Chris Marget sent me the following interesting observation:
One of the things we learned back at the beginning of Ethernet is no longer true: hardware filtering of incoming Ethernet frames by the NICs in Ethernet hosts is gone. VMware runs its NICs in promiscuous mode. The fact that this Networking 101 level detail is no longer true kind of blows my mind.
Ethernet was designed as a shared media (remember the thick coax cables with vampire transceivers?) and even though switched Ethernet (aka bridging) gave us more bandwidth, it still emulates the coax cable – frames sent to broadcast, multicast and unknown unicast addresses are flooded to all hosts (the behavior that makes brokenware like Microsoft NLB work). There is no communication between Ethernet hosts and bridges (remember: bridges emulate a single cable and a cable cannot talk to stations attached to it) and thus the switches have no way of knowing whether those frames are important to the end hosts or not (IGMP snooping is one of those kludges that is supposed to make a broken design a bit less dreadful).
In the shared media environment of early Ethernet it was very important that the frames not meant for an individual end-station do not burden its CPU. Frame filtering based on destination MAC address was thus always implemented in hardware (the same is true for almost all multi-access L2 technologies). An Ethernet NIC was always able to listen to one (or a few) unicast MAC addresses and a few multicast MAC addresses (the limits are vendor-dependent); everyone would obviously have to process broadcast frames.
A typical end-host has a single MAC address. Some hosts running DECnet Phase IV would have two (DECnet Phase IV had nothing like ARP or ND and changed MAC address to match its L3 address) as would clustered hosts sharing a single IP address (for example, Windows servers running Microsoft NLB or even Cisco routers running HSRP or GLBP).
The server virtualization has completely changed the Ethernet NIC requirements – with tens of VMs running within the same physical host, its Ethernet NIC has to be able to receive frames for tens (or sometimes even more than a hundred) unicast MAC addresses. Most Ethernet NICs are not able to do it (if you know more, please write a comment; it would be interesting to learn whether VN-Link/VM-FEX improves things) and thus most hypervisor operating systems put the Ethernet NICs in promiscuous mode. Every single frame sent to the NIC (those sent to the VMs running in the host as well as all floods) has to be processed by the host CPU, regardless of whether it’s relevant or not.
The use of promiscuous mode has made hypervisor hosts slightly more vulnerable to flooding storms. In the pre-virtualization days, the links in the network could become overloaded after a forwarding loop, but the Ethernet NICs would do most of the damage control (unless, of course, you experience a broadcast storm, in which case nothing can save you). In a network having primarily virtualized servers running in hypervisors, the host CPU has to deal with all the looped frames. Don’t ask me whether this fact is relevant for your network or not – you might decide you have bigger problems than CPU overload if you experience a forwarding loop, or you might feel that this fact alone should push you toward smaller subnets (and thus broadcast domains).
If you’re interested in details of VMware networking, check my VMware Networking Deep Dive webinar (register here or buy a recording). If you want to learn more about modern data center architectures, buy a recording of my Data Center 3.0 for Networking Engineers webinar and check the Data Center webinar roadmap. Both webinars are also part of the yearly subscription package.