you can't "flash" your way out of virtual machine (VM) performance issues without overspending on hardware if you ignore the two big I/O taxes in a virtual environment that inflates IOPS (Input/Output Operations per Second) requirements and steals bandwidth from server to storage.
The two big performance penalties in virtualized environments to be aware of are the "Windows I/O tax" and the "I/O blender" tax. No matter how many spindles or how much flash is added to the infrastructure, much of that performance is ultimately robbed due to I/O characteristics that are much smaller, more fractured, and more random than it needs to be and that steals up to 50% available throughput from server to storage.
Small, Fractured I/O Tax
As the most virtualized operating system, Windows suffers from free space allocation inefficiencies at the logical disk layer that inflates the IOPS requirements for any given workload as the relationship between I/O and data begins to break down over time. This occurs because when Windows NTFS writes data in a SAN storage environment, file allocations become unnecessarily fractured across different addresses at the logical disk layer, resulting in every piece of the file requiring its own I/O operation to process. Instead of carrying an optimal amount of data with every I/O request, a single file may take multiple I/O to process instead of single I/O had Windows first employed intelligence about choosing the best allocation instead of the next available allocation. Consequently, this results in the first I/O tax: I/O that is smaller and more fractured than necessary.
It's not just every write that is subsequently penalized, but every subsequent read as well. It is common for a 32K file to be efficiently processed with a single I/O on day one when a file system is fresh and new, but as time goes on as files are re-written, erased and extended, I/O density suffers and ultimately systems require four 8K I/O operations or eight 4K I/O operations to process the whole 32K file. More fractured environments will experience hundreds of I/O operations to process a single file which is akin to pouring molasses on systems.
"I/O Blender" Tax
The second tax is that of the "I/O blender." This tax occurs when mixing multiple VMs on one server, and then connecting those various servers to shared storage. The result is a highly random I/O stream that diminishes the entire virtualized environment.
To understand this, think about the consequence of disparate VMs sharing a single host, routing otherwise sequential I/O traffic to the hypervisor where those I/O streams become "blended." The resulting random I/O pattern then gets sent to storage, which further dampens storage performance.
Clearly, while it hurts systems to be taxed with small, fractured I/O from Windows due to free space allocation inefficiences, it's even more damaging to take all that small, fractured I/O and randomize those I/O streams when they become mixed at the hypervisor. When virtualized organizations hit an I/O ceiling that requires higher performance than the company's storage infrastructure can deliver, administrators commonly think they need to buy more IOPS, when in fact the Windows I/O tax and the "I/O blender" effect has robbed throughput, making systems more IOPS intensive than they need to be. By focusing on trying to solve the root of I/O inefficiencies first, organizations can get to the bottom of the real issue that's wasting their current and future hardware resources.
A Better Solution
As an alternative solution, these I/O inefficiencies can be easily remedied by using I/O reduction software that targets the root cause problem so administrators get the most performance possible from their hardware infrastructure after virtualizing. Today's software has been shown to result in up to 300 percent faster application performance on existing systems. By optimizing the I/O profile, software intelligence can increase I/O density and sequential writes and subsequent reads while also leveraging available DRAM to target the I/O the steals the most bandwidth from VM to storage – small, random I/O. This reduces latency and frees the infrastructure from performance-diminishing I/O.
This approach not only protects a company's investment in its existing hardware infrastructure, but it also solves performance bottlenecks without disruption and ensures organizations can maximize future storage system investment. In short, I/O optimization software can more effectively solve the application performance issues for virtualized environments — without requiring any new hardware.
