We mentioned in Chapter 1 that there are three kinds of IP addresses: unicast, broadcast, and multicast. In this chapter we discuss broadcasting and multicasting in more detail.
Broadcasting and multicasting only apply to UDP, where it makes sense for an application to send a single message to multiple recipients. TCP is a connection-oriented protocol that implies a connection between two hosts (specified by IP addresses) and one process on each host (specified by port numbers).
Consider a set of hosts on a shared network such as an Ethernet. Each Ethernet frame contains the source and destination Ethernet addresses (48-bit values). Normally each Ethernet frame is destined for a single host. The destination address specifies a single interface-called a unicast. In this way communication between any two hosts doesn't bother any of the remaining hosts on the cable (except for possible contention for the shared media).
There are times, however, when a host wants to send a frame to every other host on the cable-called a broadcast. We saw this with ARP and RARP. Multicasting fits between unicasting and broadcasting: the frame should be delivered to a set of hosts that belong to a multicast group.
To understand broadcasting and multicasting we need to understand that filtering takes place on each host, each time a frame passes by on the cable. Figure 12.1 shows a picture of this.
First, the interface card sees every frame that passes by on the cable and makes a decision whether to receive the frame and deliver it to the device driver. Normally the interface card receives only those frames whose destination address is either the hardware address of the interface or the broadcast address. Additionally, most interfaces can be placed into a promiscuous mode whereby they receive a copy of every frame. This mode is used by tcpdump, for example.
Today most interfaces can also be configured to receive frames whose destination address is a multicast address, or some subset of multicast addresses. On an Ethernet, a multicast address has the low-order bit of the high-order byte turned on. In hexadecimal this bit looks like 01:00:00:00:00:00. (We can consider the Ethernet broadcast address, ff:ff:ff:ff:ff:ff as a special case of the Ethernet multicast address.)
If the interface card receives the frame, it is passed to the device driver. (One reason the interface card might discard the frame is if the Ethernet checksum is incorrect.) Additional filtering is performed by the device driver. First, the frame type must specify a protocol that is supported (IP, ARP, etc.). Second, additional multicast filtering may be performed, to check whether the host belongs to the addressed multicast group.
The device driver then passes the frame to the next layer, such as IP, if the frame type specifies an IP datagram. IP performs more filtering, based on the source and destination IP addresses, and passes the datagram up to the next layer (such as TCP or UDP) if all is well.
Each time UDP receives a datagram from IP, it performs filtering based on the destination port number, and sometimes the source port number too. If no process is currently using the destination port number, the datagram is discarded and an ICMP port unreachable message is normally generated. (TCP performs similar filtering based on its port numbers.) If the UDP datagram has a checksum error, UDP silently discards it.
The problem with broadcasting is the processing load that it places on hosts that aren't interested in the broadcasts. Consider an application that is designed to use UDP broadcasts. If there are 50 hosts on the cable, but only 20 are participating in the application, every time one of the 20 sends a UDP broadcast, the other 30 hosts have to process the broadcast, all the way up through the UDP layer, before the UDP datagram is discarded. The UDP datagram is discarded by these 30 hosts because the destination port number is not in use.
The intent of multicasting is to reduce this load
on hosts with no interest in the application. With multicasting
a host specifically joins one or more multicast groups. If possible,
the interface card is told which multicast groups the host belongs
to, and only those multicast frames are received.
In Figure 3.9 we showed four different forms of IP broadcast addresses. We now describe them in more detail.
The limited broadcast address is 255.255.255.255. This can be used as the destination address of an IP datagram during the host configuration process, when the host might not know its subnet mask or even its IP address.
A datagram destined for the limited broadcast address is never forwarded by a router under any circumstance. It only appears on the local cable.
An unanswered question is: if a host is multihomed and a process sends a datagram to the limited broadcast address, should the datagram be sent out each connected interface that supports broadcasting? If not, an application that wants to broadcast out all interfaces must determine all the interfaces on the host that support broadcasting, and send a copy out each interface.
Most BSD systems treat 255.255.255.255 as an alias for the broadcast address of the first interface that was configured, and don't provide any way to send a datagram out all attached, broadcast-capable interfaces. Indeed, two applications that send UDP datagrams out every interface are routed (Section 10.3) and rwhod (the server for the BSD rwho client). Both of these applications go through a similar start-up procedure to determine all the interfaces on the host, and which ones are capable of broadcasting. The net-directed broadcast address corresponding to that interface is then used as the destination address for datagrams sent out the interface.
