At work, we had a problem where a computer was spamming the network with DHCP requests. It ended up physically locking us out; the embedded system which controls the card reader and unlocks the door locked up due to the load. I had to use an alternative way to get in, isolated the maglock, and later power-cycled the embedded system.
Needless to say, we enabled storm control on the switch after that.
We didn't even have VLANs enabled on the switch back then.
And even then, the storm control is useful; the same broadcast storm also made the office wireless unusable (most APs use a basic rate for broadcast and multicast).
I feel like this couldn't happen soon enough. All those mysterious networked AV devices (why does my TV want to be on ethernet anyway? It's not even a smart TV) and their mysterious poorly written firmware.
I'll point out that the original vmstat output also revealed 3000 interrupts per second, in the column right next to syscalls. Many times it can be faster to look around at the high level instead of immediately diving deep. In this case, checking to see which device was causing the interrupt storm would have led to "something on the network" pretty quickly.
Is mDNS really so useful? Of all the people I know who know about it, everyone has disabled it when possible due to all the useless network traffic it produces.
Apparently Canon's wireless printers are well known for doing mDNS floods:
mDNS has utility in small networks for your average Joe. In these circumstances (1) it shouldn't be a problem, (2) it shouldn't flood (there's logarithmic decay and there actually has been thought put into this), (3) names are discoverable without relying on anything else, (4) and it does its job in being the mechanism for DNS-SD to piggyback on so that services can be discovered automatically. Being independent of other services working or not — e.g., DHCP, (unicast) DNS — means "it just works", even if other network services fail.
There is a beauty in the ability to just have all of your computers, printers, and peripheral devices otherwise plug into a switch (without any additional network service equipment — no router/dns/dhcp multi-combo device added in the mix) and that everything just ends up working. Link-local addressing takes care of IPs, mDNS takes care of names, and DNS-SD through mDNS takes care of automatic service advertisement — "Hey, I'm a printer, use me if ya wanna!"
Further, failure of proper mDNS implementation in a device (perhaps resulting in malformed packets or over-zealous flooding) is just a failure of that device's design, not mDNS per se.
Having said that, still... yes, I think the home/SOHO router/switch/access-point/dns/dhcp multi-combo device should handle DNS-SD registrations and DNS, and that mDNS shouldn't be needed.
In medium/large networks (those actually managed by someone putting on an IT hat), it's perhaps not as appropriate, but still shouldn't be a problem assuming everything is subnetted with mDNS in mind. Of course, name resolution is handled by unicast DNS, and with regard to service discovery in this context, someone manages the DNS-SD records on the unicast DNS server so you don't need mDNS to be the mule for DNS-SD to ride on in those circumstances.
>someone manages the DNS-SD records on the unicast DNS server so you don't need mDNS to be the mule for DNS-SD to ride on in those circumstances.
Are you talking about TXT records for services or something like that? I thought that most of the consumer level tools (for adding printers) were only looking for records via mDNS.
That said, in $CURRENTJOB, network printers are found via whatever the current version of NETBIOS is called - I don't think I've seen any mDNS traffic around.
For a simple DNS-SD setup on unicast DNS, yes, it's basically a combination of PTR and TXT records. See http://www.dns-sd.org/serversetup.html for more information.
For example, in one network, multicast was completely disabled — no mDNS happening there — and so on our local (unicast) DNS server (BIND on a 'BSD), we added DNS-SD entries (PTR + TXT records) for printers (to a CUPS print server with Active Directory authentication) which gave all the iOS devices AirPrint capability to all the printers on the network (even AirPrint to old dot-matrix printers), entries for Apple TV/AirPlay, etc. Apple calls it "Wide-Area Bonjour" and it works with Macs and iOS devices just fine — it's Bonjour with no mDNS involved (more specifically, it's just DNS-SD with a unicast DNS server).
In a sense (and completely avoiding all of the details and particulars of the matter here, because that opens up a whole can of worms and a much longer discussion), mDNS is basically supposed to be the modern IETF Internet Standards branded replacement for the name service aspect of NetBIOS (and other local network name service protocols). See also http://lwn.net/Articles/252323/ ... Also, in a sense, Zeroconf (an umbrella term which includes IPv4 link local addressing + mDNS + DNS-SD) is replacement for SLP ... and SSDP & UPnP in a sense too (although they're not directly comparable concepts)... Read http://www.zeroconf.org/zeroconfandupnp.html for a better explanation on the UPnP side of things... (Again, using a really vague phrase "in a sense" here and glossing over the details and doing a hand-wavy gesture.) There are different matters of naming, service discovery, and higher layer protocols — and drawing lines of correlation aren't always direct nor as clear cut as some would like.
A service instance is a combination of an SRV record containing the service hosts name and port and a TXT record containing meta-information (key=value pairs). Services instances are found via PTR records.
For HTTP services in a browse domain called example.com, you'd query for _http._tcp.example.com PTR records and get back a list of instance names such as HN._http._tcp.example.com. If you then wanted to connect to the "HN" HTTP service you'd then lookup the SRV and TXT records at HN._http._tcp.example.com.
For anyone interested, dns-sd.org hosts some example services — try `dig -t ptr _http._tcp.dns-sd.org` and go from there.
What were teh effects? How did the author notice? Was his server lagging? his OSX machine? If would be pretty awesome if you could dos MACs with 250KB/s flood.
That's a good question. I should have mentioned that in the post perhaps. The reason I initially ran vmstat was just to make sure my dev machine was "zeroed out". Recently my kernel builds had taken twice the time as normal and I wasn't sure why. That's when I noticed this. It turns out this had no significant effect on my build times but was rather an effect of a different change I made weeks prior. I also noticed no ill effects on my laptop that I recall.
At work, we had a problem where a computer was spamming the network with DHCP requests. It ended up physically locking us out; the embedded system which controls the card reader and unlocks the door locked up due to the load. I had to use an alternative way to get in, isolated the maglock, and later power-cycled the embedded system.
Needless to say, we enabled storm control on the switch after that.