During the last two decades, the data center world has been moving to a “Software Defined Everything” paradigm. This has been taken care of mostly by hypervisors running on the x86 up to recently.
In parallel, a new communication protocol to interface with SSDs has been specified from the ground-up, allowing to fully exploit the levels of parallelism and performances of all-flash storage: NVMe, and NVMe-oF. NVMe-oF promises to enable the performances of direct attached all-flash storage with the flexibility an TCO savings of shared storage. To fully unlock the benefits of NVMe-oF while keeping the software defined paradigm, we believe a new kind of processor is needed: the Data Processing Unit, or DPU.
Emerging and existing applications with cloud computing, 5G, IoT, automotive, and high-performance computing are causing an explosion of data. This data needs to be processed, moved, and stored in a secure, reliable, available, cost-effective, and power-efficient manner. Heterogeneous processing, tiered memory and storage architecture, infrastructure accelerators, and infrastructure processing units are essential to meet the demands of this evolving compute and storage landscape. These requirements are driving significant innovations across compute, memory, storage, and interconnect technologies. Compute Express Link* (CXL) with its memory and coherency semantics on top of PCI Express* (PCIe) is paving the way for the convergence of memory and storage with near memory compute capability. Pooling of compute, memory, and storage resources with CXL interconnect will lead to rack-scale efficiency with efficient low-latency access mechanisms across multiple nodes in a rack with advanced atomics, acceleration, smart NICs, and persistent memory support. In this talk we will explore how the evolution in load-store interconnects will benefit the compute and storage infrastructure of the future.
When designing a new system from the ground up, there are multiple ways to implement Computational Storage (CS) in a new architecture, depending on the form factor, the computing type (CPU or FPGA), the environment (edge, cloud).
This talk will present a review of the available CS technologies, and will explain how to evaluate which CS technology provide the best benefits for specific workloads, including storage services, database and AI. System architecture examples will be described with use case simulation results.
This talk will give an overview of what’s been added to the SMB3 protocol documentation for the latest release as well as a deep dive into the using SMB3 over the QUIC transport. The SMB3 protocol is broadly deployed in enterprise networks and contains strong protection to enable its use more broadly. However, historically port 445 is blocked and management of servers on TCP have been slow to emerge. SMB3 now is able to communicate over QUIC, a new internet standard transport which is being broadly adopted for web and other application access. In this talk, we will provided updated details on the SMB3 over QUIC protocol and explore the necessary ecosystem such as certificate provisioning, firewall and traffic management and enhancements to SMB server and client configuration.
Learning Objectives: You will learn the details about SMB3 on QUIC; how it works, the architecture and performance comparison with SMB on TCP. Documentation updates related to SMB3 on QUIC will also be presented.
Data science and machine learning are hugely popular topics at the moment, but those topics don’t do the actual work involved justice. Most of the time is not spent on tweaking the features of the model or pruning it to optimise for edge deployment. It's spent finding, assessing and them wrangling the data into something meaningful and useful that the algorithm can get to work on. In this talk we will explore some of the challenges facing the data hunters and look to some methods that can help speed the process along and allow for repeatable data operations.
Bucket notification has become widespread in many applications including AI/ML and Edge Computing. We have demonstrated that Ceph Rados Gateway (RGW) bucket notifications trigger Serverless functions including Knative and AWS Lambda. Such workflows can be characterized as event push and event driven.
We’ve identified yet another workflow that allows Serverless computing frameworks to be more preemptive in auto scaling. The workflow is based on the event queue polling model, the opposite of event push, that allows Serverless computing frameworks, such as KEDA, to preemptively scale the functions based on the queue length. Such workflow is more advantageous and lightweighted in dealing with highly bursty event traffic and reducing Serverless function cold start overhead.
In order to support this workflow, we propose a new set of message queue APIs, modeled after AWS SQS, for Ceph RGW. In this talk, we’ll present the overview and planning of this technology to the community.
Presentation will also try to explore what are the best approaches for getting data from the Edge to the Cloud.
Artificial Intelligence (AI) in medical applications is an emerging area that aims to analyze medical big data to improve screening, diagnosis and treatment of disease. One challenge of this area is providing storage systems to efficiently store and preserve the data for future analysis within the patient lifetime and beyond. The European Union H2020 BigMedilytics project includes a pilot to analyze clinical and imaging data for early prediction of breast cancer treatment outcomes. In this talk, we’ll describe SIRF and how it can be applied in the BigMedilytics breast cancer pilot.
