Let MindShare Bring "NVM Express over Fabrics (NVMe-oF)" to Life for You
MindShare's NVMe-oF (Non-Volatile Memory Express – over Fabrics) course begins with a brief review of NVMe basics, discusses the forces driving the migration of NVMe into network fabrics, followed by an overview of the different fabrics NVMe-oF will work over such as Ethernet, InfiniBand and Fibre Channel. Finally, the details of NVMe over fabrics are described.
MindShare Courses Related to NVMe:
All of MindShare's classroom and virtual classroom courses can be customized to fit the needs of your group.
NVM Express over Fabrics (NVMe-oF) Architecture Course Info
You Will Learn:
- An overview of the storage fabrics in use today, and their strengths and weaknesses
- The changes needed for NVMe to work in a fabric environment
- How an NVMe host discovers and accesses drives on the fabric
- How NVMe-oF provides a combination of attractive features for a storage fabric
- How NVMe protocol will work over Ethernet, InfiniBand and Fibre Channel
- An overview of Ethernet including RoCE and iWARP, InfiniBand and Fibre Channel Protocols
Who Should Attend?
This course is hardware-oriented, but is suitable for both hardware and software engineers because the registers used to control the hardware are described in detail. The course is ideal for RTL-, chip-, system- or system board-level design engineers who need a broad understanding of NVMe-oF.
Course Length: 1 Day
This course is hardware-oriented, but is suitable for both hardware and software engineers because the registers used to control the hardware are described in detail. The course is ideal for RTL-, chip-, system- or system board-level design engineers who need a broad understanding of NVMe-oF.
Course Outline:
- NVMe Background
- NVMe basics: registers, queue management, commands and completions
- HCI shared-system-memory model
- NVMe over PCIe
- Why move NVMe to fabrics?
- Access to storage networks of thousands of SSDs
- Avoid protocol translation to get to those networks
- Optimize access latency across the network
- Allow systems to work using only remote storage (no local storage)
- Improving network access with RDMA (Remote DMA)
- Kernel bypass – reduce CPU involvement
- Zero-copy operation
- Reliable, lossless delivery
- Credit-based flow control
- Use of common primitives
- Pinned remote memory and storage
- Encapsulation avoids need for protocol translation
- Ethernet overview – popular, high speed, inexpensive
- TCP/IP: addressing, layers, packets
- Advantages and disadvantages compared to other fabrics
- RDMA not natively supported by Ethernet, so two competing models were developed to add support for it: RoCE and iWARP. Both require RDMA-capable NICs (RNICs).
- Overview of RoCE
- Version 2 packets include routing information, version 1 did not
- Use of Data Center Bridging (DCB) to manage routing
- Flow control
- Soft RoCE
- Overview of iWARP
- Added to TCP/IP, so packets are routable and don’t need DCB
- Messages
- Layers
- InfiniBand overview
- Designed for RDMA
- Messages
- Discovery
- Packets
- Routers
- Physical layer encoding schemes
- Changes to support NVMe
- Fibre Channel overview
- Addressing
- Initialization
- Discovery
- Packets
- Physical layer encoding schemes
- Changes to support NVMe over FC (NVMe-oFC)
- NVMe over Fabrics (NVMe-oF) Details
- Transport mapping/binding: Fibre Channel or RDMA
- Discovery Service
- Connections
- Properties
- Authentication
- Capsules
- Queue pairs
- Data transfer: Command, Response, Data, Flow Control
Recommended Prerequisites:
NVMe architecture knowledge and previous exposure to PCIe is needed, as is some general knowledge of PC architectures.
Supplied Materials:
- Downloadable PDF version of the presentation slides.
- Optional: Comprehensive NVMe 1.2a eLearning course
|
|