|By Tom Leyden||
|November 8, 2012 09:00 AM EST||
It’s probably a good idea to state I wrote this blog while employed by Amplidata, but during my own time. This article reflects my own opinion, not necessarily that of Amplidata or its partners.
As I am writing this, I am crossing the Atlantic for the seventh time in about two months. I’m on my way to CloudExpo West in Santa Clara, one of the few technology trade shows that are still growing. At the event I will be sitting on the last Object Storage for Big Data panel of the season. Robin Harris – aka StorageMojo – and I have been working hard this fall educating the industry on the benefits, challenges and opportunities of Object Storage. We’ve been trying to explain how the current generation of Object Storage platforms is so much different from the first attempt at it (EMC’s Centera), how it enables companies cope with the massive amounts of unstructured data that we are all generating and how companies can even monetize archived data by re-activating their archives.
Unlike StorageMojo and some other people who I have been working with lately, I don’t have decades of experience in the storage industry. However, being located in Belgium, I’ve had the privilege of working with people who used to be part of the Filepool team (and spent years at EMC after the acquisition). Those were the earliest object storage days, I had no idea of what was coming. Later, at Sun, I learned a lot about Object Storage when we were working on the Sun Cloud project. The architecture (ZFS) was different of what we are seeing on the market today, but the concept was – as was often the case at Sun – promising. This article is not another take at describing Object Storage and the benefits it brings, it’s more an overview of what we have learned at the past four Object Storage for Big Data panels. The setup for each of the panels was mostly the same: Robin Harris would challenge between 4 and 6 Object Storage specialists (technology vendors or users) and try to have the audience participate with. We did expect the topics of the panels to be different as we were hosted by trade shows with different audiences, but we never expected the discussions to vary as much as they did.
The common thread for each panel was the challenge companies have to store different types of Big Data and more particularly Big Unstructured Data. The latter represents up to 90% of the digital data that we will be generating over the next decades and will put traditional storage technologies under heavy stress as they are hitting their scalability limits. Unstructured data is currently mostly stored in file system based storage infrastructures. File systems will not only be unable to scale as required – try setting up a file structure for 5 petabytes of data – but they will also become obsolete as applications can provide a lot more features to keep your unstructured data organized (structured?), to analyze that information and potentially monetize what is today stored in (dead) tape archives. Rich applications that talk directly to a large and (infinitely) scalable storage pool make a lot more sense than maintenance-intensive files systems. Also, properly designed Object Storage (with erasure coding technology instead of RAID to protect the data) requires a lot less overhead, consumes a lot less power, can easily be implemented over multiple sites and does not require migration to new systems when a system cannot be further scaled. So what else did we discuss at the panels?
The first panel after summer was at Intel’s IDF in San Francisco. Panel members came from Intel and Quanta, who with Amplidata built an Object Storage reference architecture. We also had Michelle Munson of Aspera, who presented a couple of perfect use cases of Object Storage in the media and entertainment industry. Aspera developed a very smart way to transfer large amounts of data over the WAN in a much more efficient way than how it is currently done. Aspera’s bandwidth optimization software practically enables this new generation of Object Storage by taking away the latency issue, e.g. to stream high res movies over a long distance. Once we had explained the drivers for Object Storage, the opportunities and best practices, most of the discussion (questions from the audience) was about why RAID is not the right technology to architect an Object Storage platform with. We discussed the benefits of erasure coding in much detail and spent a lot of time on the differences with RAID. In short: in Erasure Coding based systems, all disks are equal (all parity) and there is no need to rebuild a disk when broken: when codes are lost due to bit errors or hardware failures, new codes can be generated spread over the whole pool, not just one system. A recent and very good independent deepdive in the Amplidata erasure coding technology can be found here.
A lot less RAID and erasure coding at the Createasphere DAM Show in New York a few weeks later. The show focusses on Digital Asset Management and the attendees are more interested in the applications and content than the actual data. That did not make the discussion any less interesting. From Sarah Berndt of Johnson Space Center we learned a *lot* about the importance of metadata, an issue that would be discussed at SNW Europe as well (see further). Interesting newcomer on the panel was Dalet, a DAM vendor who integrate with many Object Storage platforms and see a clear benefit of having their platform interface with a scale-out storage pool directly (REST) rather than through an additional file system. Dalet is the perfect valet in my car analogy that is becoming more and more popular: a file system is like a public parking lot where you have to go find your car yourself (this once took me a few hours in Paris’ CDG airport). Object storage is much more like valet parking, where you get a ticket when you leave your car and use that ticket to get it back later. The application, Dalet, is the valet.
At SNWUSA in Santa Clara in October we had David Chapa of Quantum on board for the firs time. David is an authority to explain the use cases where tape is the better alternative and when it is better to use Object Storage, or Wide Area Storage (WAS) as Quantum calls it. WAS is Quantum’s attempt to take away the confusion caused by the name Object Storage, a term first used by EMC almost a decade ago. I think it’s a good idea of Quantum to try to introduce a new term, I’m not sure WAS is the best choice though. Maybe something new will come up next month at Greg Duplessie’s Object Storage summit, although I doubt it. Once we kind of agreed that this generation of Object Storage, or whatever it will be called later, has very little or nothing to do with EMC’s product line that was most famous for locking-in customers, the conversation took a very sudden change. In an attempt to spice up the discussion, Ranajit Nevatia of Panzura claimed Object Storage provides very bad performance. This was very much true for the first generation of Object Storage platforms we just discussed and might be true of the platforms they currently promote (including Atmos, EMC’s second attempt at Object Storage), but not at all for the technologies that are most successful on the market today. Scality have been promoting their high IOPS (smaller files, IO intensive workloads). Amplidata focus more on large file storage, which is IMO the more obviouse use case for Object Storage, but I may be biassed. In a recent independent test, Amplidata demonstrated throughout numbers that can only be called “extremely high-performant”. Howard Marks confirmed Amplidata provides 1 GB/s of throughput with a single controller. But it gets better: Amplidatas scale throughput linearly by adding more controllers. So a system with 6 controllers provides 6 GB/s of throughput.
Last week’s panel at SNW Europe, which is traditionally well attended by press and analysts, was again very interactive. Robin Harris set the stage explaining how this generation of Object Storage is different from earlier products. This led to a lengthy discussion about API’s, a call for one standard API (I say let’s just all standardize on Amazon) and complaints about lock-ins by … yes, EMC. Vendors be warned, that trick is getting old and is not getting any respect. The audience included some of the better analysts and bloggers, including the451′s Simon Robinson and Storagebod. The latter, known for being a critic of the Object Storage paradigm (with great arguments), helped us bring the discussion to the next level by bringing up interesting topics such as the importance of metadata for the applications: who/what will enter metadata? The application? People? The panel acknowledged that, while applications already generate quite some metadata, companies will have to make business decisions on how much metadata they need. Adding more metadata comes at a cost as it will require manual work. The day after the panel, it was interesting to see Chris Mellor be critical of Object Storage in his review of the show (how dare the Object Storage vendors doubt the many benefits of tape?). Chris, join us on the panel next time!