One of the reasons that the storage supporting the virtualized environment becomes a roadblock is that it is often the same storage that was designed to support the one-app one-server model of five to fifteen years ago. This storage was designed to be very structured and did not need to adapt to rapidly changing and shifting workloads. Most  servers were segregated to a particular set of disks or at least no two servers accessed the same LUN at the same time. In the virtual environment you want to and even need to have the physical hosts have access to the same storage partitions at the same time. This is to enable features like live machine migration, automated load balancing and high availability. This access pattern of multiple systems accessing the same storage areas at the same time creates an I/O Mix Master that makes performance tuning extremely hard to do.

The Storage Network Roadblock

As we discussed in our article “High Performing Virtualized Workloads Require A New Storage Protocol” some storage protocols try to use sophisticated tagging schemes to be able to assign priority to various workloads. The problem with this technique is it assumes that you know in advance what workloads will need the most performance and when. This obviously is very hard to diagnose. While a few applications can have this label applied to them, they often don’t need high performance all the time, they just need it when they need it. Also, what about workloads that you don’t know need extra performance? This fine tuning to the virtual machine may not meet the flexibility demands of the virtual environment.

Instead of turning sophisticated dials or creating storage tiers on the network that brings a rigidness to a free flowing environment, why not just make the whole storage network fast – one tier for all? The ideal solution is to be able to build a network that can provide native performance such that no additional tuning is needed. As a famous football coach once said, “you can’t teach speed”. This is what fibre channel does, it is fast and robust but unfortunately it is also complex and expensive. The compromise on the cost side has been IP based protocols like iSCSI and NAS (NFS) but they lose the speed and bring complicated fine tuning and storage tiering back into the equation so that performance service levels can be met.

It is situations like these where a new protocol, ATA over Ethernet (AoE), has an advantage when compared to iSCSI and NAS. AoE, similar to FC, does not have the TCP conversion overhead of iSCSI and NAS but it provides the performance on standard Ethernet cable infrastructures. In other words costs are kept down, performance is kept up and fine tuning complexity is kept out. One tier for all.

The Storage Controller Roadblock

As we will cover in more detail in an upcoming article, storage also has become a virtualization roadblock because as more virtual machines are created more server instances are accessing the storage at the same time and the storage engines, known as controllers, can’t scale to meet the demand. In the past a shared storage system would typically have only a few dozen hosts attached to it and those hosts were seldom busy 100% of the time. In the virtualized server environment those few dozen physical hosts may have at least 10 virtual servers, meaning a potential 360 instances all accessing the same storage at the same time. As a result the storage controller can become a roadblock even before the storage system has reached its physical capacity.

Most storage systems are limited in their ability to scale because only two storage controllers were available. While these controllers are high end (and expensive) processing engines they were not designed for the density of workloads described above. As more workloads are added there are limited ways to add additional processing power. While it’s true that another storage system could be added the expense of these storage processors made it very costly to add those additional units. Again the virtualization project stalled because it was out of performance and in many cases out of budget. Not only did the additional workloads per storage system cause performance problems but these additional workloads require additional capacity. In a traditional architecture, as more disks are added performance declines. The combination meant that running out of storage processor performance could become a reality for data centers of all sizes.

What is needed is a scalable architecture where the individual cost of a storage system was not expensive. This new scaling model needs to be a more granular approach to scaling but not too granular as it is with some “scale-out” storage systems. Storage systems need to strike the right balance of performance and capacity. In short you want to use cost correct storage processors to drive high performance but you want to efficiently use all of that processor and not leave the processor idle as more storage systems are added.

AoE based systems like those from Coraid can help address the storage roadblocks. We invite you to view the articles related to this subject as well as attend our webinar “Stopping The Storage Roadblock To Server Virtualization” in which I and Robert Przyhucki, Director of Product Management from Coraid, will identify the issues that are causing the storage roadblock and provide key techniques for working around them.