HPCwire | By John Russell | November 26, 2018
Is BeeGFS – the parallel file system with European roots – on a path to compete with Lustre and Spectrum Scale worldwide in HPC environments? Frank Herold’s unsurprising answer is yes. Herold is CEO of ThinkParQ, the company created in 2014 to commercialize BeeGFS. You may remember that BeeGFS got its start as an in-house project (2005) at the Fraunhofer Institute for Industrial Mathematics (ITWM) and was called the Fraunhofer Gesellschaft File System (FhGFS). It was later spun out as BeeGFS under ThinkParQ’s control although a fair amount of development still occurs in collaboration with ITWM.
Leaving aside the history aside for a moment, Herold was busy at SC18 this month making the case that BeeGFS’s mature technical strength combined with a solidifying channel landscape make it a worthy alternative (rival) to Lustre and Spectrum Scale. In making the case, Herold highlighted a recent win with partner Dell EMC to deploy BeeGFS at CSIRO (Commonwealth Scientific and Industrial Research Organization) in Australia.
“We started going more international a year and a half ago. It’s interesting that the number of international projects are not high, but from a project size perspective those projects are quite significant. As of today, combining North America and Asia Pacific, roughly 45 percent of our revenue is from those regions,” Herold told HPCwire. “We have a very significant deal with our partner and Dell EMC at CSIRO. It’s a 2-Petabyte, all NVMe, storage solution built for AI. This is in deployment right now.”
That’s a big win and potential proof point. ThinkParQ doesn’t publicly report revenues so it’s hard to accurately judge market traction and because BeeGFS is open source it is difficult to know how many organizations are using it or what it is being used for. Nevertheless, the timing of ThinkParQ’s push may be spot on.
Lustre’s future has seemed hazy lately. It is very strong in supercomputing centers but not so much in commercial HPC. Intel purchased Lustre provider Whamcloud in 2012 intending to build up a Lustre business but exited the business in 2017. This past June Intel sold its Lustre assets to DDN which as a leading HPC storage provider with several Lustre appliances may become a better steward for Lustre. (see HPCwire articles, Intel Open Sources All Lustre Work, Brent Gorda Exits, DDN Acquires Lustre File System Capability from Intel.)
In 2015 IBM rebranded its popular General Parallel File System (GPFS) as IBM Spectrum Scale. GPFS has also been a strong player in HPC and especially strong in high end enterprise computing.
Practically speaking, no parallel file system dominates commercial HPC noted analyst Addison Snell, CEO, Intersect360 Research, “BeeGFS has grown in popularity, particularly in Europe. However, no parallel file system is yet to be broadly adopted across commercial HPC segments, which constitute the majority of HPC usage.”
One observer directly involved in HPC storage technology selection and deployment agrees interest in BeeGFS is rising.
“With the current transition and uncertainty present both for Lustre and Spectrum Scale offerings in HPC, it is no surprise that this year there has been more buzz (forgive the pun) around BeeGFS than before,” said Aaron Gardner, director of technology, BioTeam, a research computing consultancy specializing in life sciences. “HPC has long been searching for a distributed parallel file system alternative. We’ve watched BeeGFS evolve over the past few years into a potential contender.” (diagram of architecture below.)
Ease of use, scalability, and powerful metadata handling capabilities are among the attributes that distinguish BeeGFS from Lustre and Spectrum Scale contends Herold. In February of 2016 the BeeGFS source code was “open sourced.” That said, ThinkParQ’s business model depends on providing supported versions with added functionalities and a few observers grumbled the terms of BeeGFS’s open source agreement impede maximizing its value without going through ThinkParQ. Major releases occur roughly yearly with minor upgrades roughly quarterly. Version 7 was released last May. One of the major new features, said Herold, was the addition of storage pools.
“Storage pools give customers capabilities that are across the name space. They can split up data and decide where the data are going on fast or high density storage underneath,” said Herold. Storage pools allow the cluster admin to group targets and mirror buddy groups together in different classes. For instance, there can be one pool consisting of fast, but small solid state drives, and another pool for bulk storage, using big but slower spinning disks. Pools can have descriptive names, making it easy to remember which pool to use without looking up the storage targets in the pool. The SSD pool could be named “fast” and the other “bulk” (diagram below).