Data and Storage management across Edge, On-Premise, and Cloud is becoming a key requirement for any use case these days. This session will provide a brief introduction to Open Data Framework from SODA Foundation, which is a completely open-source solution and how it helps to address the key challenges of hybrid data management, cloud-native storage, and heterogeneous storage management. It will provide the overall architecture and key projects for different use cases of data management.
Object stores are known for ease of use and massive scalability. Unlike other storage solutions like file systems and block stores, object stores are capable of handling data growth without increase in complexity or developer intervention. Apache Hadoop Ozone is a highly scalable Object Store and is a spiritual successor of HDFS. It can store billions of keys and hundreds of petabytes of data. With the massive scale there is a requirement for it to have very high throughput while maintaining low latency.
This talk discusses the Ozone architecture and design decisions which were significant for achieving high throughput and low latency. With petabytes of data and billions of keys Ozone has a scalable metadata layer. The talk will detail how Ozone supports this layer without compromising throughput or latency. Such a massive scale requires Ozone to be scalable in terms of client connections and amount of data read and written to the store. The talk will discuss the challenges faced and the corresponding design solutions. Also it would touch upon Ozone’s goal of reaching trillion objects and possible challenges.
Here at Micron, we've been analyzing the effect NVIDIA GPU Direct Storage (GDS) can have on storage I/O in GPU enabled system. In this session we discuss why this technology is important, the performance improvements from using GDS, and an exciting new architecture for datacenter NVMe-Ethernet-attached Bunch Of Flash (EBOF) for NVMe Over Fabrics (NVMe-oF).
Marginal links and congestion have plagued storage fabrics for years and many independent solutions have been tried. The Fibre Channel industry has been keenly aware of this issue and, over the course of the last two years, has created the architectural foundation for a common ecosystem solution. Fabric Notifications employs a simple message system to provide registered participants with information about key events in the fabric that are used to automatically address link integrity and congestion issues. This new technology has been embraced by the Fibre Channel community and has demonstrated a significant improvement in addressing the nagging issues. In this informative session, storage experts will discuss the evolution of this technology and how it is a step toward a truly autonomous SAN.
Active archives enable media companies to quickly, efficiently and cost-effectively search their vast volumes of content to retrieve and process assets they need. Erasure coding can be used to protect this content, but it typically requires more resources and increases content retrieval latencies, limiting performance. This paper proposes a technology that offers higher content durability and availability with minimum latencies and less storage overhead than a two-copy scheme. INSIC Tape Roadmap forecasts 35% Areal Density and 25% Track Density CAGR, requiring future track pitches to be below 1000nm, such as a recently published 580TB tape technology demo requiring 50nm track pitch. At these dimensions, considering that tape must operate in open environmental conditions, drives and media failures will be dominated by correlated errors requiring more protection than an internal format erasure correcting code (ECC) and two-copy scheme can provide to achieve more than 11-nines durability.
Current replication techniques don’t provide resource-friendly protection and can’t meet 11-nines durability unless more than two copies are used. In a typical tape archival scenario, a carefully designed multi-dimensional erasure coding mechanism can bring novel efficient resource utilization, better availability and durability, and more environment-friendly operation for distributed tape systems with less overhead than two-copy protection. To reduce repeated read cycles and I/O bandwidth requirements, increase durability, and improve local recoverability, we propose multi-dimensional, hierarchical overlay erasure coding with interleaved codewords. Its unique mathematical coding structure and special interleaving allows local repair and almost uncorrelated errors, helping maximize local tape operations - i.e. minimize the system’s tape, drive and robot requirements.
Thanks to special metadata handling, the proposed system’s byproducts are the self-describing erasure-coded tapes friendly to disaster recovery and content transportation. These combined contributions help increase the lifetime of different subsystems of tape and improve the archival storage system’s performance.
The vast spread of computational storage use cases continues to raise. NGD Systems, a leader in computational storage drive will go over the journey of computational storage usage cases, starting from the obvious, the cloud, and heading into new expending territories of edge compute, low earth orbits small satellites to crypto currency plotting. All have once common thread of how can computational storage increases efficiency by reducing data movement.
Challenges facing system designers, mother board designers are: Space, Performance and Cost. Increasing the data-rate of electrical interfaces on PCB, increases cost of manufacturing and add to other engineering challenges. Dedicated pins on the CPU sockets for DDR like interfaces are limited.
EDSFF standard SFF-TA-1002, 1009 standardize the electrical and mechanical outline of devices. This ensures same server chassis can support combination of profiles from different vendors. Individual form-factor spec like 1006 and 1008 define power, thermal budget, and airflow conditions enforcing commonality between modules from various vendors.