The Host Requirements RFC takes no stand on the issue of whether a multihomed host should send a limited broadcast out all its interfaces.
The net-directed broadcast address has a host ID of all one bits. A class A net-directed broadcast address is netid.255.255.255, where netid is the class A network ID.
A router must forward a net-directed broadcast, but it must also have an option to disable this forwarding.
The subnet-directed broadcast address has a host ID of all one bits but a specific subnet ID. Classification of an IP address as a subnet-directed broadcast address requires knowledge of the subnet mask. For example, if a router receives a datagram destined for 184.108.40.206, this is a subnet-directed broadcast if the class B network 128.1 has a subnet mask of 255.255.255.0, but it is not a broadcast if the subnet mask is 255.255.254.0 (0xfffffe00).
An all-subnets-directed broadcast address also requires knowledge of the destination network's subnet mask, to differentiate this broadcast address from a net-directed broadcast address. Both the subnet ID and the host ID are all one bits. For example, if the destination's subnet mask is 255.255.255.0, then the IP address 220.127.116.11 is an all-subnets-directed broadcast. But if the network is not subnetted, then this is a net-directed broadcast.
Current feeling [Almquist 1993] is that this type of broadcast is obsolete. It is better to use multicasting than an all-subnets-directed broadcast.
[Almquist 1993] notes that RFC 922 requires that
an all-subnets-directed broadcast be sent to all subnets, but
no current routers do so. This is fortunate since a host that
has been misconfigured without its subnet mask sends all its "local"
broadcasts to all subnets. For example, if the host with IP address
18.104.22.168 doesn't set a subnet mask, then its broadcast address
normally defaults to 22.214.171.124. But if the subnet mask should
have been set to 255.255.255.0, then broadcasts from this misconfigured
host appear directed to all subnets.
The first widespread implementation of TCP/IP, the
4.2BSD system in 1983, used a host ID of all zero bits for the
broadcast address. One of the earliest references to the broadcast
IP address is IEN 212 [Gurwitz and Hinden 1982], and it proposed
to define the IP broadcast address as a host ID of one bits. (lENs
are the Internet Experiment Notes, basically predecessors
to the RFCs.) RFC 894 [Hornig 1984] commented that 4.2BSD used
a nonstandard broadcast address, but RFC 906 [Finlayson 1984]
noted that there was no Internet standard for the broadcast address.
The RFC editor added a footnote to RFC 906 acknowledging the lack
of a standard broadcast address, but strongly recommended that
a host ID of all one bits be used as the broadcast address. Although
Berkeley adopted the use of all one bits for the broadcast address
with 4.3BSD in 1986, some operating systems (notably SunOS 4.x)
continued to use the nonstandard broadcast address through the
12.3 Broadcasting Examples
The first widespread implementation of TCP/IP, the 4.2BSD system in 1983, used a host ID of all zero bits for the broadcast address. One of the earliest references to the broadcast IP address is IEN 212 [Gurwitz and Hinden 1982], and it proposed to define the IP broadcast address as a host ID of one bits. (lENs are the Internet Experiment Notes, basically predecessors to the RFCs.) RFC 894 [Hornig 1984] commented that 4.2BSD used a nonstandard broadcast address, but RFC 906 [Finlayson 1984] noted that there was no Internet standard for the broadcast address. The RFC editor added a footnote to RFC 906 acknowledging the lack of a standard broadcast address, but strongly recommended that a host ID of all one bits be used as the broadcast address. Although Berkeley adopted the use of all one bits for the broadcast address with 4.3BSD in 1986, some operating systems (notably SunOS 4.x) continued to use the nonstandard broadcast address through the early 1990s.
How are broadcasts sent and what do routers and hosts do with broadcasts? Unfortunately this is a hard question to answer because it depends on the type of broadcast address, the application, the TCP/IP implementation, and possible configuration switches. First, the application must support broadcasting. If we execute
|sun % ping 255.255.255.255
/usr/etc/ping: unknown host 255.255.255.255
intending to send a broadcast on the local cable, it doesn't work. But the problem here is a programming problem in the application (ping). Most applications that accept either a dotted-decimal IP address or a hostname call the function inet_addr(3) to convert the dotted-decimal character string to its 32-bit binary IP address, and if this fails, assume the character string is a hostname. Unfortunately this library function returns -1 to indicate an error (such as a character other than a digit or decimal point in the string), but the limited broadcast address (255.255.255.255) also converts into -1. Most programs then assume that the character string is a hostname, look it up using the DNS (Chapter 14), and end up printing an error such as "unknown host."