Another interesting capability is BeeGFS On Demand (BeeOND). The concept here is to use storage on client machines rather than main storage.
“We have storage server on one end and then you have hundreds or thousands of clients. We build on the fly a kind of temporary BeeGFS file system which can take off some nasty workloads from the primary storage area and to those temporary workspaces. It’s is really a nice model where you have your primary datacenter running [but] you can also spin up temporary work spaces, and you can decide whether on all clients or just portions of clients for specific jobs,” said Herold.
As noted in BeeGFS documentation, the problem with the internal drives in compute nodes is that they provide neither the advantage of a single name space across multiple machines nor the flexibility and performance of a shared parallel file system. BeeOND solves this problem by creating a shared parallel filesystem on a “per-job basis” across all compute nodes that are part of the particular compute job, exactly for the runtime of the job (diagram below).
Best to review the BeeGFS documentation for a fuller view of its capabilities. HPCwire asked Ari Berman, BioTeam VP and GM of consulting services, for a quick assessment of BeeGFS versus its main competitors:
“BeeGFS is promising in a lot of ways. The major disadvantage of Lustre in our space has historically been the serial metadata access model that it uses, which makes the many concurrent file operations required for many life science workloads incredibly slow. Serious Lustre shops actively discourage users from running codes that do this, but that is really only tractable when you have direct control over how the codes running in your environment are written. The Lustre community has made modifications to provide a distributed namespace model that pseudo-parallelizes metadata access across directories, but it still doesn’t fully address all use cases for concurrent file operations.
“Spectrum Scale is a little better in that the NSD (network shared disk) servers can more easily be tuned for concurrent metadata access across metanodes, but you make those tuning choices at the expense of other performance gains. BeeGFS has the advantage of having a bit faster metadata out of the gate, while also being able to distribute metadata operations per directory and subdirectory across metadata nodes in a simplified manner. Like GPFS, you can add many more metadata targets (servers), as many as you need, and it will scale well and is a bit simpler to setup than multiple Lustre MDS.
“Another major advantage is the ability to thread the metadata request server using a built-in queue, and the ability to specify how many threads to spawn on each metadata server as needed. This avoids the serial request bottleneck when many (or even one) servers are making millions of requests for small files on the filesystem. One final plus for BeeGFS is while the server is written in user space, it does have a native kernel client that has been capable for the past few years of saturating 100Gbps client connections. The latest Spectrum Scale 5 or Lustre 2.10 LTS releases also get there with tuning, but there is a lot of variability in which versions of Lustre or Spectrum Scale the vendor and channel spaces are currently providing to customers.
“So, for us, BeeGFS is promising, but all of this is theoretical. We haven’t been able to get our hands on it to test drive it and test how well these features would work in this space. The fact that the software is free to download and use is a big advantage, and makes it much more accessible to our user base. But we note that certain features require licensed support from ThinkParQ, and the source code license while open is less permissive than Lustre for example.”
Like other parallel file systems, BeeGFS is mostly hardware agnostic. “From the CPU level we support everything on the market, whether it’s Power, Arm, Intel (x86), it doesn’t matter. From a storage perspective, as long as it shows up as a device, we can manage it. Also from an infrastructure level we play on TCP/IP and gigabit Ethernet and InfiniBand,” said Herold.
ThinkParQ is working to push to expand its market reach and Herold singled out HPC (scientific), life sciences, AI, and oil & gas as the key targets. Predictably, a key element in expansion plans is channel development. Currently, ThinkParQ is a services and software play offering commercial versions and support of BeeGFS. Getting into the hardware business for such a small company poses too many challenges. ThinkParQ does, however, have a few partners either bundling BeeGFS with systems or building appliances. Penguin Computing (serving the U.S. and Europe) and Taiwan-based QCT (Asia Pac) are two.
Said Berman, “Right now they are focused on building channel partners, but they currently are still far behind the depth present in the Lustre and Spectrum Scale ecosystems. Another disadvantage is that BeeGFS has yet to be used with the volume of customers and at the scale that Lustre and Spectrum Scale have been, so there are still likely edge and corner cases with BeeGFS that are yet to be encountered.”
Herold, who became CEO last February, is hoping engagements such as the CSIRO project will build confidence and demand. BeeGFS is hardly swarming, yet, but there’s a definite hum in the air.