The modern workload set has fundamentally changed the nature of and also how we handle data sets. Previously, critical business applications were typically driven by the processing of large, monolithic databases that were comprised of large files. Today, more and more, businesses are being driven by the ingest and processing of many small files. Large data sets are frequently broken up into smaller working sets, with the end objective of processing data to feed applications, and consequentially, business users, as quickly as possible. More cutting-edge use cases such as artificial intelligence (AI), machine learning (ML) and Internet of Things (IoT) typically come to mind when we think of these workloads. But the reality is that this has become the norm for the vast majority of workloads.
The result of – and ultimately a key problem with – this paradigm shift is that legacy data center architectures can’t keep up. The practice of moving very large data sets to serve application requests is maxing out storage input/output (I/O) resources and adding substantial network latency. Both of these issues significantly slow application performance. Moving from a shared to a direct-attached storage model will not solve the problem because resources will inevitably be under or over-provisioned. Even adopting a premium NVMe-over Fabrics (NVMe-oF) networked storage implementation will not solve the core challenge, which is the need to move entire data sets the host CPU to serve every application request.
Treating storage as more than an “unitelligent” (when compared to counterpart server infrastructure), siloed data repository through adopting computational storage can help to address this problem. As defined in a previous Storage Switzerland blog, computational storage is the process of placing compute capabilities directly on storage to enable processing of data in place where it resides. Computational storage effectively “pre-qualifies” the relevance of data to an application request, on the storage drive itself, before migrating data to the main CPU. As a result, it can vastly reduce the volume of data that needs to be migrated between the storage media and the host compute node. The result is better (read: faster) application performance, especially for data-intensive workloads, alongside a more cost-effective compute, storage and networking footprint.
To learn more about how computational storage might benefit your applications, access our recent 15 minute on demand webinar with NGD Systems, “Rethinking Storage Performance for Modern Applications,” and download our eBook, Taking the Load Off of Networks with Computational Storage.