Slashdot Mirror
Slashdot.org Self-Slashdotted
Slashdot.org was unreachable for about 75 minutes this evening. Here is the post-mortem from Sourceforge's chief network engineer Uriah Welcome. "What we had was indeed a DoS, however it was not externally originating. At 8:55 PM EST I received a call saying things were horked, at the same time I had also noticed things were not happy. After fighting with our external management servers to login I finally was able to get in and start looking at traffic. What I saw was a massive amount of traffic going across the core switches; by massive I mean 40 Gbit/sec. After further investigation, I was able to eliminate anything outside our network as the cause, as the incoming ports from Savvis showed very little traffic. So I started poking around on the internal switch ports. While I was doing that I kept having timeouts and problems with the core switches. After looking at the logs on each of the core switches they were complaining about being out of CPU, the error message was actually something to do with multicast. As a precautionary measure I rebooted each core just to make sure it wasn't anything silly. After the cores came back online they instantly went back to 100% fabric CPU usage and started shedding connections again. So slowly I started going through all the switch ports on the cores, trying to isolate where the traffic was originating. The problem was all the cabinet switches were showing 10 Gbit/sec of traffic, making it very hard to isolate. Through the process of elimination I was finally able to isolate the problem down to a pair of switches... After shutting the downlink ports to those switches off, the network recovered and everything came back. I fully believe the switches in that cabinet are still sitting there attempting to send 20Gbit/sec of traffic out trying to do something — I just don't know what yet. Luckily we don't have any machines deployed on [that row in that cabinet] yet so no machines are offline. The network came back up around 10:10 PM EST."
I believe you mean: Buffalo buffalo Buffalo buffalo buffalo buffalo Buffalo buffalo. The caps matters. In other words, Buffalo from the city of Buffalo that are pushed around by (other) buffalo from the city of Buffalo in turn push around (still more) buffalo from the city of Buffalo. And you thought this was unrelated to the recursive dupe comment.
I'm surprised STP was off by default. I remember in 1999 or so I had some trouble that resulted in my having to turn STP off on Cisco switches (they shipped with it on (these were 3524s and a 5505). I can't actually remember why. I think it had something to do with a Novell server?
The problem likely was that the machine required network at boot (typical Netware clients were like that, I've been told). STP started when the link went up, but it took a rather long time, so forwarding had not been enabled when the client required the network.
Since then, I have seen exactly that situation many times in small office environments. Also, the classic plugging in while also being on the wireless side of the network.
Port security helps a lot.
STP is also not fail-safe because typical switches happily forward traffic even if the STP process running on the CPU has died. If you build a L2 core, one broken switch (or OS glitch on a switch) can still take down your entire network easily (it's one of those pesky distributed, multiple single points of failure). In general, L3 networks are somewhat more robust in this regard, so it's often a good idea to avoid switch-to-switch connections (but that might be difficult, as it is difficult to tell L2 devices from L3 devices these days).
Depends how good your out-of-band management is.
Finally! A year of moderation! Ready for 2019?
There are no buffalo living in the US. Only bison. ;-)
== Jez ==
Do you miss Firefox? Try Pale Moon.
I don't know about you, but I'm suing for compensatory damages. Do you have any idea much pain and suffering the work I did in that time caused me?!
Fixed that for you. Sorry, law student.
If you can read this, it means that I bothered to log in.
Our network engineer lives a couple of states away from the data center. The work he's talking about doing, he did from home.
I'll be sure to when I get to the data center next week and am able to get my hands on the angry switch in question. I do love how it just sat there quietly for two weeks w/o doing anything and then decided randomly to just start blasting out 20 Gbit.. sigh.. hardware..
- U
Failed ASIC on the switch most likely.
I've see an issue just like that about once a year, but working with a sick number of systems globally the chances of seeing one offs becomes fairly regular.
Depending on the failure it might have logged what it was doing, but I'll presume since your monitoring didn't catch the spike it was because it was just random garbage.
Fun times!
"You should always go to other people's funerals; otherwise, they won't come to yours." -- Yogi Berra
The point is, they hadn't already given him direct access to those connections before yesterday, and he had to spend a large chunk of those 75 minutes getting the authorization to access the equipment so he COULD fix it.
That's not how I read it at all. The switches were so overloaded that he had to "fight" to get into the box. He, more than likely, already had access to the box, but the network was working against him.
core = core switch = a main switch that most of the edge switches/devices are plugged into.
reboot core = reboot a core switch.
For Slashdot staff, I think the generally accepted nominal is "It"...