Organizations need to periodically rethink their high-performance storage strategies due to changing requirements for high-performance computing (HPC).
In the past, a lab or corporation might run a limited set of HPC applications. So a system and its storage could be optimized for a narrow set of workloads. Today, HPC workloads in most organizations are highly variable. One group might need to train a machine learning model, another might run very granular finite element simulation models, and yet another might require the lookup and high-speed search of a massive genomics database.
Each of these workloads has vastly different computational requirements; organizations can no longer rely on a solution optimized for a specific workload.
Further complicating matters is the type of data being used in HPC applications. Many more workloads today use large numbers of small files. In general, the number of small files is going up, and the average size of the files is going down. This has great implications when trying to match a suitable high-performance storage system to HPC processing power and memory systems.
Organizations must find a balance where appropriate storage is matched to file workloads to ensure a cost-effective solution that delivers the needed performance. Let’s take a look at what’s at issue and how an intelligent data placement architecture can help.
Why are these changes an issue?
With any HPC application, the challenge is how to keep the CPUs (and now GPUs) satiated to make the most efficient use of expensive HPC hardware. The combination of variable workloads and the growing number and small size of data files makes this harder to achieve.
If a system just uses HDD storage, its performance suffers if an application makes use of many small files. Conversely, SSDs will solve the performance problem for small files, but it is cost-prohibitive to use SSDs for everything at the high capacities HPC applications require.
Typical approaches use a small number of SSDs for so-called hot data (data that has recently been read or written), independent of the file sizes involved. While this can help workloads where data reuse is high, is doesn’t help workloads like AI where there is very little data reuse during any given training session. It’s especially challenging to use so-called “LRU” (Least Recently Used) caching with such a widely mixed workload.
Solution: The Panasas ASCA Intelligent Data Placement Architecture
Panasas ActiveStor® solutions provide high-performance scale-out storage for HPC applications. They address mixed workload and file size issues using a data placement architecture called Adaptive Small Component Acceleration (ASCA) https://www.panasas.com/company/why-panasas/high-performance/.
The ActiveStor solution lets organizations use the right kind of storage for each kind of data. Large files are stored on low-cost, high-bandwidth SATA HDDs, small files are stored on cost-effective, high-IOPs SATA SSDs, and metadata is stored in a database on low-latency NVMe SSDs. The solution includes just enough of the more expensive types of storage media, minimizing overall cost while still delivering the best performance.
ASCA is a hands-off, automated solution that dynamically moves data between SSDs and HDDs based on the mix of file sizes and the fullness of the SSDs. It ensures that any file smaller than about 128KB will be kept on SATA SSDs, while larger files are stored on HDDs, but ASCA automatically adjusts that 128KB line. If the SSDs are too full, the largest files stored on the SSDs are migrated to the HDDs, and if the SSDs are not full enough, the smallest files on the HDDs are migrated to the SSDs.
ASCA addresses performance issues in several ways. File performance is a combination of accessing the metadata for each file plus accessing the data in that file. The Panasas solution improves the raw speed of accesses to metadata by keeping the metadata in a database and storing that database on NVMe SSDs. SATA SSDs are a cost-effective and high-performance way to store and deliver small files without seek penalties. And finally, ASCA allows HHDs to consistently deliver their highest bandwidth because they are not burdened with any head seeks for small files or for metadata.
The net result of this combination is that the more costly device types are minimized, and every device is only doing what it is good at. Small file accesses will not impact the performance of large files, and metadata accesses never wait behind file accesses.
The performance benefits of using the intelligent data placement architecture of ASCA complement the many ease-of-management, and scalability benefits already delivered with Panasas ActiveStor and Panasas PanFS®, the operating environment for Panasas ActiveStor systems.
Simply put, Panasas provides scale-out storage with limitless scaling offering optimal data placement and an internally balanced architecture to boost efficiency.
For more information about matching storage to today’s HPC workloads, visit www.panasas.com