Building a High Performance, Scalable Object Service: Three Strategies to Leverage

New pressure on CIOs to deliver high performance storage at unprecedented scale Today’s data solution managers are stretched in more directions than ever before, attempting to meet all the needs for performance while balancing all of the issues of cost at scale. The rapidly changing face of enterprise storage. In the past, high-performance enterprise storage solutions could be measured in the 10s of TB of capacity. Storage appliances designed to meet the performance requirements did not need high capacity, so they were designed for high-availability on a single server only. Getting larger capacity was done by adding capacity serially with additional servers. By contrast, applications today are demanding both high performance and capacity in excess of 10 PB, a trend driven primarily by two use cases.

• • Analytics: In the past, large analytic workloads required that map-reduce architectures distribute the analytic workloads into the level of many thousands of nodes through products like Hadoop. Over time, these workloads have shifted to a more disaggregated architecture that flexibly allows compute to scale separately from the storage. Traditionally at least 1 PB, these workloads can grow beyond 1 exabyte.
  • AI/ML training: Enterprises are finding new ways daily to use data that was put into cold storage in the past and assumed to have little value. Often maintained in object stores, things like images and other blobs bring not only an increase in data volume, but that data is getting “hotter” and more valuable to the enterprise – driving up storage performance requirements. More data needs to be accessed faster by more applications.

For both these use cases, object storage has emerged as a simple option for PB-scale storage, and simplicity is key for both successful setup and long-term data center deployment. Modern architectures and business models pave the way to a solution. In response to the challenge of providing both high performance and orders of magnitude with greater capacity without breaking the bank, the key storage market requirements point to three interdependent strategies:

• • Design storage servers from the bottom-up to focus on high density and low cost, without the compromises needed for a general-purpose server.
    Goal: Hitting price/performance targets for petabyte-scale applications while at the same time addressing sustainability issues.
  • Utilize cloud-like business models that reduce upfront capital expenses for additional capacity.
    Goal: Affordability at whatever the current scale happens to be.
  • Follow an approach that abstracts data storage from the underlying hardware through Software-Defined Storage (SDS). SDS software simplifies storage management and scaling, and achieves higher storage utilization and high resiliency/availability without specialized hardware.
    Goal: Scalability and flexibility for large workloads.

Samsung has been working diligently with partners to co-design solutions that leverage these three strategies. PBSSD: all-Flash storage server designed for petabyte-scale workloads When designing servers, most OEMs create designs to be as broadly usable as possible. While this approach retains maximum flexibility, it also means that for specific purposes, they may have to incorporate many features unnecessary for storage – leading to designs that are not optimal for large object or large workload storage applications. The advent of a new generation of CPUs brought an opportunity to improve storage server design, setting the stage for petabyte SSD (PBSSD) architecture. PBSSD is based on the 4th Generation AMD EPYC CPU, using a single socket to offer 32 to 84 cores and 128 PCIe Gen 5 lanes. There were a number of design considerations made.

• • High capacity: While many servers in the past handled up to 100 TB of storage, Samsung wanted to push the limit much higher. PBSSD can handle up to 16 E3.s data SSDs in a 1RU form factor server. Each SSD is capable of up to 15.36 TB of storage for a maximum of 244 TB in a 1 RU chassis. An upcoming system design will support 1 PB in a 2 RU chassis.
  • Simplicity of design: Many servers rely on PCIe switches, HBA buses, or at least retimers, to accommodate enough SSDs. PBSSD architecture is able to do away with all these, connecting 16 data SSDs directly to the CPU for maximum performance.
  • Reduced power: One of the inherent advantages of the simpler design of PBSSD is that using fewer components also means less power draw. Samsung takes it a step further with additional SSD-level power reduction techniques.
  • Reduced TCO: The design philosophy of PBSSD is to reduce storage TCO without sacrificing performance. While each of the above points is important individually, they collectively point to decreasing both the CAPEX and OPEX of managing storage.

Delivering storage-optimized servers. PBSSD uses NVMe protocol and connects PCIe SSD directly to the PCIe interface of the server CPU, eliminating retimers and PCIe switches. This architecture delivers a staggering performance. Addressing sustainability through performance-to-power ratio and power management. By providing incredibly high performance, the PBSSD design is easily able to offset increased power consumption.

