With soaring data volumes and insatiable computing driving nearly every facet of economic, social and scientific progress, data storage is seizing the spotlight. Hyperion Research analyst and noted storage expert Mark Nossokoff looks at key storage trends in the context of the evolving HPC (and AI) landscape in this in-depth Q&A. One of the themes discussed is the rise of cloud storage.

HPCwire: Has HPC storage become more important with respect to the overall HPC ecosystem in recent years?

Mark Nossokoff: Servers and compute drive the majority of user spending and should rightly be top-of-mind in terms of requirements and procurement conversations. Historically our research indicates that servers drive just under 50 percent of user-spend on HPC-related on-prem infrastructure. That said storage is historically the second largest area of on-prem HPC spend at approximately 20 percent of users’ HPC technology investments. Alternatively you can view it as for every $1 spent on on-prem compute, another $0.20 is spent on storage. Storage also historically projects the highest growth rate of users’ on-prem HPC-related infrastructure. For completeness, I should also note HPC personnel costs are important and increasingly scarce but we don’t track non-technology items.

HPCwire: Are there reasons other than financial that drive HPC technology conversations more towards compute than storage?

Nossokoff: HPC has historically been dominated by what’s now referred to as “traditional” HPC modelling and simulation (mod/sim) workloads. These mod/sim workloads were driven by primarily compute-intensive applications and as such, attention and investment were largely focused on making compute threads faster, getting more compute threads running in parallel, and minimizing time spent doing things outside of the server. This included provisioning larger and larger amounts of memory on the technical servers and accessing networks and storage as little as possible. Clock speed, core counts, and flops were the primary topics of conversation. Despite that, traditional mod/sim did drive two primary areas of storage requirements and innovation. First is the capacity to store the ever increasing amounts of data required to drive the simulation and store the resultant simulation outcomes. Second is write bandwidth performance (Gb/s) for checkpointing to quickly save the state of long-running mod/sims in the event of catastrophic system failures interrupting the job.

HPCwire: Is there anything happening to elevate the mindshare and importance of HPC storage?

Nossokoff: Yes. There are several maturing areas across technology and business processes that are dramatically changing the HPC storage conversation, including the emergence of cost-effective solid-state storage media, the adoption of data-intensive AI workloads by both traditional HPC and enterprise IT datacenters, increased utilization of the cloud to run production HPC workloads, and the advent of viable consumption-based business models.

HPCwire: Can you briefly elaborate on each of those?

Nossokoff: Sure, but I guess it depends on your definition of “briefly”…

Let’s first look at the emergence of cost-effective, high performance solid-state storage media.

Mass storage for HPC has long been deployed on magnetic media, both tape and rotating magnetic disk-based hard disk drives (HDDs), with working sets and local data storage on HDDs and archive data storage on tape. The mechanical nature of both tape and HDDs provided inherent latencies and performance bottlenecks that effectively capped the economically feasible performance of most HPC systems. The emergence of cost-effective solid-state storage, initially in the form of high performance flash-based solid-state drives (SSDs) began to change that. No longer throttled by HDDs, compute nodes can now move more data in and out of the servers at a faster speed, enabling much faster completion times of the even the most demanding mod/sim applications. The notion of tiering emerged (“hot” storage for frequently accessed data, including both metadata and simulation data, “cold” storage for less-frequently accessed data, and “archive” for long-term data retention), as did further opportunities for read-intensive and write-intensive workloads.

Next is the adoption of data-intensive AI workloads. The value of AI is highly dependent on the accuracy and precision of the AI model. The accuracy and precision of the AI model is largely dependent on the quality and the amount of data input into the model. The creation of and updates to the model, leading to the speed at which inferencing can be derived form the model, are then primarily of function of how quickly data can be fed into the systems doing the modeling and inferencing. The adoption and deployment of GPUs to implement AI-based solutions quickly exposed the storage performance challenges of traditional HPC mod/sim machines. Expensive compute nodes were sitting idle or underutilized, waiting for data to be read from and written to the storage system. These storage challenges created storage innovation opportunities to elevate the value and importance of storage within the ecosystem of HPC-enabled AI infrastructure. Innovations had occurred across all facets of storage (e.g., file systems, interconnects, tiering SW, storage class memory supporting  subsequent in-memory computing), creating the ability to design cost-effective systems traditional enterprise IT datacenters can employ to address their data-intensive AI and HPDA workloads, in addition to augmenting traditional mod/sim application, aw well.

Now let’s turn to the increased utilization of the cloud to run production HPC workloads. Our research not only indicates that approximately 20 percent of users’ AI-related workloads are run in the cloud, but it also suggests users are starting to delay and divert on-prem HPC spending to run their workloads in the cloud. Regardless of how many or what types of workloads are being run in the cloud, the business process and workflow management of hybrid-cloud, multi-cloud and cloud-native HPC workloads has a direct impact on HPC storage. Where should the storage be placed? When should it be moved, if at all? What are the cost implications? Are there security or data locality issues? Am I looking for just cloud storage or do I have an equal need for cloud computing, as well? These are but a subset of considerations for users and HPC datacenter managers to consider relative to HPC storage and the cloud.

The cloud storage offerings of the CSPs are critical to understand, as well, not only between the primary CSPs (e.g., Alibaba, AWS, Google Cloud Platform, IBM Cloud, Microsoft Azure, Oracle Cloud Infrastructure), but also between the different types of CSPs (e.g., the aforementioned “primary,” HPC-specific [e.g., ReScale, Verne Global, UberCloud], and domain-specific [e.g., DUG]).

