Mistakes When Trying to Solve the I/O Problem With SSDs for Virtualized Servers

One of the biggest problems in virtualized computing is I/O performance. Increasingly powerful multi-core servers have the CPU horsepower to host dozens of virtual machines but the I/O capability has not kept pace with this mammoth increase in computing power, and the aggregate demands of all virtual machines pose a bottleneck to unlocking the full potential of consolidation and cost savings promised with server virtualization.

Flash-based solid state disk provides an inflection point in server storage with the potential to solve the I/O bottleneck issue for virtualized infrastructures. But to take full advantage of the performance potential of SSDs, a different approach to I/O is required, according to IO Turbine, Inc., an early-stage company preparing to introduce a software solution to address I/O bottleneck issues in VMware environments.

To date, the cost of using SSDs has been too high and the utilization rate too low to make SSDs economically viable in virtualized environments. Companies have pursued a laundry list of techniques without success, searching for a method to simply and efficiently deploy SSDs in VMware and other virtual infrastructures.

Here are five common mistakes to avoid
when trying to solve
virtualization-imposed I/O bottlenecks:

1. Using locally attached Flash SSDs as block devices to provide high IOPS storage: Using flash SSDs in this configuration loses vMotion functionality and the ability to migrate live virtual machines. It also requires management-intensive sizing of SSDs for each virtual machine and a reconfiguration of application and primary storage. Flash capacity is wasted as VMs start and stop when dedicated to a specific machine and requires a lot of manual intervention each time you deploy more SSDs.
2. Adding high-performance hard disk drives and extra spindles to increase IOPS: This option requires some or all drives in the storage array to be upgraded to expensive 15K Fibre Channel or SAS drives. Additionally, hard disk arrays are typically over-provisioned with unwanted capacity in an attempt to wring out extra IOPS, making it a poor choice in terms of return on investment to solve I/O bottlenecks.
3. Adding Flash as special primary storage using an SSD array: This option also requires the intervention of storage administrators to specify which applications use SSD storage and then tuning the system to take advantage of the SSD performance potential. I/O performance improvements can easily be offset by the higher administrative overhead to make the system run properly. While many enterprise-class storage providers offer automatic tiering with data migration to and from the SSD storage, these solutions typically take place well after the need for the I/O acceleration has passed. Implementing SSD within the primary storage array does not eliminate the network latency incurred going to and from the application and SAN or NAS storage and thus negates the low latency advantage of SSDs.
4. Not Running I/O Intensive Applications on a Virtualized Host: One sure way to avoid I/O bottlenecks in virtualized environments is to restrict the applications that run on virtual machines to those that don’t stress the I/O limits of the system. This again adds to management overhead by forcing administrators to carefully evaluate applications based on their I/O requirements and offsets the server consolidation and cost savings generated by server virtualization.
5. Adding More Fibre Channel SAN or IP Bandwidth: This increases hardware costs without solving the problem. Increasing bandwidth between virtual machines and storage doesn’t make the data pipe faster. It is of limited value without also increasing IOPS at the storage layer. This approach only addresses part of the problem with all of the cost.

The virtualized server I/O problem will only get more acute as enterprises continue to pack more virtual machines on servers with faster processors with higher core counts. Flash SSD has largely bridged the historical performance gap between processor and storage speed, but VMware and other virtualization software solutions have yet to incorporate tools and techniques to enable virtual machines efficiently and easily incorporate SSD without manual intervention into the workflow and allow I/O-intensive applications to be deployed without recreating traditional bottlenecks.

What is needed is a different approach to the I/O bottleneck problem, a new class of solutions that fully leverage the I/O potential of flash SSD without the need to over-provision storage or add more hardware to increase network bandwidth while enabling the CAPEX and OPEX benefits of server virtualization on mission critical applications.