Posts tagged switching

At the office, we reprovision servers like it’s going out of style.  It happens so often that my cabling documentation rarely matches what’s actually out in field, which is a pretty big problem when you’re trying to find to what switch port a server is connected.  I finally relegated myself to asking for the MAC address of the server, having the admin ping something, and then tracing it down through the CAM table entries of the switches.  It works, but the guys really don’t know how a switch populates its CAM table, so they always say “Why can’t you just look on the switch?  I shouldn’t have to ping anything.”  Here’s one just for the aspiring system admin.

The Content Addressable Memory (CAM) table on a switch keeps track of MAC addresses and on what port they appear, along with some other stuff like age.  When a device that’s plugged into a particular port sends a frame to the switch, the switch makes note of the source MAC and the port and checks the CAM table.  If it’s a new MAC, it adds an entry in the CAM table; if it’s an existing on a different port, it removes the old entry and adds a new one; if it’s an old MAC on the same port, it updates the age.  By default, Cisco switches keep CAM entries for 300 seconds, so they don’t stay there forever.

What about the destination MAC?  Good question.  That’s a pretty important field when sending a packet, but, when generating a CAM entry, the destination MAC is ignored.  If  a host talks, a switch knows exactly from where the frame came, but there’s no way to know exactly where it should go without the destination first speaking up.

Let’s set up an example.  You have a Cisco 2950 switch that you’ve just powered on with nothing plugged into it except the console cable. If you do a show mac-address-table, you’ll see the CAM table — a table of MAC addresses that the switch knows; you would think that it would be empty since nothing’s plugge din, but the switch has its own MACs, so it always knows those guys.  There’s not much to see here yet, though, since we haven’t hooked anything up to the switch yet.

Switch#sh mac-address-table
Mac Address Table
-------------------------------------------
Vlan    Mac Address       Type        Ports
----    -----------       --------    -----
 All    000a.f43b.ddc0    STATIC      CPU
 All    0100.0ccc.cccc    STATIC      CPU
 All    0100.0ccc.cccd    STATIC      CPU
 All    0100.0cdd.dddd    STATIC      CPU
Switch#

Next, let’s plug a Linux desktop up to it.  Once that box has booted, what should you see in the CAM table?  If you guessed the MAC of the Linux box, you may be right; it all depends on if the server sent a frame or not.  There’s lots of things that run on a Linux box that could send frames on startup — DHCP requests, multicast services, network-based storage — so, more than likely, a frame did get sent.  The only way to know it to take a gander.

Switch#sh mac-address-table
Mac Address Table
-------------------------------------------
Vlan    Mac Address       Type        Ports
----    -----------       --------    -----
...
   1    001c.0cbb.ada2    DYNAMIC     Fa0/1
Switch#

Ah…it worked. That’s good, but it’s boring with only a single device. Let’s plug in a simply-configured and fully-booted Cisco router and see what happens. More than likely the router won’t speak until spoken to, so the CAM table won’t update, and the switch won’t know where to send the frame, right? Yes, but the frame still gets sent. If the switch doesn’t know where the destination MAC lives (i.e., it’s not in the CAM table), then it floods the frame out every port except the one on which it was received.

When I first learned that this is how it worked, I immediately wondered why LANs weren’t flooded out constantly. I didn’t think long enough to realize that the host being sent the packet will actually respond within several milliseconds (hopefully), so the CAM table will then have an entry for that guy. In reality, it’s even simpler than that. Since most of the world runs TCP/IP, we have this wonderful thing called ARP. When a host needs to talk to another host on the same network segment (IP subnet), it checks its ARP table, and, if it doesn’t know a MAC for that IP, it will actually ask what MAC it should use via a broadcast message. In a well-behaved network, the mystery host will answer with a “Here I am!” type message, which causes the switch to generate a CAM entry. In a “perfect world”, you should only have a few floods on a switch per day/week/month/year/decade.

Here’s a couple items of note.