NVMe power efficiency:	608MB/s per Watt
Compared to SATA:	186MB/s per Watt

Advanced power management and built-in power consumption reporting capabilities tailor power and performance to specific applications and data-center limitations. Cloud-like business model: affordable monthly subscriptions at your chosen location Taking its cue from the affordability and scalability of cloud business models, Samsung does them one better. By offering a hardware-based, on-site storage subscription model that reduces upfront costs of storage capacity, the Samsung approach helps focus capital expenses on building core cloud functionality and value. There are various attractive facets to this cloud-like business model:

• • Customers place an order for the desired storage capacity in increments of 244TB (one PBSSD unit) and subscription duration (1, 3, or 5 years) at a fixed monthly fee. No hidden or usage-based costs.
  • Customers are in full control: Samsung ships PBSSDs for the customer to install. Security is fully under the control of the customer at all times. Samsung does not monitor or access the hardware remotely.
  • When ordering products, customers can take advantage of Samsung partnerships with software-defined storage providers to deliver pre-certified object and file solutions like MinIO, Weka, and vSAN. Customers are also free to engage separately with the SDS software provider of their choosing, or even to use open source software like Ceph.
  • Samsung provides channels for customers to work with managed service providers like Opsara to provide them both hardware and software bundled as a managed object service.

PBSSD subscriptions deployed by private cloud owners and operators in their data centers or co-location facilities, benefit from the most reliable and highest density flash storage infrastructure, achieving lowest possible operating expenses by saving rackspace, power, and maintenance costs. SDS: balancing capacity and performance with demand, on-the-fly It’s easy to state that capacity and performance should be readily scalable and software-defined. But when “SDS” can refer to so many different design and implementation aspects, what specific ones are important to consider for large object storage? Perhaps it would be easier to specify the selected solution first, wherein the advantages will become immediately apparent. Samsung has chosen to partner with MinIO, the leader in “Amazon S3-compatible” object storage. What is S3 compatibility, and why is it important? Many applications are already using a well-understood set of REST API calls to access AWS S3 cloud storage for objects such as images, log files, videos, PDF, artifacts, ML models, and LLMs. Being able to simply redirect these same API calls to MinIO-based servers makes it an almost trivial matter to convert existing applications to use MinIO-supported storage. MinIO uses familiar HTTP GET and PUT calls to initiate object transfer and management sequence. This cloud native approach is far more efficient than using half-century old file-oriented commands like copy, move, and delete. Moreover, MinIO bakes content, metadata, version, and security into the recipe for handling objects, simplifying the task of maintaining data integrity and recoverability across servers anywhere. A MinIO-based solution provides an incredibly flexible and scalable platform for object storage that works seamlessly across public and private clouds – offering robust and secure storage even at exabyte scale. Why Samsung and MinIO? In choosing Samsung and MinIO, you benefit from the two market leaders. But MinIO works on any platform – so why choose Samsung? Because Samsung and its partners have been implementing scalable, performant, and affordable storage that supports new workloads and data center storage demands in a way that is unlike any other competing solution.

• • Samsung is the undisputed leader in NAND/SSD technology and business, and truly understands the potential and tradeoffs and has a unique insight into the future direction and costs of Flash storage. Its PBSSD hardware architecture enables servers purpose-built for large object storage at the right price.
    • • Direct-connect SSDs to CPU
      • High capacity E3.s for density, low power/TB
      • Gen 5 SSDs for performance
      • Right-sized core count
      • Flexible location - you choose where to deploy the hardware on-premises or at your co-location facility, wherever your data is
  • MinIO is the leader in S3-compatible object storage, providing scalable solutions from PBs to EBs that effectively eliminate the challenges of data storage, redundancy, backup, rebuild, and revision control. Unchecked, these can put incredible demands on the underlying storage technology to achieve the necessary performance.

MinIO and Samsung create a seamless storage fabric for the most demanding workloads. Sixteen directly connected PCIe Gen 5 lanes, compute power up to 84 cores, and sophisticated power reduction techniques result in breathtaking performance with sustainability. MinIO operations, such as maintaining data integrity and rebuilding after a data loss event, take place with virtually no performance impact on active requests. Teaming up Samsung PBSSD architecture and MinIO object storage changes the paradigm from “frantically recovering from a data loss event” to a more future-forward goal of “optimizing secure object access with alacrity.” MinIO manages the objects such that they are encrypted, erasure-coded, securely backed up, and revision-controlled; Samsung provides storage mechanisms that eagerly serve those objects, supporting MinIO on an exabyte scale, with the greatest performance possible. Conclusion Choosing to move to a scalable ecosystem using Samsung PBSSD with MinIO provides broad flexibility with virtually no ramp-up effort. If you are:

• • A CIO or data center storage manager
  • Seeing new workloads and facing storage performance or cost challenges in your data center
  • Looking to prepare for changes you see coming, but don’t know yet what the pace of scale-up will be
  • Attempting to leverage any hardware investments your company makes by having cloud-based capacity expansion options

We’d like to hear more and see if we can help. Meet us at NVIDIA GTC, booth #528, or watch our video on-demand session #S63226. You can also visit the event page here to watch the video live at 4:00pm PDT on March 18, 2024.