About a year ago, I wrote about the ever-decreasing role of spinning disks in highly transactional OLTP storage workloads. Back then, Texas Memory Systems had just released test results for its RamSan-630 SSD storage product that achieved more than 400,000 SPC-1 I/Os per second (IOPS) at a cost of $1.05 per I/O -- a truly remarkable performance result that was (and is) impossible to achieve with traditional disks at the same price.
Today, the Storage Performance Council (SPC) published new test results for Kaminario's K2 SSD storage product. Its record-shattering performance is in excess of 1.2 million SPC-1 IOPS for just 40 cents per I/O. That Kaminario can produce these results at such an economy is a huge testament to the advances made in SSD technology and further proof that solid state rules the roost when it comes to OLTP workloads.
[ Also on InfoWorld: See how far solid-state disks have come since Matt Prigge proclaimed them the reigning champs for storage. | Sign up for InfoWorld's Data Explosion newsletter for news and updates on how to deal with growing volumes of data in the enterprise. ]
However, that's not all it tells us. Although the full details aren't yet available, it appears that Kaminario was able to more than triple Texas Memory Systems' performance not by using an entirely new type of SSD but by using a different architecture with current SSDs. Unlike Texas Memory's single-appliance approach or the scale-up architecture used by many traditional storage vendors, new entrants to the storage space such as Kaminario tend to develop scale-out architectures. That key difference is likely why Kaminario now holds, at least temporarily, the transaction-performance crown.
Where scale-up falls short
Where scale-out works better -- and where it doesn't
Scale-out architectures avoid some of those pitfalls by achieving scalability through the addition of many much smaller and much cheaper chunks of storage -- typically attached to industry-standard server hardware. This approach removes much of the forecasting uncertainty for the initial purchase; you often need to consider just your near-term storage requirements. As your capacity and performance outlook dictates, you can add storage resources on a pay-as-you-go basis.
Also, the ability to use largely off-the-shelf hardware components substantially decreases the hardware and development costs for the storage vendor, in turn making the storage products less expensive.
However, scale-out architectures have their downsides. Because performance and storage resources are typically linked to one another (after all, a single storage building block includes both), the ratio of performance to capacity can get out of line with an organization's requirements, resulting in either wasted storage or wasted capacity. Many scale-out vendors try to overcome that issue by offering various models with different performance-to-capacity ratios, some using SSDs, some using SATA hard disks, and some using near-line server-attached storage. They also tier data along the storage media types as workloads dictate.
Another substantial challenge that faces scale-out vendors is how to efficiently get performance from many individual bricks of controller resources while still providing rock-solid reliability. This involves solving many of the same challenges that must be overcome in high-performance computing environments: safely and effectively parallelizing a complex and demanding workload among many compute and storage nodes while making the result appear as though it's a single, unified architecture. That challenge places much more emphasis on the quality and innovation that goes into the software -- which, in large part, is what you're paying for.