Caching with solid state
The other basic implementation is to use solid-state storage as a cache in front of spinning disks. This method has the advantage of always accelerating the hot data in real-time, since only the hot data is likely to be in cache. And because the solid-state storage is acting as a cache, there's no need for an administrator to decide what data should be placed on it. The basic questions here are what size cache is appropriate and which workloads should be directed toward the cache to make the best use of the solid-state device.
Some solid-state caching solutions are built into existing storage systems, while others are delivered as external appliances. Adding flash memory as a cache inside a storage subsystem in effect provides a "level 2" cache not unlike the L2 cache found on many processors inside today's computers. This added cache capacity improves performance for most if not all operations. In addition, because flash memory is non-volatile, this cache provides some extra protection in the event of power loss. But issues such as cache coherency, and whether the cache is DRAM based or flash memory based, remain. Generally, a cache is tied to one processor or controller, and there are various cache management functions that can be applied to allow caches to work properly with multiple processors or controllers. In addition, storage systems that use caching can add special features to their internal OSes that are aware of the cache and can provide additional flexibility, such as the ability to assign different I/O priorities for I/O going to different volumes on the storage system.
The caching appliances add the benefits of cache without requiring changes to any existing servers or storage systems. These appliances fit easily into the storage network and can accelerate all I/O going through them, even sending data to different storage subsystems at the same time. Many of the appliances can be set to write-back, write-through or pass-through for any given volume they accelerate. Some of the caching appliances are constructed in such a way as to allow their memory modules to be hot-plugged, so maintenance or growth can occur without taking down the entire appliance.
The big question for a caching implementation is how much cache is enough. For many workloads and applications, a relatively small amount of cache (5% to 20%) relative to the total storage allocated to that application is enough to provide significant performance improvements. For other workloads, the cache needs to be large enough to hold the entire volume to achieve appreciable performance gains.
It's all about performance
Solid-state storage, however it's deployed, offers the promise of significant performance gains. We've seen results of seven to nine times overall performance gains in our lab testing for various real-world applications (email, database, etc.) when configured optimally for the application.
With performance gains of that magnitude possible, what's not to like? Certainly, pricing is a factor. However, consider some of the current methods that are used to increase performance for spinning disk drives, such as "short stroking" spinning disk drives. Short stroking spreads data over many disk drives by using only a portion of the capacity of each drive for data, so that as many "spindles" as possible can be applied to improve performance. To achieve desired performance goals, some users short stroke some of their enterprise disk drives using ratios of 7:1, 8:1 or 9:1, which means they're using only 1/7th, 1/8th or 1/9th of the available capacity on each drive. If the price of an enterprise SSD is 10 to 15 times the price of the spinning drives being short stroked, it may make sense to move that application data to enterprise SSDs and get the required performance while using much less power and space.
Almost all data storage system vendors now offer configurations that use a combination of solid-state storage and enterprise SATA storage instead of arrays full of enterprise spinning disk drives. These new configurations typically offer higher performance, equivalent capacity, lower power consumption, smaller space requirements and lower total hardware costs.
BIO: Dennis Martin has been working in the IT industry since 1980, and is the founder and president of a computer industry analyst organization and testing lab.
This was first published in June 2010