A good friend of mine is friends with someone who works at Panasas and suggested I might like to hear from them. I had the opportunity to speak to some of the team, and I thought I’d write a brief overview of what they do. Hopefully I’ll have the opportunity to cover them in the future as I think they’re doing some pretty neat stuff.
It’s HPC, But Not As You Know It
I don’t often like to include that slide where the vendor compares themselves to other players in the market. In this case, though, I thought Panasas’s positioning of themselves as “commercial” HPC versus the traditional HPC storage (and versus enterprise scale-out NAS) is an interesting one. We talked through this a little, and my impression is that they’re starting to deal more and more with the non-traditional HPC-like use cases, such as media and entertainment, oil and gas, genomics folks, and so forth. A number of these workloads fall outside HPC, in the sense that traditional HPC has lived almost exclusively in government and the academic sphere. The roots are clearly in HPC, but there are “enterprise” elements creeping in, such as ease of use (at scale) and improved management functionality.
[image courtesy of Panasas]
It’s Really Parallel
The really value in Panasas’s offering is the parallel access to the storage. The more nodes you add, the more performance improves. In a serial system, a client can access data via one node in the cluster, regardless of the number of nodes available. In a parallel system, such as this one, a client accesses data that is spread across multiple nodes.
What About The Hardware?
The current offering from Panasas is called ActiveStor. The platform is comprised of PanFS running on Director Blades and Storage Blades. Here’s a picture of the Director Blades (ASD-100) and the Storage Blades (ASH-100). The Director has been transitioned to a 2U4N form factor (it used to be sit in the blade chassis).
[image courtesy of Panasas]
Director Nodes are the Control Plane of PanFS, and handle:
- Metadata processing: directories, file names, access control checks, timestamps, etc.
- Uses a transaction log to ensure atomicity and durability of structural changes
- Coordination of client system actions to ensure single-system view and data-cache-coherence
- “Realm” membership (Panasas’s name for the storage cluster), realm self-repair, etc.
- Realm maintenance: file reconstruction, automatic capacity balancing, scrubbing, etc.
Storage Nodes are the Data Plane of PanFS, and deal with:
- Storage of bulk user data for the realm, accessed in parallel by client systems
- Also stores, but does not operate on, all the metadata of the system for the Director Nodes
- API based upon the T10 SCSI “Object-Based Storage Device” that Panasas helped define
Storage nodes offer a variety of HDD (4TB, 6TB, 8TB, 10TB, or 12TB) and SSD capacities (480GB, 960GB, 1.9TB) depending on the type of workload you’re dealing with. The SSD is used for metadata and files smaller than 60KB. Everything else is stored on the larger drives.
DirectFlow is a big part fo what differentiates Panasas from your average scale-out NAS offering. It does some stuff that’s pretty cool, including:
- Support for parallel delivery of data to / from Storage Nodes
- Support for fully POSIX-compliant semantics, unlike NFS and SMB
- Support for strong data cache-coherency across client systems
It’s a proprietary protocol between clients and ActiveStor components, and there’s an installable kernel module for each client system (Linux and macOS). They tell me that pNFS is based upon DirectFlow, and they had a hand in defining pNFS.
Scale out NAS is exciting but us enterprise types want to know about resilience. It’s all fun and games until someone fat fingers a file, or a disk dies. Well, Panasas, as it happens, have a little heritage when it comes to disk resilience. They use a N + 2 RAID 6 (10 wide + P & Q). You could have more disks working for you, but this number seems to work best for Panasas customers. In terms of realms, there are 3, 5 or 7 “rep sets” per realm. There’s also a “realm president”, and every Director has a backup director. There’s also:
- Per-file erasure coding of striped files allows the whole cluster to help rebuild a file after a failure;
- Only need to rebuild data protection on specific files instead of entire drives(s); and
- The percentage of files in the cluster affected by any given failure approaches zero at scale.
Thoughts and Further Reading
I’m the first to admit that my storage experience to date has been firmly rooted in the enterprise space. But, much like my fascination with infrastructure associated wth media and entertainment, I fancy myself as an HPC-friendly storage guy. This is for no other reason than I think HPC workloads are pretty cool, and they tend to scale beyond what I normally see in the enterprise space (keeping in mind that I work in a smallish market). You say genomics to someone, or AI, and they’re enthusiastic about the outcomes. You say SQL 2012 to someone and they’re not as interested.
Panasas are positioning themselves as being suitable, primarily, for commercial HPC storage requirements. They have a strong heritage with traditional HPC workloads, and they seem to have a few customers using their systems for more traditional, enterprise-like NAS deployments as well. This convergence of commercial HPC, traditional and enterprise NAS requirements has presented some interesting challenges, but it seems like Panasas have addressed those in the latest iteration of its hardware. Dealing with stonking great big amounts of data at scale is a challenge for plenty of storage vendors, but Panasas have demonstrated an ability adapt to the evolving demands of their core market. I’m looking forward to seeing the next incarnation of their platform, and how they incorporate technologies such as InfiniBand into their offering.
There’s a good white paper available on the Panasas architecture that you can access here (registration required). El Reg also has some decent coverage of the current hardware offering here.