If we fix this programming shortfall in the ping program, however, the results are often not what we expect. On six different systems tested by the author, only one handled this as expected and generated a broadcast packet on the local cable. Most looked up the IP address 255.255.255.255 in the routing table, applied the default route, and sent a unicast packet to the default router. Eventually the packet was thrown away
A subnet-directed broadcast is what we should be using. Indeed, in Section 6.3 we sent datagrams to the IP address 126.96.36.199 for the bottom Ethernet in our test network (inside front cover), and got replies from all the hosts on the Ethernet. The subnet-directed broadcast address associated with each interface is the value used with the ifconfig command (Section 3.8). If we ping that address, the result is what we expect:
sun % arp -a ARP cache is empty
sun % ping 188.8.131.52
PING 184.108.40.206: 56 data bytes
64 bytes from sun (220.127.116.11): icmp_seq=0. time=4. ms
64 bytes frombsdi (18.104.22.168): icmp_seq=0. time=172. ms
64 bytes from svr4 (22.214.171.124): icmp_seq=0. time=192. ms
64 bytes from sun (126.96.36.199): icmp_seq=l. time=1. ms
64 bytes from bsdi (188.8.131.52): icmp_seq=l. time=52. ms
64 bytes from svr4 (184.108.40.206): icmp_seq=l. time=90. ms
interrupt key to stop
--220.127.116.11 PING Statistics--
2 packets transmitted, 6 packets received, -200% packet loss
round-trip (ms) min/avg/max = 1/85/192
sun % arp -a check ARP cache again
svr4 (18.104.22.168) at 0:0:c0:c2:9b:26
bsdi (22.214.171.124) at 0:0:c0:6f:2d:40
IP looks at the destination address (126.96.36.199), determines that it is the subnet-directed broadcast address, and sends the datagram to the link-layer broadcast address.
We mentioned in Section 6.3 that this type of broadcast means all the hosts on the local network, including the sender. We see here that we do get a reply from the sending host (sun) in addition to the other hosts on the cable.
In this example we've also shown the ARP cache before and after the ping of the broadcast address. This is to show the interaction between broadcasting and ARP. "The ARP cache is empty before we execute ping, but full afterward. (That is, there is one entry for every other host on the cable that responded to the echo request.) How did this happen when we said that the Ethernet frame is sent to the link-layer broadcast address (0xffffffff)? The sending of these frames by sun does not require ARP.
If we watch ping using tcpdump, we see that it is the recipients of the broadcast frames that generate an ARP request to sun, before they can send their reply. This is because the reply is unicast. We said in Section 4.5 that the receiver of an ARP request (sun in this example) normally adds the requestor's IP address and hardware address to its ARP cache, in addition to sending an ARP reply. This is on the assumption that if the requestor is about to send us a packet, we'll probably want to send something back.
Our use of ping is somewhat special because the type of programming interface that it uses (called "raw sockets" on most Unix implementations) always allows a datagram to be sent to the broadcast address. What if we use an application that was not designed to support broadcasting, such as TFTP? (We cover TFTP in more detail in Chapter 15.)
|bsdi % tftp||start the client|
|tftp> connect 188.8.131.52||specify the IP address of the server|
|tftp> get temp.foo||and try to fetch a file from the server|
|tftp: sendto: Permission denied|
|tftp> quit||terminate the client|
Here we get an error immediately, and nothing is sent on the cable. What's happening here is that the sockets API doesn't allow a process to send a UDP datagram to the broadcast address unless the process specifically states that it plans to broadcast. This is intended to prevent users from mistakenly specifying a broadcast address (as we did here) when the application was never intended to broadcast.
With the sockets API the application must set the
SO_BROADCAST socket option before
sending a UDP datagram to a broadcast address.
Not all systems enforce this restriction. Some implementations
allow any process to broadcast UDP datagrams, without requiring
the process to say so. Others are more restrictive and require
a process to have superuser privileges to broadcast.
Not all systems enforce this restriction. Some implementations allow any process to broadcast UDP datagrams, without requiring the process to say so. Others are more restrictive and require a process to have superuser privileges to broadcast.
The next question is whether directed broadcasts are forwarded or not. Some kernels and routers have an option to enable or disable this feature. (See Appendix E.)
