Guaranteed Storage Performance Requires a New Approach to QoS

Derek Leslie is Senior Product Manager at SolidFire.

It is essential for cloud providers or private cloud operators to secure storage performance under the most diverse conditions, including error scenarios, peak loads, variable workloads, or increasing capacity requirements. However, new QoS approaches to storage are necessary.

Whether in public or private clouds, companies require constant performance and system availability, even in the face of increasing amounts of data and increasingly complex workloads. This is only possible with a storage QoS that is an integral component of the system design.

Storage infrastructure that is ready for the future should include at least the following four main components:

Continuous SSD Architecture Ensures Consistent Latency

The basic requirement for successful QoS implementation is to replace hard-drive-based storage systems with all-SSD or all-flash architectures. This is the only way to ensure consistent latency under any I/O load. This is because a spinning drive can only perform one I/O action at a time and every search request creates latency. In cloud environments where several applications or virtual machines share a drive, this can lead to variances in latency from 5 to over 50 milliseconds. However, in an all-SSD architecture, the lack of moving components results in consistent latency, regardless of how many applications require I/Os and whether or not the I/Os are sequential or random. In comparison to the I/O bottleneck of a hard drive, SSDs can perform I/O actions in parallel, for example, with eight or 16 channels, which ensures low latency under any I/O load.

Scale-out Architectures Ensure Linear, Predictable Performance with a Scaled System

Traditional storage architecture follows the scale-up model, where one or two controllers are linked with a set of drive units. Capacity is increased by adding drives. One problem with this is that controller resources can only be upgraded by switching to a "larger" controller. If the "largest possible" controller is used, the only way to upgrade is to purchase an additional storage system. This inevitably leads to higher costs and increased administrative work. A genuine scale-out architecture, however, links controller resources with storage capacity. Any time that the latter is increased and more applications are added, performance is also increased. This performance is available for any volume in the system, not only for new data. This is essential, not only for the administrative planning, but also for the storage system itself in terms of consistent performance.

RAID-less Data Protection Offers Guaranteed Performance in Error Situations

In regards to QoS, a RAID approach seriously affects performance when a drive fails – often by more than 50 percent. This is because an error involves a two- to five-fold increase in I/O load for the remaining storage drives. It's best to use a RAID-less data protection based on a replication algorithm. Therefore, redundant data copies of a single storage medium are evenly distributed over all remaining drives in a cluster – not just on a specific RAID system. The result of this data distribution is that if an error occurs, the I/O load of a failed drive is taken over by the other storage media in the system, i.e., the I/O load on each individual drive increases only slightly.

Balanced Load Distribution Eliminates Peak Loads that Cause Unpredictable I/O Latency

In traditional storage architectures, data is stored in a storage pool within a RAID set. This is usually on the same drive. If new drives are added, they are generally used for new data and not for load rebalancing. The result: Static storage creates unequal load distribution between storage pools, RAID sets, and individual drives. In this scenario, the manual actions of the storage administrator, often via Excel spreadsheets, is the only way to establish an efficient and balanced I/O load distribution and capacity allocation. Other approaches, however, automatically distributes the data across all drives in the cluster. If new drives are added, the data in the system is automatically arranged into several clusters, regardless of whether the data is old or new. Balanced data distribution is therefore possible without any manual intervention. If the system experiences additional workload, it is also distributed evenly. This automatic distribution is the only way to avoid peak loads, which can impair performance.

Only a continuous QoS approach is suitable to ensure predictable storage performance. Modern storage architecture with integrated QoS features comprehensively overcomes not only performance problems, but also offers quick provisioning, not to mention simplified management.

Industry Perspectives is a content channel at Data Center Knowledge highlighting thought leadership in the data center arena. See our guidelines and submission process for information on participating. View previously published Industry Perspectives in our Knowledge Library.