Knowledge

RoCE moves RDMA to Ethernet, which is extremely sensitive to packet loss. Based on the three steps of lossless premise, network card and speed selection, and online verification, this article explains how to configure PFC and ECN, how to select 100G network card and DAC, and how to test perftest.

The shared link between the iSCSI storage network and the business network is prone to slowdowns and disconnections. This article gives the construction key points in three steps: inferring bandwidth according to business load, isolating VLAN and dual switches, and leaving a margin for optical path loss, which can be used as a reference for small and medium-sized cluster selection.

The most difficult thing about upgrading a NAS to 10G is whether to choose the electrical port or the SFP+ optical port. This article gives a three-step judgment method based on transmission distance, power consumption and existing switch ports, and explains the configuration and link key points of the electrical port, DAC and optical module solutions respectively.

SR-IOV allows virtual machines to bypass virtual switches and directly access physical network cards, reducing CPU overhead and forwarding delays. This article summarizes the key points of implementation in VMware and PVE environments according to the three steps of enabling check, network card selection and pass-through selection.

Sort out the location and shutdown methods of network card energy-saving items under Windows and Linux, explain the benefits and side effects of offloading functions such as TSO, LRO, and GRO, and provide a baseline configuration list for new card installations.

Starting from the typical phenomenon of single-core soft interrupts being full, a complete tuning path is given to use /proc/interrupts and mpstat to locate bottlenecks, set the number of queues, and configure interrupt affinity and aggregation parameters.

It explains the profit boundary of MTU 9000 in storage and migration scenarios, emphasizes that all path devices must be aligned at the same time, and provides a two-level verification method of prohibiting fragmentation, ping and throughput comparison, to avoid packet loss caused by missing configurations in the middle.

Compare the applicable scenarios and switch side requirements of mode1 active/standby and mode4 LACP, and provide the configuration sequence, verification method and fiber pull-out switching test steps for dual-port optical card aggregation to avoid switching failure in the event of a real fault.

Use lspci to check the chip, ethtool to check the driver, and dmesg to check the link. The correspondence between ixgbe, i40e, ice and mainstream Intel network cards will be explained clearly, and suggestions for persistence of network configurations for each distribution will be given.

It explains clearly the selection principles and installation process of the fiber optic network card driver on Windows Server, as well as the three check items that must be checked after installation: energy-saving options, advanced properties and link speed. It also provides methods to clean up the old driver when replacing a new card.