Nimbus Data today is releasing a new software suite for their flash-based NAS/SAN boxes. Luckily for their current customers, the updated suite is free.
The new software is called HALO (High Availability Low Overhead) 2013 and it brings three major things to the table, a database driven metrics analysis system, an API for writing your own tools, and a mobile monitoring suite. If you are not familiar with the Nimbus products, the company makes one basic product, a 2U box stuffed full of SSDs with very fast interfaces on the back. This is the non-technical way of saying a SAN, NAS, or all-of-the-above device, plus the attendant software that we are focusing on today. The number of boxes, the exact feature sets, and a few other differences are what differentiate the various products, but at the most basic level, they are all flash storage modules based around a 2U building block.
Nimbus claims to have a few features that make them stand out from all the other SAN vendors out there, and at least on paper it looks good. The company doesn’t have a legacy magnetic drive market to protect so they promote flash for just about every market. They claim that flash is the best solution for almost all heavy storage workloads, something that most won’t have a hard time believing. SSDs are undoubtedly the best from a performance standpoint, but some enterprises still have lingering doubts about lifespans and TCO.
The lifespan question is the easiest part to address, take the smallest box in the Gemini line which has 6TB as a starting point. They use “enterprise” grade flash meaning a top bin that will not only meet but likely exceed the promised write lifetime by a good margin. It also has 28% of the space reserved for write endurance. If they do even the most basic wear leveling at the system level, it would be pretty safe to assume they do that and more, you can get a good idea of how long it will live.
Assuming that the box has 24/7 full speed writes at the maximum claimed rate of 12GBps from the 56Gbps from the FDR Infiniband ports, you can make some baseline assumptions. To fill the entire box without taking overprovisioning into account would take 6000/12 or 500 seconds, lets just call it eight minutes. That means 180 full writes a day. At 100K cycles of write endurance per flash cell, that would be about a year and a half, 555.56 days to be exact of solid writes before you hit 100K cycles.
While possible, filling up an expensive flash array with data then immediately writing over it is not a good use of expensive hardware. A normal enterprise would more likely save a bit of cash by not buying an array like this and instead modifying their code to store the data in /dev/null instead. If you take a more sane number like 20% writes and 80% reads, that would mean about 7.5 years before you hit the write lifetime of the smallest box. That 28% overprovisioning and any clever wear leveling will only add to this number, likely at least doubling it. The biggest box Nimbus makes stuffs 48TB into that same 2U, so those 7.5 years become 60. If you are still using your storage box after a few decades, you should probably replace that 2U unit with something that makes your wristwatch look bulky.
The other objection is cost, be it measured by dollars per TB, TCO, or any other relevant metric. Here Nimbus has a site up for cost comparisons, but this is an enterprise solution so MSRP has about as much to do with reality as a Congressional speech. Lets just say that Nimbus makes a decent case for flash having a lower TCO than magnetic storage solutions, but definitely run your own numbers before you make an impulse purchase in this sector.
There are three ways that Nimbus does things differently that makes us take their numbers seriously, software, features, and power. As we said earlier, the software is included with the box, it is all ‘free’ as are updates. If you have a Nimbus box, there aren’t the usual list of feature ‘gotchas’ that normally bloat the cost in this arena. Software add ons are also non-existent, you get everything possible with the box.
Same with the interfaces, the basic box ships with 4 ports of FDR Infiniband (56Gbps) per controller standard, and the controller is multi-protocol capable. If you want 16Gbps FC or 40GbE you just need an adapter, no cards to purchase. Nimbus says these cost <$100 per, and a full rack of Gemini servers can have up to 8 FDR ports every 2U, so $800 max if you want to go slower than stock. This is a pittance for enterprise storage devices, barely rounding error.
The last claim is low energy use, 80% less than an array of 15K drives. How they get there is a bit more interesting than most claims. Nimbus built their own controller from scratch and it does only what it needs to do. They claim that it is not only faster than generic hardware but more focused too, so it burns less power. In most other arenas like video decode, a dedicated hardware block will cream a general purpose CPU for performance per watt, and we have no doubt this could very well be the case in storage too. A quick check of the specs say the boxes draw 2A@120/1A@240V, a 2.5″ 15K RPM SAS drive will pull about 4W, add in what you feel for an enterprise enclosure and interfaces. A cursory check of the numbers seems to show that Nimbus’ power claims are in the right ballpark. At the very least this controller saves you from having to buy a closet full of storage interface boards at painful prices, not to mentioning powering them.
One other thing that Nimbus claims is that their controller allows them to have no single point of failure in their higher end E-Class and Gemini boxes. Many SANs have two interfaces, redundant drive/raid chips, and PSUs, but only one controller. Two controllers are essentially all you need for full redundancy, a good start for the failure averse enterprise crowd. If it needs to be said, the controllers also support RAID 5/6/10 and likely some odder combinations thereof, but no credible device in this space doesn’t. No cracks about Microsoft storage server from the peanut gallery, remember that we said credible.
With all that in mind we take the Nimbus claims very seriously, they have all the pieces to do the job right, and they claim 330+ customers using their servers now. Given the cost and targeted market segment, you can be pretty sure that these customers did run the numbers in detail, so they likely hold up under real scrutiny.
If you recall that in first part of this article we mentioned something about the HALO (High Availability Low Overhead) software suite being upgraded. The new version called HALO 2013 has a metrics database at its heart for automation tasks. Everything is stored in open formats and plain text accessible so if you don’t like their tools, you can pull the data into whatever you want however you want. The second part of the upgrade is a full RESTful API that exposes all of the functions that the Nimbus tools themselves use so you can write extensions, plug-ins or whatever you choose. The most likely use for this is simply to interface with existing enterprise dashboards but how far you take it is up to you.
The last bit is something that is becoming more and more common, a mobile management suite. The HALO Mobile suite is available for Android and iOS allowing you to monitor your flash on the go in a flash. Sorry, couldn’t resist that pun. For now, HALO Mobile is read only so you can look but not touch. One would guess that at least some portions of actual management will be added in the near future, but if not, remember that API? There is no reason that given a RESTful API, you couldn’t roll your own mobile app with very little effort.
And that is what Nimbus is adding to the HALO suite, a database, API, and mobile monitoring app. It may not sound like much, but since it is free most users have no reason to avoid the upgrade. In the enterprise world, not soaking customers for a feature filled upgrade is a rare thing indeed. That datapoint goes a long way to minimizing doubts about the rest of the TCO numbers Nimbus put out. Could it be that flash finally has a lower TCO than magnetic storage?S|A
Editor’s note: Article updated to reflect the software suite name is HALO 2013, not HALO 13. March 25, 2013 10:45pm.
Editor’s note 2: Article updated to add Gemini to redundant controller list. March 31, 2013 12:15pm.
Latest posts by Charlie Demerjian (see all)
- SemiAccurate has Skylake-SP die shots, sizes, and more - Jul 21, 2017
- Skylake-SP has a diverged core - Jul 19, 2017
- Intel’s Purley platform architecture is a step forward - Jul 13, 2017
- Intel launches Purley aka metal Xeons - Jul 11, 2017
- AMD’s Epyc has lots of connections - Jun 26, 2017