When Microseconds Matter: Delivering Highly Available Inline Deduplication and Consistent Low Latencies at Scale; Interview with Thomas Isakovich, Nimbus Data Systems, Inc. Chief Executive Officer and Founder, Part 2

In this second blog entry from our interview with Nimbus Data CEO and Founder Thomas Isakovich, we discuss microsecond latencies and how the recently announced Gemini X-series scale-out all-flash platform performs against the competition.
DCIG: Could you address what kind of latencies we should expect to see with the Gemini X-series?
Thomas: With the Gemini X latency is going to be around 100 microseconds, whereas the Gemini F can get as low as 50 microseconds. That latency is consistent regardless of the number of Flash Nodes, which is pretty impressive.
DCIG: Is that difference even detectable for the end user?
Thomas: We had a potential client who was looking at purchasing 100 units. Our product delivered, I think, something like 95 microseconds in their tests. The next vendor delivered 150, and then the next one delivered 190. Even though we cost a little bit more, they bought all of it with us, because to them, that app hits so hard that it’s a 2x delta between 95 microseconds and 190 microseconds. One is twice as fast. So certainly not everybody, but a lot of these web-scale guys that run these Oracle databases — they get that much transaction load, they really do care.
Also, with us if you lose a controller there’s no change in performance. The design is such that all the IO — on the regular configuration basis, anyway — is processed by one controller, even though it’s active/active. It’s what we call asymmetric active/active. So, if you lose a controller, then all the IO remains at the same rate; there’s no change in the performance. If you lose a drive with our product, it’s like a 5 or 10 percent hit.
With some other competitive offerings, if they lose a controller, they’ll lose half their performance as well. They truly relied on balancing every IO and LUN ownership between two controllers. So you lose one controller, and you lose half the horsepower.
DCIG: When you talk about the Gemini X’s ability to lose a controller without a significant performance impact, are you talking about losing one of the two controllers in a Flash Node or losing one of the Flash Directors?
Thomas: Either. In either scenario there’s no performance impact.
DCIG: The metadata management happens at the Flash Node level?
Thomas: Yes.
DCIG: And each Flash Node is independent in terms of deduplication?
Thomas: Correct. It’s possible that we could add kind of global cross-node deduplication. I’ll be honest though, at 4K I don’t think that’s really going to do much, because the metadata tables are already going to have so many 4K variations. I can’t imagine there being a slight variation from one or the other very often.
If we ended up doing it globally, then we kind of lose scale-out in a sense because you put the burden for that calculation on the Flash Director. I don’t think we’re going to do it. Given how granular the deduplication is, I think we’re fine. If we’re doing, say, 64K granularity, then you can maybe make an argument that global deduplication would help. But at 4K, I think we’re good.
In the Part 1 of this interview series, Thomas Isakovich guided us through the development of the Nimbus Data Gemini X-series, and where he sees it fitting into the current market.
In the final part of this interview series, we discuss the appeal of Gemini X to enterprise and cloud service providers.