之前使用过chatgpt和claude协助写过一些测试脚本,比如早期的:RDMA测试脚本 by ChatGPT?!,后来优化的就没更新在blog中了。以至于后来研发的同事帮忙优化的测试脚本,虽然目前可以达到一定需求,但python精度太差,都是ms级精度控制,效率不是很高。因此需要一套高性能RDMA测试脚本,满足不同场景下针对交换机的测试需求。目前看上去使用openmpi来控制perftest的思路比较合理,而且有的用户也在使用类似的方式,所以就尝试走这条路。以下是五一期间跟claude合作开发的测试系统,里面很多细节还有待完善中。
I summaried the hands-on at 2019.7, public on my blog now. So you will find the version is old (652/661), new version(721/731) include more inovation features in SR solution.
Introduction
I summarized SR/SRTE hands-on-deck that includes our highlight features and show information in the same topology, content includes the following topic, now revised version is v3:
There are two old great features that can achieve Active/Active in multi-chassis:
mLACP: ASR9k support, and by ICCP.
NV Cluster: only ASR9k support, not need other protocol (e.g ICCP), but not support from 6.x.
So could we have the similar solution on NCS55xx? The answer is “Yes”, we can deploy EVPN + Anycast IRB to achieve the similar function. For this article, no reference to real EVPN environment, only use multi-home A/A EVPN function. And include testing when link failure.
Conclusion
If you deploy EVPN + Any IRB, you must to know there are no dynamic IGP protocol between CE/mPE or Host/mPE, only Static or default GW. That due to we couldn’t deploy the same IGP process at two chassis. That should different with NV Cluster. As follow:
Due to deploy EVPN that belong to BGP, so we need underlay protocol that can be achieved by LDP or Segment Routing. In this example, will deploy Segment Routing.
Traffic(via EVPN) will not forward to the link between two PEs, so no need to deploy inter-link. Except customer has other L3 business in the same PEs.