43 Scality today announced RING XP — a special configuration of our RING object storage software that is optimized for the extreme performance demands of AI model training and fine-tuning. Applications in this domain need access to small data objects in the millions and billions (or more), where the per-object size is in the range of a few kilobytes (KB) or smaller, and the system also supports large (GB+) objects very efficiently and with high throughputs. The key storage performance goal here is to achieve extremely low access latencies. For small object (4KB and lower) workloads, RING XP achieves read and write access latencies measured in 100s of microseconds — a level of performance that has never been achieved before by any object storage solution on the market, whether on-premises or cloud-based. RING XP – eXtreme Performance solution for AI model training and fine-tuning with AI applications. How we achieved the results: Context for RING’s eXtreme Performance For context, RING XP responding in microsecond latencies can be compared to: Standard AWS S3 cloud object storage: Amazon S3 delivers latencies between 10–50 milliseconds, though this is not deterministic or bounded by an SLA, and latencies are often higher. Amazon S3 Express One Zone: AWS introduced a streamlined (less functional and not S3 API-compatible) object interface in late 2023. In exchange for reduced functionality (changed bucket model, data only in one region/zone), it promises “single digit millisecond” latencies. All-flash file systems (not object storage): All flash-file systems that are appliance-based — often only on proprietary servers — can deliver latencies in the 100s of microseconds over file (NFS) protocols. While some of these do offer S3 object storage interfaces, they don’t publish object performance data and are nearly certain to insert higher latency over object protocols than via file system interfaces. Only RING XP can scale to exabytes and provide microsecond latency. RING XP performance testing was conducted in the following environment: Software version: RING/RING XP 9.3.0.8 Platform: Dell PowerEdge R7615, AMD EPYC™ 9124 3.0GHz, 3.2TB PCI/NVMe mixed-use drives Operating system: Rocky Linux 8.6 (kernel version 4.18.0-372) Network: 1 x 100GbE Test driver: Client on Dell PowerEdge R7615 server writing and reading 4KB objects over RING XP object storage API (PUT/GET requests) Response-time latencies measured on the client include send/receive time for 4KB data object RING XP object request performance distribution Performance levels achieved: In this environment, RING XP delivered consistent access latencies under 511 microseconds (GET) and under 741 microseconds (PUT) for a majority of API requests. This level of latency makes RING XP 10–20X faster than Amazon S3 Express One Zone and 50–100X faster than standard Amazon S3, putting it in the same range as hardware-centric (not software-defined) all-flash file systems that are often more costly. We, therefore, assert that RING XP is the world’s fastest object storage solution, making it an ideal match for the needs of AI applications for model training and fine-tuning. RING XP technical details RING XP is built on Scality’s proven RING distributed software stack. It has all of the comprehensive capabilities of standard RING in terms of scalability (in multiple dimensions), data durability, high-availability, self-healing and simplified systems management at scale. The building blocks of RING XP are: High-performance object storage API Ideal target for AI applications, tools and file systems via simple object APIs Virtually unbounded, flat namespace = major simplification vs. file systems + a natural fit for AI apps built with cloud-native design principles RING storage stack Scale-out, distributed system with no single points of failure for high availability and ability to grow in multiple dimensions All advantages of RING for high data durability (erasure coding, replication, self-healing and more) Optional: Reduced data durability policies for maximum performance on small objects (ideal for scratch storage) RING software deployed on AMD EPYC™, all-NVMe flash server platforms Reference platforms from HPE, Supermicro, Dell & Lenovo at launch Making the strong tradeoff of API features for extreme performance Scality customers who are investing in AI projects tell us that they are in a race for extreme performance — as Amazon has shown the market with S3 Express One Zone — and the right tradeoff for these applications is to eliminate unnecessary API features and heaviness in exchange for low latency and high throughput. With that said, we believe that Amazon did not go far enough with S3 Express One Zone, leaving its users with a weak tradeoff that doesn’t provide sufficient performance gains. With RING XP we’re going even further by simplifying the object stack and removing features that are commonly performed at a higher level in the stack. For example, every data lake management system (or data lake house) maintains an indexed view of all data across the entire dataset. That’s