Building Scalable Data Center Networks

Editor's Notes

• #2: Thank you Vashuha. I’m Diane Patton and with me is Ravi Nittur from Dell/EMC. Today we are going to cover the benefits of using eBGP in a Data Center, along with deployment scenerios and the setup and results of a real live POC we did with RedHat using Openstack. By the end of the webinar, we are hoping you will learn why eBGP is a RP of choice in the data center, how to deploy it, along with one possible use case.
• #3: Many data centers today are Layer 2 networks. Generally they have servers that are dual homed to 2 top of rack switches, which are also often called leaf switches. To provide redundancy and increase the BW, MLAG is often used and spanning tree is deployed. Now, mLAG is a propiretary protocol but it does allow us to be able to utilitze both of these ToR switches by fooling spanning tree into thinking there is only one ToR, thereby SPT won’t block one of these links. However, in order for this to work, the 2 ToR must be connected together via peer-links, which utiltze additional ports. For connectivity between racks, a spine layer exists and the same conecpt is used there too. To provide mobility, VLANS are often used throughout the data center and this increases the failure and broadcast domains. This design does not scale well and there can be issues with it. Not as stable either. Large failure domain, no standards based. Troubleshooting – trace route
• #4: 3
• #5: Now, which RP to use? We could use link-state protocols like OSPF or ISIS. Link-state IGP implements adjacency information, maintenance and flow control .. On the other hand event propagation scope of link state is entire area, regardless of failure type. BGP just relies on TCP and underlaying transport and the flooding overhead is less BGP flooding overhead is less Greacefully direct traffic off a switch when doing an upgrade.
• #6: Mention showing 3 spines to show that routing and not MLAG is being used. Same AS on one spine helps with route convergence, makes convergence faster. However, the spines cannot talk with each other. The reason you cannot use this is if the spines need to communicate with each other, like in the case if you are running multicast along with MSDP for RP Anycast, or in the case of Cumulus LNV. If you put spines in different Ass can get around convergence issues by implementing a route policy to only announce locally originated routes on leafs and spine loopback (BGP speakers will drop if sees same AS in path) Each ToR in different AS Each as for troubleshooting – can watch AS path One AS for spine to reduce path hunting Using private as 1023 AS’s unless use 4 byte Ass About 95K AS private AS numbers S ad then strip at the edge remote-private-AS
• #7: Rememeber to mention allow-as 1 as config for this option. Mention that I show 3 to show it’s not pairs, like you would with MLAG Mention how Pod A is what we showed on the prior slide. Can easily grow by adding Super spine layer to the setup. AS numbers remain the same on Pod A – so easy scale. RFC 6793 – BGP Support for 4 Octet AS numbers
• #8: Reason each host has it’s own AS is because this is what we did for the Openstack Redhat/Dell/Cumulus trial. Will mention this is not necessary. Routing all the way to the host eliminates SPT, MLAG completely, ECMP used up to the leafs, adds mobility back. 4 octet ASNs if needed Advantage to move VTEP directly on the host, as we will talk about later during the deployment section If both ToR in different AS, need bgp bestpath as-path multipath-relax command Bgp always compare med No SPT – no MLAG, ECMP – no max number ToR switches Eliminates peer-link ports on leaf swithes One more port
• #9: The downside of any clos topology is that there are many switches to manage, however if automation is used then it’s the same for 10 swithe
• #10: Mention the title of RFC 5549 – what is it for, how it works, Menotine you don’t need ipv6 routing tfor this to work Advertising IPv4 Network Layer Reachability Information with an IPv6 Next Hop - don’t need to configure IPv6 address because it uses the automatic link-local address Config in VTYSH Uses IPv6 Router Advertisement to learn neighbors lin local address Reduces FIB size Reduces attack vector since only single reachable address As opposed to as many addresses as there are links See RFC 7404 for more details IETF, other network vendors also advocating the use of LLA: https://blog.apnic.net/2016/02/16/change-of-paradigm-with-ipv6-no-global-addresses-on-router-interfaces/
• #11: Use of ipv4 link local address to make up next hop
• #13: If you choose to not use automation, we will be coming out with a new utitily, NCLU. This adds a easy to use command line to cumullus Linux. You never leave bash, so it interoperates with the traditinoal way of doing things with Linux, but it’s nice becaue guardrails are inlcude, meaning if you make a typo you are notified and even a suggestion is made. Examples are included within, for example if you type net example bgp unnum – you will see an entier setup and config commands needed to achieve it.
• #14: Rsyslog sends to server SNMP Kernel log messages Watchquagga watches the quagga daemons. -f follow –u unit (follow quagga) In this mode, whenever a single daemon hangs or crashes, the given command is used to restart this daemon only. The only exception is the zebra daemon; in this case, the following steps are taken: (1) all other daemons are stopped, (2) zebra is restarted, and (3) other daemons are started again. Example usage: watchquagga -adz -r '/sbin/service %s restart' \ -s '/sbin/service %s start' \ -k '/sbin/service %s stop' zebra ospfd bgpd
• #15: Say which tier was this command Mention hostnames are used. Etc.
• #16: Make sure font is consolas Similar to a term mon. In this case, I purposely configured the wrong AS – configured internal when the AS number was external. Allows you to see as things are added to the log file.
• #18: Describe how VXLAN works Run VTEP on the host itself – no changes on underlay – this is what we did for the openstack trial.
• #19: Route-map set-src permit 10 set src X Ip prototocl bgp
• #20: Virtualization has revolutionized data centers, facilitating extraordinary gains in efficiency and ROI. A similar transformation is underway in networking. Break free from the proprietary restrictions of single vendor network platforms, with switches based on open standards. By adopting a Dell Open Networking platform, you can choose an operating system (OS) that’s best suited for your needs. Gaining this level of network control and flexibility is a requirement for software-defined networking (SDN) and an important step toward realizing the ultimate agility a software-defined data center delivers. You’ll notice that the hardware platforms for traditional and Open-Networking are the same. You choose the hardware that fits your requirements, whether its 1gb or 100gb, or anywhere in between with our multi-rate platforms and choose the software that best fits your requirements.
• #27: Openstack Platform Director is based on project TripleO that provides toolset for installing and managing a complete OpenStack environment. Rally is a benchmarking and profiling openstack tool used for – checking how openstack works at scale/stress Browbeat is a performance tuning and analysis tool for openstack [ open source/free]. Analyze and tune cloud for optimal performance, create Rally workloads for performance and scale testing.