Abstract
Infrastructure Processing Units (IPU) is an evolution of SmartNIC, focusing on infrastructure processing such as networking offload and storage offload. IPU is a critical ingredient in the disaggregated computer architectures and becomes a control point in the DC-oF (Data Center of the Future). In this talk, we would like to share the NVMe over TCP Initiator implementation as an example for IPU based storage offload, focusing on SPDK (https://spdk.io) support for IPU based NVMe over TCP Initiator solution.
First, we will introduce our high performance 200GbE IPU, then we will present the software components composed with SPDK based NVMe over TCP software stack. Second, we will share the performance optimizations in SPDK from last year which leverages on Intel® Ethernet 800 Series with Application Device Queues (ADQ) to improve SPDK based NVMe/TCP initiator performance. Third, we will share some performance results from Intel® Ethernet 800 Series with Application Device Queues (ADQ), which demonstrates ADQ significantly improves the performance (increase the IOPS and reduces the longtail latency) for the SPDK based NVMe over TCP initiator.
- Share the hardware and software architecture of Intel's high performance 200GbE Smart NIC/IPU with the audience.
- Share how SPDK I/O stack is integrated into the SmartNIC/IPU and leverage on Intel ADQ technology in the Intel SmartNIC/IPU to a
- Share the detailed techniques from using SPDK and ADQ technology to improve the performance for the NVMe-oF TCP initiator in the
- Discuss the possible future optimizations of TCP transport to improve storage offloading solutions in SmartNIC/IPU.
- Share the ADQ performance improvement for the NVMe/TCP initiator (e.g., increased IOPS, reduced latency) from Intel 100GbE NIC (