I can’t hide the fact I was looking forward to the Qumulo presentation at Storage Field Day 8: I love scale-out NAS systems, with all the advantages they bring in terms of performance, scalability, manageability and upgradeability. I quickly learned that the founders of Qumulo previously worked on Isilon and OneFS. I work with Isilons in the field, so my interest was peaked
Back to Qumulo: they build a data-aware, scale-out, primary storage system. And it’s software defined. Meaning you’ll have full flexibility in the hardware you want to use and how big/fast/expensive you want to make the system. Plus it also gives you real-time insight into the data on the Qumulo system. Interested? Read on!
Qumulo’s mission is making data visible on an otherwise invisible storage infrastructure. The storage infrastructure itself should never be visible, much like the power infrastructure at home. Infrastructure becomes visible when there’s a problem with it. We’re used to as-good-as 100% uptime of the power grid in The Netherlands: the only time I notice it is when a fuse blows (usually my own fault) or if an Apache attack helicopter cuts down some power lines. It’s the same in IT: users don’t care how the storage system is built, s long as data can be requested correctly and quickly, they’re happy.
Qumulo meets this demand by creating a data-aware, scale-out, software defined, primary NAS system. It includes real-time analytics to give visibility into data usage and performance. Adding more nodes to the cluster allows you to scale capacity and performance when you need it.
Qumulo Core is a software only product, which means you can run it on any commodity hardware or in virtual machines (although not all hardware platforms are qualified). New software is shipped every two weeks, with roughly 50% of their customers updating withing 4 weeks. This rapid development lifecycle is something we’ve seen more often at SFD8: the “one massive upgrade once a year” approach is being abandoned in favor of smaller but more often upgrades.
If you want to buy a pre-installed appliance, Qumulo currently offers two flavors: a 1U QC24 and a 4U QC208 appliance. There’s a 4 node minimum for each grid with no mixing of node types just yet. And it supports the main protocols like NFSv3, SMBv2.1 and REST.
The 1U nodes are created to offer a lower entry point into the market. Four QC24 appliances with several months to a year of subscription are roughly $50k. The largest deployed system back at SFD8 was 20 nodes, 4PB with 10 billion files.
It’s about the files
Qumulo discussed one of the many challenges of NAS systems: storing PBs of data is easy, but storing billions of files is something else.
They revisited the storage physics in a pretty enlightening video which you can find over here. In a nutshell: a spinning disk is actually not that slow compared to an SSD, as long as you make sure it’s being used for large mean transfer sizes (i.e. sequential I/O). There’s a number of problems driving down system and storage performance, including:
- Fragmentation of files across the media
- Small files
- Large file random I/O
- Per file protection
- Metadata scanning
If you can master these problems, it will allow you to build bigger, cheaper, faster and easier systems. One way Qumulo does this is by writing data to disk in 5GB chunks. This squeezes the maximum sequential performance out of the spinning disks.
The data stored in the Qumulo systems comes in through the protocol layer, into the file system and lands in the Pstore or Protection Store. It is then written out to the Bstore, which is an SSD component (1-5% of the storage in a Qumulo system is SSD) linked to a 5GB chunk on spinning HDD.
The above approach helps in creating a fast and affordable scale-out NAS system, but also helps in increasing the availability of the system. You want a low Annual Failure Rate (AFR) for your storage system, but unfortunately this is impacted by a number of things:
- The AFR grows with restripe time. Qumulo promises their systems will restripe a 8TB drive in 20 minute, which is indeed seriously quick!
- The AFR grows when the individual components have higher AFRs (no-brainer).
- If you introduce more components to the system, the AFR also increases.
- And it shrinks once the protection level increases. Currently the Pstores can only do mirroring (=writing to two Bstores), but Qumulo is working on erasure coding with +2 protection and 3x mirroring.
My thoughts on the Qumulo Core product
We were somewhat surprised when Qumulo started explaining storage physics to us. “We know this stuff!” But Qumulo has put a lot of research into the physics of storage. That pays off and allows them to squeeze a bit more of juice out of our trusty old spinning HDDs, which are becoming bigger but not that much faster. Contrary to many “HDDs are dead, long live SSD!” posts on the web, I agree with the mindset that HDD will stay around for several years, maybe even decades: there’s still plenty of development going on in spinning disks. So it makes sense to properly utilize their potential for cheap storage.
Qumulo also spoke about the the challenges of data analytics against billions of files. Check the Qumulo presentations (well worth the watch) and read Alex’s post or this post by Dan.
Disclaimer: GestaltIT paid for the flight, hotel and various other expenses to make it possible for me to attend SFD8. I was however not compensated for my time and there is no requirement to blog or tweet about any of the presentations. Everything I post is of my own accord.