why we made the decision to remove object listing from RING XP since it would be an unnecessary duplication of functionality and effort. One way to think of RING XP is as an infinitely scalable distributed data storage system that is addressable through a fully routed TCP/IP network, and is flexible and efficient for objects that range in size from a few hundred bytes to gigabytes and beyond. The RING XP object storage API is a simple, fast, REST-based connector optimized for demanding small object AI workloads. It is stateless, lightweight and highly scalable, providing a flexible API for interacting with RING XP through GET, PUT, HEAD and DELETE commands. Because it is so lightweight and has a low resource footprint, it can be deployed on application servers, on the storage nodes themselves, or even on dedicated proxy servers in a disaggregated deployment. The connector offers two ways to store and retrieve data (by key and by path), to provide developers with a flexible choice to fit their design patterns. For applications that require the richness and functionality of the full Amazon S3 API, RING provides a comprehensive implementation. For extreme performance, the RING XP object storage API is the right answer for AI applications requirements. Patented architecture At the core, RING XP enables microsecond access latencies through several key innovations, all of which are covered by Scality’s core patent: Low-latency DHT: Our architecture builds a distributed hash table over the storage nodes, ensuring guaranteed, bounded access times without needing a central database. This eliminates bottlenecks typical in centralized architectures. O(1) access: In normal operations (no disruptions), access nodes cache the entire topology, enabling O(1) access to any object in the system, contributing to our low-latency performance. Shared-nothing architecture: The system uses a shared-nothing architecture while still striping data across multiple servers and disks, which maximizes performance by distributing loads and avoiding single points of failure. Late materialization: This technique enables network operations to first hit memory or fast metadata flash for IO requests, optimizing performance (latency) by shielding the disk drives used to store data for all non-data operations. With NVMe flash at the storage layer, this further accelerates data access to maintain microsecond response times through the full IO stack. The architecture achieves microsecond latencies through the software stack to the physical storage media as depicted in this diagram: Flexible data durability policies As described above, the RING stack is the basis for RING XP, meaning that all of the goodness of RING is available to RING XP users in terms of data durability and data integrity assurance. Ultra high-levels of data durability are achieved through policies that implement erasure coding (optimal for large sized objects), data replication (optimal for small sized objects), checksums stored with each object (with active bit rot detection) for data integrity assurance, and self-healing after component failures (disks or entire storage servers). RING can achieve data durability levels of 14 nines through these mechanisms. In the classic high-performance computing (HPC) space, the concept of scratch storage was used for temporary, high-speed data storage and intended for files that are frequently overwritten or deleted (and, in some cases, can be recomputed quickly if needed). This is also the case in AI applications such as intermediate data during model training. For small object AI data, some applications may trade off ultra high data durability for the best possible performance, meaning the lowest latencies. RING XP can, therefore, be configured in different modes to fit business and performance requirements: For high durability (redundancy) requirements: RING XP can be configured with a dedicated data protection schema for small objects, using object level replication. This efficiently creates replicas of the data rather than erasure coding, as is commonly more suited for large objects. Optionally, for scratch storage: RING XP can be configured with zero data redundancy, allowing the optimization to be on maximizing performance of small files with absolute minimal latency. RING XP licensing and platform support Scality’s goal is to provide our customers with the ability to achieve maximum benefit of using RING XP for their high-performance AI applications. To that end, we have maintained the same licensing model as for RING. Licensing remains on per-usable-TB basis, under a subscription model for 1, 3 or 5 years with the same price as for standard RING configurations. This means there is no pricing difference between standard RING and RING XP. Scality RING customers have immediate access to RING XP and can reap the full benefits of its extreme performance by deploying the software on one of the supported reference AMD™-powered platforms from Lenovo, Supermicro, Dell and HPE. Contact Scality for full details on pricing and platform support.
