Do you like sequels? When you see or read something really great and you’re left thinking how you wish it hadn’t ended! You wonder, "Is that all there is?” And if you’ll get a change to relive — or better yet, extend — the memories? I know I do.
Then…I hear that a sequel is coming, and I’m thinking, "Nice! How cool is that?!" The best parts of the first release start to replay in my brain and excitement builds. Smell the popcorn, feel the dimming lights… Then, a double-take — will the sequel be as good? Could it be as good? What if it isn’t?
I sometimes get the same feeling in tech. No, really!
(Psssst! Do you want to jump to the last chapter and read the results? Grab the NVMe™ SSDs Future-Proof Apache Cassandra® paper from micron.com).
The First Release
In November of last year, Micron published some pretty amazing NoSQL database results using our newly-minted Enterprise SATA SSDs. With standard YCSB benchmarks, we showed how SSDs can improve multiple NoSQL workloads and use cases. In this release, we compared our 5210 ION SATA SSD to state-of-the-art 10,000 RPM hybrid hard drives in two four-node NoSQL clusters with four physical drives per node). We found exceptionally higher performance (operations per second) and much, much lower latency with SSDs. Great first release!
So -- using SSDs with NoSQL seems like an obvious fit, right? Well, yes – for performance focused NoSQL, SSDs were the clear winners (download all the results and test details).
Then, we started thinking about a sequel. What else could we do with NoSQL? What if instead of an HDD-replacement SSD we went 180-degrees the other way? What if we used our NVMe performance SSDs in a similar test? What could we get from Micron 9300 SSDs and NoSQL using the same standard benchmarks?
Researching the (Potential) Sequel Script
When I was doing some background before writing this – some sequel script research -- I ran across a really good piece by Bob Violino that’s posted to InfoWorld’s website. Bob notes one of the main reasons why NoSQL databases are so relevant and popular today: very large data sets and our reliance on the insights those large dataset hold. He writes:
“…NoSQL databases can be scaled across thousands of servers, though sometimes with loss of data consistency. But what makes NoSQL databases especially relevant today is that they are particularly well suited for working with large sets of distributed data, which makes them a good choice for big data and analytics projects…”
Scale. Distributed workload. Big Data and Analytics projects. The definitions of ‘demanding’ for a database.
That got us wondering – since we’d released our latest NVMe SSD, the Micron 9300, what might that bring to NoSQL? We knew that SATA SSDs had huge benefits compared to how we used to do NoSQL (hard drives), but would next-generation, performance-focused NVMe SSDs make much of a difference?
Intuitively? Sure. But we wanted to know. Really know. What’s the next logical step for NoSQL builds when you need the utmost performance?
Writing the Sequel Script
Let’s set the stage: We know that SSD capacity, throughput (GB/s), and IOPS vary by SSD, I’m focused on the Micron 9300 3.84 TB U.2 NVMe SSDs and leading, enterprise-class 3.84 TB SATA SSDs. Other SSD models, capacities, configurations and/or interfaces may give different results.
The main characters:
For any comparison to be fair, we have to start all tests from the same place. This is crucial for SSD testing (download this whitepaper for more background). All SSDs were restored to fresh out of box (FoB) state and preconditioned prior to measurement. This ‘testing good hygiene’ helps ensure that all test results reflect only that test run (and not prior test runs).
We set up four-node test clusters and applied YCSB workloads A-D and F and ensured that the data set was ‘spread’ across all four test nodes. Why not workload E? Only because that workload isn’t universally supported.
The Sequel is a Hit!
Not only was the sequel just as good as the first release, it was better. Here’s how much better when we compared it to Enterprise SATA SSDs:
To say the least, we LOVED the sequel! It was amazing.
Here’s an example of the Workload A results (reflects session recording). The Micron 9300 SSD results are in blue, the enterprise SATA SSD results in dark grey. At the far left is a lightly loaded test run (64 workers) and at the right a more heavily loaded test (1024 workers).
What stands out? First, the SATA tests stop at 512 workers. Why? Well…that would be giving away the sub-plot of our sequel, wouldn’t it? (grab the complete paper– all the details and background are spelled out).
Second, notice the latency lines? With latency, lower is better. The details are pretty clear (not like credits scrolling by on a screen) – the Micron 9300 SSD configuration is lower at any common worker count and increases far more slowly as the worker count increases.
More workers (heavier loading) places immense demands on NoSQL clusters and the Micron 9300 NVMe SSD excels with that extra loading. If you’re looking to build a cluster you can grow into, look at the Micron 9300 NVMe SSD.
Here’s why: When NoSQL demands drive us toward higher performance – higher than Enterprise SATA SSDs can deliver, toward forward-thinking designs, the answer is clear: the Micron 9300 PRO SSDs deliver NoSQL performance and capacity that’s lightning-fast and easily extensible. It helps us build NoSQL cluster nodes scale well when the data set does not fit into memory.
When that happens, you can realize ‘next level’ results with Micron 9300 PRO NVMe SSDs, the clear choice for faster and more consistent NoSQL database workloads.
And it isn’t just with session recording either. We saw major benefits with all the workloads we tested – more users, more performance and lower latency.