Method for optimizing performance of high-performance network card under Feiteng platform based on NUMA architecture

Abstract

The invention discloses a method for optimizing the performance of a high-performance network card under a Feiteng platform based on an NUMA architecture. The method comprises the following steps of: processing a memory node application of network interruption; expanding a data transceiving queue of the network interruption; and eliminating the network interruption reversal. According to the method for optimizing the performance of the high-performance network card under the Feiteng platform based on the NUMA architecture, the performance of the high-performance network card is improved through network card interruption affinity under the numa architecture, and meanwhile, the performance of the high-performance network card under the Feiteng platform based on the NUMA architecture is further optimized in a more targeted manner; by analyzing the numa architecture characteristics of the Feiteng platform, the performance of the high-performance network card under the Feiteng platform based on the numa architecture is improved in a more targeted manner; and the limitation of a network card driver on the number of data receiving and transmitting queues and the number of interrupts is broken through, the problems of extra routing consumption and interrupt reversal possibly occurring on a processor platform of a multi-channel over 64-core CPU are avoided, and the network I / O performance is improved.

Description

technical field [0001] The invention belongs to the technical field of Feiteng platform, and in particular relates to a method for optimizing the performance of a high-performance network card under the Feiteng platform based on NUMA architecture. Background technique [0002] With the further rapid development of network technology, in order to meet the needs of applications, high-performance network cards with higher bandwidth such as 25G, 40G, and 100G have appeared and are being used more and more. With the improvement of network bandwidth, high-performance network cards are increasingly squeezing the CPU. However, the original scheduling method of the Linux system and the way single-core CPU handles network card interrupts can no longer meet the demand. For this reason, the NUMA architecture and multi-queue network card technology are widely used in the I / O requests of high-performance network cards. This type of technology uses the characteristics of the NUMA architect...

AI Technical Summary

Problems solved by technology

Regardless of whether the network card is inserted into other numa nodes, or the network I / O request is processed at a node far away from numa node0, it will generate interrupt routing overhead and affect the performance of high-performance network cards;

[0005] 2) On the FT2500 processor platform with a total of 128 cpu cores in the main and slave roads, some high-performance network card driver logic problems may lead to poor performance of the high-performance network card: Compatible with some can only apply for 64 interrupts The total number of data sending and receiving queues registered by some high-performance network card drivers does not exceed 60, which makes it impossible for the FT2500 processor platform to obtain data sending and receiving queues from the rear 64-core CPU on the road. Network card, the I / O request of the slave network card can only be routed to the master CPU for processing, instead of being directly processed by the slave CPU, resulting in additional routing consumption;

[0006] 3) Because the total number of msix interrupts applied by some high-performance network card drivers cannot exceed 64 (to be compatible with some network cards that can only apply for 64 interrupts), and the network card driver selects the CPU for interrupt processing by extracting the cpu id from the interrupt information For the reason of the FT2500 platform, when the master and slave channels of the FT2500 platform are both connected to the network card, there will be a phenomenon that the master and slave channels are interrupted and reversed: when the network card is interrupted on the slave channel, the cpu id obtained through 64 interrupts can only be the one on the main channel. For the first 64 CPUs, because the interrupt types of the main road and the slave road are the same, this will cause the interrupt of the slave road to preempt the main road cpu (number cpu0~cpu63), and the interrupt of the main road network card is forced to reverse to the slave road. cpu (number cpu64~cpu127) to process

[0008] At present, none of the existing technologies for improving the performance of high-performance network cards in Linux systems is specifically aimed at optimizing the performance of high-performance network cards for domestic Phytium processor platforms.

Images