If you work in any aspect of the enterprise storage market, it should come as no surprise when you hear about how much faster solid state drives (SSDs) are compared to even the highest-performance enterprise hard disk drive (HDD) arrays, or about how SSDs provide much quicker application response times. Depending on what you hear, SSDs, and particularly PCIe SSDs, are considered to be somewhere between a targeted solution for relieving specific bottlenecks and a storage panacea that simply makes everything better. Some of that is true, some of it’s hype—but what’s the reality?
We wanted to know. We measured a baseline workload using a MongoDB node equipped with a 20-drive enterprise disk array and then measured the same data in an identical node, but this one equipped with a single Micron P420m PCIe SSD. The results were surprising—even to us!
Useful Work: One P420m Does 4X to 5X MoreHow much more can you do with the Micron P420m PCIe SSD? Performance in terms of operations per second is a good metric to measure useful work (work done in a given amount of time). The more operations per second we process, the more useful work we can do. Our tests measuring the same workloads show the Micron P420m improves operations per second by 4X to 5X over the HDD array, enabling 4X to 5X more useful work.
Responsiveness: One P420m SSD Is 27X to 138X BetterThe performance difference was amazing, although not that surprising. (We expect very high performance from our P420m because we know it’s quite fast.) What was surprising was what we saw in application responsiveness (average latency). With the same workloads, improvements ranged from 27X to 138X with the P420m.
Less Power: One P420m Draws Less Than 1/6th the Power With the P420m, performance is better and latency is far better—but what about power? Surely with all that NAND flash onboard, the P420m would show a significant power draw. We got a surprise there too. Assuming each HDD draws about 8 watts (which translates into a total of 160 watts for the HDD array), the P420m reduces per-node storage power consumption by more than one-sixth. (The P420m draws 25 watts when fully active.)
Smaller: One P420m Node Needs Half the SpaceLooking at the space needed to house storage, it takes only 1U to house a single P420m-based node (many 1U platforms can support two P420m SSDs, but 1U is the minimum needed to support one P420m), while it takes 2U to house 20 enterprise HDDs (2U can typically support 24 HDDs).
Designing the ClustersWhen designing a new MongoDB cluster, several approaches can be taken. For example, you can fix a workload (set a target number of operations of a given type to be completed in a given time) and then determine what sort of node and cluster scale are required to meet those needs. Or, you can look at fixed space constraints and estimate how much useful work can be accomplished by the resultant cluster.
The tests we ran show you can do both. If you use the P420m instead of an enterprise HDD array, you can build nodes that increase performance 4X to 5X, reduce latency 27X to 138X, use one-sixth the power, and occupy only half the space of the HDD array.
So What?SSDs have been mainstream in enterprise storage for some time. We’ve moved from asking if we’re going to deploy them to asking where are we going to deploy them. Are we going to use them as a cache/buffer, or should we finally make the move from specialty deployments to primary data store in our line of business applications, like MongoDB?
If we want to do 4X to 5X more useful work with each node, if we want the application doing that work to be 27X to 138X more responsive, and if we want to do all this with less power in less space, the answer is clear: We should use enterprise SSDs like the P420m where they bring the utmost value—as the main data store in our core line of business applications.