Software Quality can mean a lot of different things to different people. Most developers equate quality to testing which is very reasonable. But in the SPDK Community it’s about much more than just testing, it’s about automation, code reviews, and clear communication to name just a few. This talk will provide a bird’s eye view of the SPDK automated testing infrastructure and how it’s evolved over the years. We will also demonstrate how developer interactions are another key to SPDK’s continued commitment to quality.

Driving efficiency and performance is critical to modern data center architectures, this talk will cover how the features of SPDK (lockless design, user space lib, core affinity, bdev stacking, and debug ability) were used to provide high performance nondurable block storage with RAID 0, think provisioning QOS, Clones, Snapshots and Redfish/RSD compliant management.

SPDK was originally designed to run in a model where all memory was allocated up front, a system thread was spawned per CPU core, and the process was executed with elevated privileges. While suitable for a large number of use cases, a more flexible and dynamic model is required especially for use within cloud deployments. Today, SPDK supports fully dynamic memory allocation, a new flexible threading model design to integrate seamlessly into existing code bases, and full support for running without elevated privileges. This talk will cover that transformation process and outline how to take advantage of all of SPDK’s new-found flexibility in your application.

A SmartNIC is a NIC with on-board processing units, that can run software. Such a SmartNIC can present to the host a standard NVMe device, while implementing the device logic, using software on the SmartNIC's processors. Use cases include: bare-metal and virtual cloud providers; storage systems with proprietary protocols; storage devices security and isolation; compression and encryption over the standard NVMe interface and more. SPDK is a great match as a framework for the software processing on the SmartNIC.

NVME-OF target offload is a hardware implementation of the NVME-OF RDMA protocol at the network card. Employing this technology, the network adapter terminates the target side of NVME-OF protocol and submits the NVME requests directly to NVME devices over PCI using peer-to-peer. Integrating this technology with SPDK will allow users to manage such configurations from a known and flexible environment.

Intel is releasing an Open framework for their Cache Acceleration Software product (Open Cache Framework OCF). OCF will be maintained by the Intel team and available to all members of the community. This package can be wrapped in other storage applications to improve caching effectiveness and improve overall platform performance. Current examples from Intel include integration into SPDK and QEMU. The details of this OCF package, how it can be used, and potential benefits will be reviewed.

Dealing with data corruption is important for storage applications, including those built on SPDK. This presentation will describe the new SPDK DIF library and its use cases in SPDK. The library leverages the Intel® Intelligent Storage Acceleration Library (ISA-L) for efficient CRC calculations and provides routines to simplify integration of DIF operations into other libraries. The SPDK iSCSI target now leverages this DIF library to support insert/strip for read and write I/O. This demonstrates that end-to-end data protection can be implemented in host software without specialized hardware while still providing compelling performance and efficiency.

The Flash Translation Layer library provides block device access on top of non-block SSDs implementing Open Channel interface. Such SSDs provide more flexibility with regard to data placement decisions, over-provisioning, I/O operations scheduling, garbage collection and wear leveling. FTL library handles logical to physical address mapping, responds to the asynchronous media management events and manages the defragmentation process. Presentation will focus on explaining library core concepts and components, how the library should be used and what benefits could be obtained. Future plans to support power loss protected 4K writes and Zoned Namespace interface will be covered.

Solarflare has a mature and proven user space TCP/IP stack that has been integrated into every capital market, powering exchanges for over a decade. We recently ported this stack, called Onload®, onto SPDK. Solarflare has spent the past 24 months testing and proving the value of NVMe over TCP. Solarflare will present our work and roadmap going forward while exploring ways to work closer with the  SPDK ecosystem.

The goal of integrating Oracle database with SPDK is to build highly available and scalable applications where the database is serving millions of IOPS at very low latencies. However, there are a few challenges with integrating SPDK with the Oracle database. The multi-process Oracle database architecture quickly exhausts the PCIe hardware queues. The DPDK memory model conflicts with Oracle's shared memory model and introduces inefficiencies in the RDMA data access path.

In this talk, we will go over the Oracle dispatcher and Oracle environment abstraction layer (EAL) design and implementation. Additionally, we will also describe other enhancements that enhance the security model for SPDK memory management, provide better isolation between different databases and improve the fault tolerance of the database. These features help address the above challenges and enable creation of an enterprise-grade solution that can unlock the full potential of the NVMe devices. Lastly, we will talk about some of the deployment and packaging related challenges with delivering this solution for on premise and cloud deployments.

Since its release in 2016, The SPDK NVMe-oF target has leveraged innovative features to provide low latency, high throughput access to remote NVMe drives. This last year has provided many opportunities to expand on that foundation to broaden the applicability of the NVMe-oF target, further increase performance, and decrease it’s memory footprint. In this talk, we introduce the NVMe-oF TCP target and initiator, describe their position in the SPDK NVMe-oF architecture and provide benchmarking data for the new transport. We also detail several new features of the RDMA transport including shared receive queue and scatter gather list support and their impact on performance and memory consumption.

Since the SPDK community introduced OCSSD FTL bdev in January 2019, Circuit Blvd and Marvell have been jointly evaluating it on Marvell's Zao SSD SoC platform. We will first present Zao platform technical details including OCSSD SDK features. We measured various performance metrics from the initial prototype in comparison with the standard NVMe devices. In conclusion, we will demonstrate how SPDK OCSSDs can solve the noisy neighbor problem in multi-tenant environments.

Our two-years practice to develop an Enterprise AFA storage with high-speed RDMA reveals that its complicated storage stack requires a threading environment to reduce implementation complexity. However, it is a big challenge to implement a threading environment including synchronous communication primitives without degrading the performance of SPDK/DPDK.

RMI threading environment, we have developed on SPDK 16.12, supports over 16K threads, and provides dynamic load-balancing and simple synchronous communication primitives using RDMA. Our preliminary evaluation show that the cost of load balancing is quite low, and the communication performance scales well on modern multicore systems.

NVMe-over-Fabrics (NVMe-oF) is an architecture that allows low latency access over RDMA and TCP to remote NVMe-attached storage. It is fast becoming a key enabler for disaggregated storage and NVMe pooling in data centers, driving increased storage efficiency. Intel, Mellanox and Mirantis joined forces to enable NVMe-oF in Openstack Cinder and Nova - a new SPDK NVMe-oF driver was recently merged to upstream, which paves the way for consuming SPDK-based high performance storage in OpenStack (and soon to be enabled, Kubernetes) cloud deployments.

Last year we introduced the crypto bdev module which made use of DPDK’s existing variety of drivers to usher in the capability. This year we are expanding our use of DPDK with the addition of a compression bdev module. This talk will outline the overall architecture of the compression module and explain in detail how we are managing the layout of the device and leveraging PMDK to store metadata in super-fast persistent memory.