Lastly let’s look at the advent of consumption-based business models. While related to utilization of the cloud, this item refers to bringing cloud-based-consumption business models and practices to infrastructure placed on the users’ premises. Whether it be more of a utility model (e.g., capacity-based) or quality of service service-level-agreement-based (SLA) model (e.g., reliability, availability, performance), tools and management hooks are required within the storage system to properly instrument and provide insight and transparency into the operation and behavior of the system.

HPCwire: Is there anything architecturally that can be done to better address storage and data I/O needs?

Nossokoff: While there’s no “one size fits all” architectural answer, one philosophical area with a couple of dimensions come to mind: homogeneity vs. heterogeneity, first relative to interconnects and second, the storage system itself. From the interconnect perspective, there are two choice. The first being between multiple, separate interconnects for system traffic (e.g., interprocessor communication, MPI message passing) and storage traffic (e.g., transferring data between the compute nodes and storage subsystems), and the second being a single, converged network to support both system traffic and storage traffic

Relative to the storage system itself, the question is effectively a tiering and/or partitioning tradeoff  between multiple storage tiers or partitions, with each tier optimized for a particular I/O profile and a single, flat tier that is capable of adequately – if not optimally – all types of I/O profiles.

Examples of I/O profile-based tiers include 1) Small, fast I/Os for metadata and message passing 2) Large, fast I/Os for datasets for large simulations and AI processing, 3) Large, moderate I/Os for nearline storage of recent simulations and accessed data and active archiving, and 4) Large storage capacity for less-active archiving and long-term data retention.

The pragmatic answers to both interconnects and storage system design are likely “it depends” and “some combination”, but it is critical to understand what workloads will be primarily run on the machine, along with the overall system utilization expectations to determine the proper architectural starting point.

HPCwire: Are there any emerging storage-related technologies that could materially impact the landscape in five or more years, if not sooner?

Nossokoff: There are several areas of storage research and investment occurring at both ends of the HPC spectrum. On the high-performance end, the continued evolution of solid-state media (e.g., storage class memory, MRAM, ReRAM) and associated in-memory computing could provide the next step-level of application performance improvement for latency-driven workloads whose datasets can fit entirely in memory. Also relative to latency-driven applications are the conversations occurring around POSIX compliancy for storage I/O. POSIX support has been a solid requirement since the advent of multi-threaded applications to prevent one application from conflicting with and overwriting data from another. With certain systems becoming more specialized and more control being exerted over the specific utilization of the system, certain aspects of POSIX compliancy may be able to be relaxed to improve system data I/O latencies and overall system application performance.

On the capacity-driven end of the spectrum, file systems are emerging to address and optimize for the growing active archive market. Additionally, new storage media (e.g., synthetic DNA) is being evaluated to address the cost, reliability, durability, and longevity requirements of the long-term data retention space relative to its magnetic counterparts.

HPCwire: Can you comment further on the rise of storage/memory hierarchies and how this complexity is being addressed?

Nossokoff: The increased attention being given to memory in the data path hierarchy is due to the recognition of how improvements in this area can drive further meaningful improvements in application completion performance. Low-hanging-fruit is already being harvested with the adoption of more cost-effective flash-based technologies, supporting memory sizes that that allow complete datasets to be contained in the higher speed memory footprint. Additional complexities are being addressed by what some are characterizing as Big Memory; that is, the deployment of data services seen traditionally on enterprise-class external storage such as replication, snapshots, dedupe, and compression on data residing in memory. Others recent innovations addressing memory are the evolution of High Bandwidth Memory (HMB), adoption of newer intranode interconnects such as CXL, and newer interfaces and protocols such as GPU Direct to eliminate unnecessary movement and copying of data.

HPCwire: Are there any vendors emerging as the leader in providing best-in-class HPC storage?

Nossokoff:  That’s a bit of a loaded question. And it’s dependent on whether the discussion is around on-prem storage or cloud storage.

On-prem HPC storage providers largely fall within two categories: system vendors that also provide their own storage and independent storage vendors. While HPC system vendors provide the lion’s share of on-prem storage, recent surveys suggest a subtle shift towards independent storage vendors. When users were asked in a recent survey from whom they acquired storage for their primary HPC system, Dell, HPE, and IBM lead on the system side while DDN and NetApp lead on the independent storage vendors side. That said, many other independent storage vendors are heavily innovating to address specific opportunities (e.g., filesystem for heterogenous traditional mod/sim and data-intensive AI workloads; all-solid-state systems) and specific markets (e.g., AI, active archive, long-term data retention).

From the cloud storage perspective, AWS currently leads but the others are not sitting idly by. Google Cloud Platform and Microsoft Azure are investing and leveraging their respective strengths to further capture cloud storage market share.

Bio

Mark Nossokoff, senior analyst, Hyperion Research, has over 30 years of experience in the data storage technology field and directs his research and consulting efforts in associated areas. Nossokoff’s experience includes hardware design, application engineering, product marketing, product management and strategic planning and co-inventor of (four) patents. Before joining Hyperion Research in 2020, Nossokoff worked primarily in the OEM Storage B2B market for NCR, AT&T, Hyundai/Symbios, LSI (Engenio), NetApp (Engenio), LSI again (server IO and PCI RAID), Dot Hill and most recently with Seagate.