• A trunk interface will have a whole bunch of MACs listed as attached to that port.  This is quite normal, so don’t freak out.
• If someone plugs a switch or hub into one of your ports, you will see multiple CAM entries for the same port.  This is a good way to see who brought in their Linksys hub from home.
• If a host hasn’t sent a frame in more than 5 minutes, it disappears from the CAM table, so the whole discovery process starts over again.
• There’s a limit to the size of a CAM table, so it’s possible to fill it up and then every new destination gets flooded.  Wow, I can see your packets.

Have fun.  Be safe.  Practice safe computing.  Lock down your network.

We run a large number of LANs all over the country that are “controlled” by the particular business unit. We manage the gear, but, since they have the money and have to pay for anything we do, they make the final decision on what gets put in. Sometimes that gets out of hand, as you can well imagine.

A good terrible example came up a few months ago. It seems that, at some time in the past, one site needed some more LAN ports, but, instead of calling us and having us send them another switch, one of the “technical people” there brought in a hub from home. It really irks me to see a hub on the switched LAN, but we really have no control over those decisions. They plugged the hub into one of the existing drops somewhere in the building and plugged everyone in. It worked…until somebody moved one of the machines. The machine was at a desk near the hub, and the network cable, still with one end plugged into the hub, was just left lying there. A good Samaritan came by, saw that the hub was not plugged into the network (though it was through another path), and plugged it back in for us — providing a nice second link from the hub to the switch stack in the closet. Take one switch stack, add a hub, insert a switching loop, bake at 350F for a few milliseconds, and you have a broadcast storm. If you don’t know already, broadcast storms are bad and eat switch CPU like the yummy cookies we baked. In this case, several 3750s were taken completely down.

How does one prevent such from happening again? Well, the first thing to do is to get the CTO to tell everyone that they can’t plug hubs into the network. That works about 0% of the time, though, so we had to find a solution that was enforceable. One of my coworkers found the traffic storm control mechanism built into Cisco switches. This mechanism allows you to set thresholds based on broadcast, multicast, and unicast traffic and take action when those are reached.

Here are the gory details. I need to mention, though, that storm-control is configured very differently across platforms and IOS versions. I would say your mileage may vary, but it’s probably more accurate to say that this won’t work on your switch. A 6500 is configured differently than a 4500. A 2900XL is different from a 2950. This will get you going, but you’re going to have to do some research on your own to find out what works on your platform.

interface FastEthernet 0/1
storm-control broadcast level 50
storm-control action shutdown

What just happened? Good question to ask. If broadcast traffic on F0/1 utilizes 50% of available bandwidth, the port is shutdown. That means that if broadcast traffic takes up 50Mbps of bandwidth on this port, the port is admined down just as if you did a shutdown on it.. You should probably do the same for multicast or unicast as well to make sure you don’t get bitten by those. If you don’t want to shut down the port, you can also use the trap action to just send an SNMP trap with the port and information, but that doesn’t prevent very much; the storm will probably wreak havoc before an email for the trap lands on your Crackberry.

Here’s another big disclaimer. Finding a good level for a port can be very, very difficult. A linux box is going to have very different broadcast/multicast/unicast traffic than a Windows box which is different than a Mac. You may have to spend a lot of time analyzing SNMP counters to find out what a good level is. God help you if you have a hub like we did with mixed computer platforms on it.

In my professional life at some point, I came across someone who had a stack of Catalyst 2950 switches all trunked together with their Internet routers connected to the top of the stack. This was all well and good until they kept adding hosts to the “middle” of the stack, then they had all sorts of latency and packet loss.

The old adage of your chain only being as strong as your weakest length holds true in this case. Here, the weakest link is actually the most-congested trunk, though. Let’s step through to see. A 2950 is a 10/100 switch, so a single trunk can handle 100Mbps of traffic. We have 10 of these guys, Switch1 to Switch10, all trunked to the one above and below. If a server in the center of the stack on Switch5 is sending a lot of data to the Internet routers on Switch1, the trunks off of Switch5 will start to get saturated. Switch4 has a few hosts doing the same thing, so traffic from both Switch4 and Switch5 heads towards Switch1, further filling the trunks. Same for Switch3. Same for Switch2. Next thing you know, there’s 184Mbps or so trying to go across a 100Mbps link.