If we enable this feature on our router bsdi and run ping from the host slip, we can see if the subnet-directed broadcasts are forwarded by bsdi. Forwarding a directed broadcast means the router takes the incoming unicast datagram, determines that the destination address is the directed broadcast for one of its interfaces, and then forwards the datagram onto the appropriate network using a link-layer broadcast.
slip % ping 184.108.40.206
PING 220.127.116.11 (18.104.22.168): 56 data bytes
from 22.214.171.124: icmp_seq=0 ttl=255 time=190 ms 64 bytes
from 126.96.36.199: icmp_seq=0 ttl=254 time=280 ms (DUP!)
64 bytes from 188.8.131.52; icmp_seq=0 ttl=254 time=360 ms (DUP!)
64 bytes from 184.108.40.206; icmp_seq=l ttl=255
64 bytes from 220.127.116.11: icmp_seq=l ttl=254 time=270 ms (DUP!)
64 bytes from 18.104.22.168: icmp_seq=l ttl=254 time=360 ms (DUP!)
type interrupt key to
- 22.214.171.124 ping statistics -
3 packets transmitted, 2 packets received, +4 duplicates, 33% packet loss
round-trip min/avg/max = 180/273/360 ms
We see that this does indeed work. We also see that the BSD ping program checks for duplicate sequence numbers and prints DUP! when this occurs. It normally means a packet was duplicated somewhere, but here we expect to see this, since we sent the requests to a broadcast address.
We can also run this test from a host much farther away from the network to which the broadcast is directed. If we run ping from the host vangogh.cs.berkeley.edu (14 hops away from our network), it still works if the router sun is configured to forward directed broadcasts. In this case the IP datagrams (carrying the ICMP echo requests) are forwarded by every router in the path as a normal datagram. None of them knows that it's really a directed broadcast. The next to last router, netb, thinks it's for the host with an ID of 63, and forwards it to sun. It is the router sun that detects that the destination IP address is really the broadcast address of an attached interface, and turns the datagram into a link-layer broadcast on that network.
Broadcasting is a feature that should be used with
great care. In many cases IP multicasting will prove to be a better
IP multicasting provides two services for an application.
In this section we'll take a look at multicast addresses, and the next chapter looks at the protocol used by multicasting hosts and routers (IGMP).
Figure 12.2 shows the format of a class D IP address.
Unlike the other three classes of IP addresses (A, B, and C), which we showed in Figure 1.5, the 28 bits allocated for the multicast group ID have no further structure.
A multicast group address is the combination of the high-order 4 bits of 1110 and the multicast group ID. These are normally written as dotted-decimal numbers and are in the range 126.96.36.199 through 188.8.131.52.
The set of hosts listening to a particular IP multicast address is called a host group. A host group can span multiple networks. Membership in a host group is dynamic-hosts may join and leave host groups at will. There is no restriction on the number of hosts in a group, and a host does not have to belong to a group to send a message to that group.
Some multicast group addresses are assigned as well-known addresses by the IANA (Internet Assigned Numbers Authority). "These are called permanent host groups. This is similar to the well-known TCP and UDP port numbers. Similarly, these well-known multicast addresses are listed in the latest Assigned Numbers RFC. Notice that it is the multicast address of the group that is permanent, not the membership of the group.
For example, 184.108.40.206 means "all systems on this subnet," and 220.127.116.11 means "all routers on this subnet." The multicast address 18.104.22.168 is for NTP, the Network Time Protocol, 22.214.171.124 is for RIP-2 (Section 10.5), and 126.96.36.199 is for SGI's (Silicon Graphics) dogfight application.
The IANA owns an Ethernet address block, which in hexadecimal is 00:00:5e. This is the high-order 24 bits of the Ethernet address, meaning that this block includes addresses in the range 00:00:5e:00:00:00 through 00:00:5e:ff:ff:ff. The IANA allocates half of this block for multicast addresses. Given that the first byte of any Ethernet address must be 01 to specify a multicast address, this means the Ethernet addresses corresponding to IP multicasting are in the range 01:00:5e:00:00:00 through 01:00:5e:7f:ff:ff.
Our notation here uses the Internet standard bit order, for a CSMA/CD or token bus network, as the bits appear in memory. This is what most programmers and system administrators deal with. The IEEE documentation uses the transmission order of the bits. The Assigned Numbers RFC gives additional details on the differences between these representations.
This allocation allows for 23 bits in the Ethernet address to correspond to the IP multicast group ID. The mapping places the low-order 23 bits of the multicast group ID into these 23 bits of the Ethernet address. This is shown in Figure 12.3.
Since the upper 5 bits of the multicast group ID are ignored in this mapping, it is not unique. Thirty-two different multicast group IDs map to each Ethernet address. For example, the multicast addresses 188.8.131.52 (hex e0.80.40.20) and 184.108.40.206 (hex e0.00 40.20) both map into the Ethernet address 01:00:5e:00:40:20.
