For any company in the high performance computing storage market, sixteen years is quite a run. Still private and still growing past its early years as a venture-backed, scale-out NAS company, Sunnyvale-based Panasas has managed to carve out a niche at the highest end of computing—and has turned an eye to the broader world of high performance commercial markets.
With approximately 150 employees, Panasas is still led by the company’s founder and chief scientist, Garth Gibson, who raised the small company through the HPC storage ranks in the early 2000s to its stable position today. The basis for the Panasas ActiveStor appliance that packages its NAS under the PanFS file system umbrella is rooted in early research work at Carnegie Mellon University, where Gibson and team worked on refining RAID, which was later spun out to meet the needs of academic and government HPC sites. According to Geoffrey Noer, VP of product management at Panasas, its list of customers is growing and well beyond the confines of research HPC.
In 2010, Noer says that around 60 percent of the company’s revenue came from its traditional business in academia and government. However, over the last five years, that number has flipped, with 70 percent of Panasas revenue coming the commercial market. To be fair, this is not general enterprise—these are what one might call commercial HPC segments, including oil and gas, life sciences, manufacturing, and other areas, but these have been hot growth areas, beginning with the seismic simulation push that began during their commercial turnaround and leading up to today, where the primary focus is still on capturing life sciences as well as media and entertainment.
In both of these target markets, however, Panasas has an uphill climb, not only against its rivals on the file system front (Lustre and GPFS), but more from the EMC Isilon systems that excel in particular in both media and life sciences.
Noer agrees that there are shakeups coming for vendors in its target markets. While are some are positive for companies that have deeper roots in the traditional HPC space, including the fact that high performance computing is no longer so isolated from larger enterprise spheres, others mean they will need to continue investing in research and development to keep an edge. “There has been a continued move toward scale-out storage architectures and parallel file systems because it’s increasingly critical to get levels of performance that are in ever more demand. But from a protocol and file system standpoint, it means a move away from NFS, even though it’s still widely used.”
That ecosystem is filtering down to three main file systems, says Noer, although there are big differences in what kind of market share these file systems actually capture (not to mention rather widely different numbers on the analyst and vendor sides). To put this in context, one analyst group, Intersect360, finds in its survey of commercial and academic HPC sites that 44 percent of sites were implementing parallel file systems for their high performance applications overall. As the firm’s CEO, Addison Snell, tells The Next Platform, “among these, PanFS was named as the parallel file system used by 10 percent of the respondents. That’s significantly behind GPFS and Lustre, but well ahead of the others in the rest of that pack, including OneFS, GlusterFS, Ibrix, and others.”
To be fair, Intersect360’s survey is distinctly focused on HPC, but for many sites in media and entertainment and life sciences that are not part of those surveys, finding a true breakdown is difficult. Nonetheless, it shows that the Panasas pNFS approach is still finding a fit. But Noer agrees the company faces stiff competition in media and entertainment in particular. Still, this understanding is not without a bright spot. He says that companies in life sciences and media and entertainment are faced with increasing demands on their I/O systems. Accordingly, he says key Panasas technology, including the ability to tap into direct access to storage, something Panasas has been pushing (while others, including IBM with GPFS have as well) to improve I/O performance in these markets.
The company’s DirectFlow, which has been a long-standing part of its core approach, allows all object-level access from Linux clients and allows any client to talk to all the storage blades in its appliances to get the different fragments of files directly without going through intermediary layers.
“In some other systems, users access the system via NFS or SMB and there is a backend InfiniBand network that shuffles pieces of data around so users are load balanced to one specific node, but chances are, your data isn’t there, so the system has to find it and send it back, which adds chatter and reduces performance.” One can assume that Noer is referring to Isilon in this case, but as one might imagine, having objective benchmarks to look at to compare how these two approaches compare is difficult (although we are certainly open to taking a gander at your findings here at The Next Platform).
This chatter, by the way, is inherent to NFS and SMB as protocols. When using NSF as an example, users mount a specific IP address as an accessing client. If there are multiple IP addresses, as is the case with any sort of scale-out approach, the clients are distributed across all those IP addresses, which ideally should add performance. However, that balancing act does not always work out well. It’s possible to have hot clients mounted to the same node together, as well as others that are idle, which means there is always the potential for hot spots. All of this affects performance scaling, as one might imagine. If load balancing favors a specific node and it only contains a small fraction of the scale-out file system data, for instance, then the rest is elsewhere. Hence the chatter aspect that Panasas seeks to address.
This is all done through its ActiveStor and PanFS file system bundles, which are only sold as appliances at this point. When asked whether or not Panasas might consider chucking the hardware side of its business to offer its capabilities as a standalone software service, he says that the hardware element does not add much overhead for Panasas. It is all commodity hardware, there are no custom ASICs, and the model has worked for the last sixteen years to the point where Panasas does not see value in operating as a software company alone. He does see how the market might favor such approaches in the future, however, he says for the markets they are going after that favor reliability, availability, and ease of management, customers are glad to have an integrated solution.