Compute Express Link™ (CXL™) is a high-speed CPU-to-Device and CPU-to-Memory interconnect designed to accelerate next-generation data center performance. CXL is an open industry-standard interconnect offering coherency and memory semantics using high-bandwidth, low-latency connectivity between host processor and devices such as accelerators, memory buffers, and smart I/O devices. CXL technology is designed to address the growing needs of high-performance computational workloads by supporting heterogeneous processing and memory systems for applications in Artificial Intelligence, Machine Learning, communication systems, and high-performance computing.
Released in November 2020, the CXL 2.0 Specification adds new features – including support for switching, memory pooling for increased memory utilization efficiency, and persistent memory – all while maintaining full backwards compatibility with CXL 1.1 and 1.0. CXL 2.0 provides standardized management of the persistent memory interface and enables simultaneous operation alongside DDR, freeing up DDR for other uses. In this presentation, attendees will learn how CXL 2.0 supports persistent memory, CXL use cases, and what’s ahead for the Consortium.
In this talk, we’ll cover the latest updates of the Microsoft Protocol Test Suites for File Services. Microsoft Protocol Test Suites are a group of tools that were originally developed for in-house testing of the Microsoft Open Specifications. Microsoft Protocol Test Suites have been used extensively during Plugfests and Interoperability (IO) Labs to test against partner implementations.
Learning Objectives: How do the Test Suites work and what scenarios do they cover? How have the Test Suites changed over the past year? What new testing scenarios are covered for File Services?
The Linux SMB3 client continues to be the most active network/cluster filesystem on Linux over the past year, and the progress on Samba server and the Linux kernel server has helped add new features to the SMB3.1.1 client in Linux, and made these improvements even more important.
It has been a great year for SMB with the addition of many security improvements, many performance improvements including to caching and RDMA (smbdirect) as well as dramatic improvements to multichannel, and new sparse file features. Support for the Witness protocol (allowing transparent movement to a different server) has been added, as well as the new more feature rich Linux mount API. In addition support for the final piece of the optional SMB 3.1.1 POSIX protocol extensions was completed. Tooling has been improved with many new features added to tools like smbinfo, and support for easily getting and setting more auditing and security information.
This presentation will go through some of the new features added to the Linux client over the past year, and demonstrate the great progress in access various types of network storage, including the cloud (e.g. Azure), and Samba and Windows server and also the new Linux kernel server (ksmbd).
ksmbd(cifsd) is a new SMB3 kernel server which implements server-side SMB3 protocol. The target is to provide optimized performance, GPLv2 SMB server, better lease handling (distributed caching). The bigger goal is to add new features more rapidly which are easier to develop on a smaller, more tightly optimized kernel server. Many changes and improvements have been made since cifsd(ksmbd) was introduced to earlier SDC 2019. This talk will give ksmbd overview and the current status update.
The SMB protocol was designed long after Unix was created, and as a result supported concepts like globally unique identities and rich ACLs that are in Windows, but not in Linux. User identity and access control are very relevant to the Linux SMB3 client, as it acts as a bridge between the world of Windows-like-filesystems (including the cloud) and the world of Linux filesystems, and has the hard task of translating security information from the more complex Samba and Windows world, to the simpler Linux/POSIX model.
There are three key problems:
Id-mapping: Who the user is? And how does it map to the user that the server understands?
Authentication: Can the user prove his/her identity?
Access control: What permissions does the user have for this file?
This talk will discuss and demonstrate the different ways that the Linux client can be configured to map POSIX permissions (mode bits) to ACLs, and the implications of using these configurations. It will discuss the different authentication choices, especially how to leverage Samba’s winbind for easy to use and highly secure Kerberos authentication and key refresh. In addition it will discuss how to integrate with Samba’s winbind to map user identities (from the local Linux client’s UIDs to globally unique SIDs) and the various alternatives like “idsfromsid”. Recent improvements in cifs-utils for managing ACLs and auditing information remotely will also be discussed, which can make managing Samba server easier in some cases.
Under the raised floor tiles of data centers around the world you will find tangles of multi-colored cables and strange, other-worldly interconnects. The swill from the overflowing bitbuckets seeps its way down to this level, where even DevOps fear to venture. In the darkness of these digital depths, strange new creatures evolve. The new new breed of SmartNICs is an example. SmartNICs are similar to the familiar TOE and iSCSI cards, but the latest wave of evolution are based upon multi-core RISC-based I/O processors known as Data Processing Units, or DPUs.
This talk will cover a little project that started up just about a year ago. The Zambezi project aims to create an SMB3 Offload Engine, aimed at DPUs and SmartNICs. We will look at SMB3 message processing from the Syntactic to the Semantic layers and discuss how this processing can be broken down into layers so that much (if not all) of it can be offloaded to (general case) SmartNIC or other network infrastructure device.