High-Performance Computing (HPC) is going mainstream. Commercial organizations, with increasing frequency, are creating HPC architectures. These organizations have large amounts of unstructured data that they need rapid sequential access to for analysis. Also, many organizations require “Big Manufacturing” which is the design, simulation, assembly and testing of physically large and complex products like airliners. Finally, in our content-driven world, many organizations have media and entertainment workloads which are made more complicated because of the need for special effects, augmented reality and virtual reality.
Commercial organizations have a unique set of requirements that differ significantly from those of HPC’s traditional home of academia. These organizations don’t have the time or resources to assemble the various components of the traditional HPC project. Instead, they are looking for a more turnkey solution that can deliver value immediately after implementation.
A commercial HPC solution has two essential elements; software, and hardware. The software delivers the filesystem, data management and cluster management capabilities of the solution. The hardware delivers the computing that the software needs to execute and the storage media to store the data and respond to user requests. Most traditional HPC environments select open-source filesystem software and select the hardware of their choosing. The result is almost too much flexibility for the organization. The open source software solution is sometimes difficult to support, especially if the customer identifies bugs. Even the hardware is problematic. Differences in CPU BIOS and network interface firmware can dramatically impact the stability and performance of the cluster.
To address the challenge of too much flexibility in the commercial HPC space, vendors have offered turnkey solutions, which are actually pre-assembled systems that still use open source filesystems and commodity hardware. The pre-assembly does often limit initial exposure to cluster stability problems as a result of firmware incompatibility. As the cluster expands, however, the customer finds themselves in the same situation of managing the hardware nuances in addition to supporting open source software.
Another possible method to address the problem is to take the extreme approach of selecting a vendor with an HPC specific filesystem and customized, even proprietary hardware. Assuming the vendor maintains price competitiveness, this approach does solve several problems, but it does limit the speed at which the customer can take advantages of new advancements in hardware. The commercial HPC customer may find themselves behind the curve, while the vendor tries to integrate support for new hardware advances. The vendor may also have a more difficult time remaining price competitive since it too is limited in its OEMs supplier relationships.
A third approach is to take the middle ground and create a filesystem targeted at HPC with specific advantages over its open-source competitors but making that software portable across multiple hardware vendors. The HPC storage vendor still supplies the hardware, but the more inclusive approach to hardware enables them to adopt new technologies and to leverage multiple suppliers to remain price competitive.
Panasas ActiveStor Ultra – A Portable Filesystem for Commercial HPC
Panasas has a reputation of being the ideal commercial HPC product. Organizations in the commercial HPC market typically place a higher value on manageability and reliability than on performance. Panasas was closer to the second option of an HPC specific filesystem and customized hardware. The problem for Panasas is an increasing number of their customers also needed the highest levels of performance.
The third approach, portable HPC filesystem software, is precisely the approach that Panasas is taking in the latest release of its software. ActiveStor Ultra is a portable filesystem, not a software-defined filesystem. The portable nature of the filesystem enables Panasas to enhance flexibility by using more modular components. Also, because the software can leverage the latest technology it increases performance. Panasas claims 75GB/s of bandwidth per rack. The use of industry-standard hardware also reduces Panasas’ cost, enabling them to be one of the most competitive solutions in the HPC market.
The Panasas Filesystem (PanFS) stores metadata and data differently since the access patterns of metadata are so different than that of data. Metadata makes up over 80% of the IO requests in the typical HPC environment so to accommodate PanFS every storage node in the Panasas cluster stores its metadata in its own dedicated database. Panasas stores that database on NVMe flash media inside the storage node. PanFS uses a highly scalable object storage backend. PanFS 8 is optimized to match the object access and update patterns common in mixed workloads.
ActiveStor Ultra Hardware
An ActiveStor Director controls the ActiveStor Ultra hardware. The Director is the brains of the operation and interfaces with the compute layer network. The director is a 2U chassis with four nodes to handle filesystem and cluster management. The ActiveStor Ultra storage nodes come in 4U enclosures. Each storage enclosure contains four storage nodes. Each node has six hard drives; capacities can range from 4TB to 12TB per drive. Each node also has between 2 and 4TB of SATA SSD for small files and a 2TB NVMe drive for metadata.
Commercial HPC, as well as traditional HPC, are multiplying both in use cases and criticality to the organizations that use them. Before ActiveStor Ultra customers might have to choose either a fast system but sacrifice manageability and reliability, or they had to choose a manageable and reliable system but sacrifice performance. Now Panasas gives them the potential to have it all, manageability, reliability and high performance.