I fixed the problem using EtherChannel. EtherChannel, sometimes called trunking (not Cisco trunks…don’t get confused) or bonding, takes up to 8 interfaces of a switch and binds them logically as one Port-channel interface. The switch then load-balances (not really, but can be pretty close) the traffic across all the links, so, in essence, if you have 2 100Mbps ports in an EtherChannel, you get 200Mbps of bandwidth. If you put in 8 ports, you’ll get 800Mbps. Or, at least, close to it.

There are, as usual, stipulations for using it.

• All ports on the EtherChannel must terminate to the same device on both ends. You can’t have one port go to one switch and another to another switch.
• The ports must use identical media. You can’t have a copper port and a fiber port in the same group.
• The ports must have the same capabilities. You can’t put a 100Mbps port and a 1Gbps port in the same group.
• The ports must be configured identically or it won’t work. This is actually pretty easy to maintain since configuring the Port-channel actually sets the config on all participating ports.
• Both ends of the ports must be EtherChannel. You can’t run it on one switch but not the other.

Enough of that. Let’s configure. The scenario is that you have SwitchA and you want to bind F0/1 and F0/2 into an EtherChannel. Of course, you want to carry traffic from all VLANs, so let’s make it a trunk.

int F0/1
channel-group 1

int F0/2
channel-group 1

int Port-channel 1
speed 100
duplex full
switchport trunk encapsulation dot1q
switchport mode trunk

Easy as pie. Do the same thing on the other end and you have a 200Mbps, bonded trunk.

If you didn’t now already, trunks are connections between switches that carry traffic for all VLANs. It allows you to have, say, VLAN 10 and VLAN 20 on two switches appear as the same network. Unless you’re a really small shop, you’ve already dealt with trunks, so there’s no need for an introduction.

Let’s say we have a Catlyst 2950 switch with multiple VLANs connected to another 2950 configured with those same VLANs. We’ll say we have VLANs 10, 20, and 30 and that the switches are connected to port F0/24 of each switch. First, let’s turn on the trunk.

interface F0/24
switchport trunk encapsulation dot1q
switchport mode trunk

Quite easy there. With this configuration on each switch, the connection between them will carry traffic for all VLANs. The encapsulation directive tells the switches to use the IEEE standard 802.1Q for the trunk, which is VLAN tagging. Cisco has its own trunk encapsulation called ISL, but that’s not compatible with non-Cisco gear. If you have a mix of switches, just use the dot1q encapsulation so you don’t hurt yourself later.

A note on the word “encapsulation” here. Dot1q does not actually encapsulate; it adds 4 bytes to the frame header that marks the VLAN the frame is for. ISL, however, does encapsulate; it takes the whole frame, shoves it into an ISL frame, and sends it on. Since Cisco’s preferred method for a trunk is an encapsulation method, we have the directive “encapsulation” in the configs.

At this point, all VLANs are being carried across the trunk, but what if you want to use multiple trunks and send different traffic across each one? For example, let’s say that you want to have VLAN 10 traffic use a second trunk while the other VLANs use our original trunk. To do that, you get into pruning.

interface F0/24
…
switchport trunk allowed vlan 20, 30

interface F0/23
switchport trunk encapsulation dot1q
switchport mode trunk
switchport trunk allowed vlan 10

The switchport trunk allowed vlan directive says that only traffic on VLANs 20 and 30 are allowed across F0/24 and only VLAN 10 across F0/23. I use this type of setup to give high-bandwidth VLANs (like VLANs for backups) their own trunk so they won’t eat all the bandwidth of the other VLANs. To use the terminology, F0/24 is pruned to VLANs 20 and 30, while F0/23 is pruned to VLAN 10.

I also want to mention that the word trunking is used differently across different platforms. We have a nearly-totally Cisco LAN, and trunks are the connections that carry all VLANs as described. On other LAN gear, trunking is actually the act of combining port, links, cables, whatever, together to form a single logical connection (Cisco calls this EtherChannel). VLAN tagging is what other manufacturers call a Cisco trunk. It makes sense if you remember that 802.1Q simply tags the frame.