Server virtualization has delivered a significant return on investment (ROI) for those companies that have embraced it, and most data centers have. The problem they are facing is that one of the primary goals of virtualization, increased server resource utilization, is not being met. Most reports show that current utilization levels of server hosts are less than 35%. This is actually good news, since it means that there is an even greater potential ROI available from VMware projects. The challenge is how to best extract that ROI.

The most obvious way to increase server host utilization is to increase the virtual machine (VM) density of each host. Doing so will increase resource utilization on that host and save the expense of buying another server (improving the ROI). However, the problem with loading additional VMs on each server shows itself in two areas, first in storage I/O performance and second, in server RAM limitations. In both cases the intelligent use of flash installed in the server can help.

The Storage I/O Problem

The potential storage I/O problems of a server host with even a dozen virtual machines is well documented as the I/O Blender effect. As more VMs are added the randomization of storage traffic becomes increasingly worse bringing hard-drive based storage systems to their knees. Even when SSD is added to the storage systems downstream, the I/O problem is still not totally resolved. The type of network interface can be an issue. While the overhead of the storage network is only a small fraction of total latency when hard drives are used, it is more significant when accessing fast SSDs.

Another issue can be limitations within the storage system’s controller. Most controllers were designed around hard drive performance. After replacing hard drives with SSDs, users may not achieve the full performance boost they were expecting.

One option is to upgrade to a storage system that’s either designed to be flash optimized or supports all-flash. At the same time the network may need to be upgraded to sustain the storage I/O requests from the connecting hosts and the now-faster storage system. For many companies this may not be the right time in terms of budget for such a storage system and network refresh. Additionally, these expenses may cancel any additional ROI that comes from increasing the utilization of the server host.

The Server-Side SSD Challenge

As a result of the costs and complexity of networking solid state storage, many companies have turned to server-side flash memory to increase VM density without breaking the storage budget. PCIe SSDs have a dedicated controller designed specifically for flash storage. The problem with many PCIe SSD solutions is that they may prevent vMotion. Moving a VM from one physical server to another using vMotion typically requires both servers share the same storage resource over a network.

The second challenge with flash-based storage in general is that it still priced at a premium compared to HDD on a per capacity basis and so it must be used efficiently. One of the best methods to accomplish this is to use at least a portion of it for a cache, so that the most active data is automatically placed into this higher cost resource. Caching relies on the fact that many applications only access a small portion of storage on a frequent basis. Accelerating access to that portion can result in significant performance improvement.

If implemented correctly, the caching software should allow the PCIe SSD to accelerate SAN accessed storage allowing vMotion to operate. To support this, write traffic updates the cache and the shared storage. Read traffic may receive data quickly from the PCIe SSD, if it was determined to be cached. A VM moving to another physical server will operate normally because the shared storage always is up to date. The SSD cache on the new server will begin to provide quick responses as frequently accessed data is identified and retained in the cache. Although only read traffic is accelerated, significant performance gains are achievable.

Solving the PCIe SSD Problem

Caching implementations are typically closely connected to the operating system. This means that a dozen VMs will each need their own caching software and associated PCIe SSD partition, which will add to the complexity of the installation. Companies like LSI are solving this problem with their Nytro XD Application Acceleration Solution by implementing caching in the Hypervisor. This allows all storage traffic from all the VMs to be cached on one or more PCIe SSDs.

Conclusion

Increased virtual machine density is the answer to realizing greater VMware ROI because it eliminates the expense of additional server hosts. But as the ratio of VMs to hosts increases, the resulting storage I/O performance problems can force the IT team to consider SSDs.

Performance may not increase when SSDs are installed in shared storage due to issues with the storage network interface or controller limitations. While typical PCIe SSDs will provide better performance gains, they may not allow vMotion to be supported because data is stored only on that physical server. Caching implementations using PCIe SSDs allow vMotion by still directing write traffic to shared storage. Caching has the additional cost benefit of allowing SSDs to be used on frequently accessed data instead of completely replacing hard drives.

LSI is a client of Storage Switzerland 

Previous Entry: “The Benefits of a Flash Only, SAN-less Virtual Architecture”