Since the mapping is not unique, it implies that
the device driver or the IP module in Figure 12.1 must perform
filtering, since the interface card may receive multicast frames
in which the host is really not interested. Also, if the interface
provide adequate filtering of multicast frames, the device driver may have to receive all multicast frames, and perform the filtering itself.
LAN interface cards tend to come in two varieties.
One type performs multicast filtering based on the hash value
of the multicast hardware address, which means some unwanted frames
can always get through. The other type has a small, fixed number
of multicast addresses to listen for, meaning that when the host
needs to receive more multicast addresses than are supported,
the interface must be put into a "multicast promiscuous"
mode. Hence, both types of interfaces still require that the device
driver perform checking that the received frame is really wanted.
Even if the interface performs perfect multicast
filtering (based on the 48-bit hardware address), since the mapping
from a class D IP address to a 48-bit hardware address is not
one-to-one, filtering is still required.
Despite this imperfect address mapping and hardware
filtering, multicasting is still better than broadcasting.
Even if the interface performs perfect multicast filtering (based on the 48-bit hardware address), since the mapping from a class D IP address to a 48-bit hardware address is not one-to-one, filtering is still required.
Despite this imperfect address mapping and hardware filtering, multicasting is still better than broadcasting.
Multicasting on a single physical network is simple. The sending process specifies a destination IP address that is a multicast address, the device driver converts this to the corresponding Ethernet address, and sends it. The receiving processes must notify their IP layers that they want to receive datagrams destined for a given multicast address, and the device driver must somehow enable reception of these multicast frames. This is called "joining a multicast group." (The reason we use the plural "receiving processes" is because there are normally multiple receivers for a given multicast message, either on the same host or on multiple hosts, which is why we're using multicasting in the first place.) When a multicast datagram is received by a host, it must deliver a copy to all the processes that belong to that multicast group. This is different from UDP where a single process receives an incoming unicast UDP datagram. With multicasting it is possible for multiple processes on a given host to belong to the same multicast group.
But complications arise when we extend multicasting beyond a single physical network and pass multicast packets through routers. A protocol is needed for multicast routers to know if any hosts on a given physical network belong to a given multicast group. This protocol is called the Internet Group Management Protocol (IGMP) and is the topic of the next chapter.
FDDI networks use the same mapping between the class
D IP address and the 48-bit FDDI address [Katz 1990]. Token ring
networks normally use a different mapping, because of limitations
in most token ring controllers [Pusateri 1993].
Broadcasting is sending a packet to all hosts on a network (usually a locally attached network) and multicasting is sending a packet to a set of hosts on a network. Basic to these two concepts is an understanding of the different types of filtering that occur when a received frame passes up a protocol stack. Each layer can discard a received packet for different reasons.
There are four types of broadcast addresses: limited, net-directed, subnet-directed, and all-subnets-directed. The most common is subnet-directed. The limited broadcast address is normally seen only when a system is bootstrapping.
Problems occur when trying to broadcast through routers, often because the router may not know the subnet mask of the destination network. The results depend on many factors: which type of broadcast address, configuration parameters, and so on.
A class D IP address is called a multicast group address. It is converted to an Ethernet address by placing its lower 23 bits into a fixed Ethernet address. The mapping is not unique, requiring additional filtering by one of the protocol modules.
12.1 Does broadcasting increase the amount of network traffic?
12.2 Consider 50 hosts on an Ethernet: 20 running TCP/IP and 30 running some other protocol suite. How are broadcasts from one protocol suite handled by hosts running the other protocol suite?
12.3 You login to a Unix system that you've never used before and want to find the subnet-directed broadcast address for all attached interfaces that support broadcasting. How can you do this?
12.4 If we ping the broadcast address with a large packet size, as in
|sun % ping 220.127.116.11 1472
PING 18.104.22.168; 1472 data bytes
1480 bytes from sun (22.214.171.124): icmp_seq=0. time=6. ms
1480 bytes from svr4 (126.96.36.199): icmp_seq=0. time=84. ms
1480 bytes from bsdi (188.8.131.52): icmp_seq=0. time=128. ms
it works, but increasing the packet size by 1 byte gives us the following error:
sun % ping 184.108.40.206 1473
What's going on?
12.5 Redo Exercise 10.6 assuming the eight RIP messages are multicast instead of broadcast (assume RIP Version 2 is being used). What changes?