Basic networking¶

Ethernet¶

Ethernet is a networking protocol, specified by the IEEE 802.3 standard. Most wired network interface cards (NICs) communicate using Ethernet.

In the OSI model of networking protocols, Ethernet occupies the second layer, which is known as the data link layer. When discussing Ethernet, you will often hear terms such as local network, layer 2, L2, link layer and data link layer.

In an Ethernet network, the hosts connected to the network communicate by exchanging frames. Every host on an Ethernet network is uniquely identified by an address called the media access control (MAC) address. In particular, in an OpenStack environment, every virtual machine instance has a unique MAC address, which is different from the MAC address of the compute host. A MAC address has 48 bits and is typically represented as a hexadecimal string, such as 08:00:27:b9:88:74. The MAC address is hard-coded into the NIC by the manufacturer, although modern NICs allow you to change the MAC address programmatically. In Linux, you can retrieve the MAC address of a NIC using the ip command:

$ ip link show eth0
2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc pfifo_fast state UP mode DEFAULT group default qlen 1000
    link/ether 08:00:27:b9:88:74 brd ff:ff:ff:ff:ff:ff

Conceptually, you can think of an Ethernet network as a single bus that each of the network hosts connects to. In early implementations, an Ethernet network consisted of a single coaxial cable that hosts would tap into to connect to the network. Modern Ethernet networks do not use this approach, and instead each network host connects directly to a network device called a switch. Still, this conceptual model is useful, and in network diagrams (including those generated by the OpenStack dashboard) an Ethernet network is often depicted as if it was a single bus. You’ll sometimes hear an Ethernet network referred to as a layer 2 segment.

In an Ethernet network, every host on the network can send a frame directly to every other host. An Ethernet network also supports broadcasts, so that one host can send a frame to every host on the network by sending to the special MAC address ff:ff:ff:ff:ff:ff. ARP and DHCP are two notable protocols that use Ethernet broadcasts. Because Ethernet networks support broadcasts, you will sometimes hear an Ethernet network referred to as a broadcast domain.

When a NIC receives an Ethernet frame, by default the NIC checks to see if the destination MAC address matches the address of the NIC (or the broadcast address), and the Ethernet frame is discarded if the MAC address does not match. For a compute host, this behavior is undesirable because the frame may be intended for one of the instances. NICs can be configured for promiscuous mode, where they pass all Ethernet frames to the operating system, even if the MAC address does not match. Compute hosts should always have the appropriate NICs configured for promiscuous mode.

As mentioned earlier, modern Ethernet networks use switches to interconnect the network hosts. A switch is a box of networking hardware with a large number of ports, that forwards Ethernet frames from one connected host to another. When hosts first send frames over the switch, the switch doesn’t know which MAC address is associated with which port. If an Ethernet frame is destined for an unknown MAC address, the switch broadcasts the frame to all ports. The port learns which MAC addresses are at which ports by observing the traffic. Once it knows which MAC address is associated with a port, it can send Ethernet frames to the correct port instead of broadcasting. The switch maintains the mappings of MAC addresses to switch ports in a table called a forwarding table or forwarding information base (FIB). Switches can be daisy-chained together, and the resulting connection of switches and hosts behaves like a single network.

VLANs¶

VLAN is a networking technology that enables a single switch to act as if it was multiple independent switches. Specifically, two hosts that are connected to the same switch but on different VLANs do not see each other’s traffic. OpenStack is able to take advantage of VLANs to isolate the traffic of different tenants, even if the tenants happen to have instances running on the same compute host. Each VLAN has an associated numerical ID, between 1 and 4095. We say “VLAN 15” to refer to the VLAN with numerical ID of 15.

To understand how VLANs work, let’s consider VLAN applications in a traditional IT environment, where physical hosts are attached to a physical switch, and no virtualization is involved. Imagine a scenario where you want three isolated networks, but you only have a single physical switch. The network administrator would choose three VLAN IDs, say, 10, 11, and 12, and would configure the switch to associate switchports with VLAN IDs. For example, switchport 2 might be associated with VLAN 10, switchport 3 might be associated with VLAN 11, and so forth. When a switchport is configured for a specific VLAN, it is called an access port. The switch is responsible for ensuring that the network traffic is isolated across the VLANs.

Now consider the scenario that all of the switchports in the first switch become occupied, and so the organization buys a second switch and connects it to the first switch to expand the available number of switchports. The second switch is also configured to support VLAN IDs 10, 11, and 12. Now imagine host A connected to switch 1 on a port configured for VLAN ID 10 sends an Ethernet frame intended for host B connected to switch 2 on a port configured for VLAN ID 10. When switch 1 forwards the Ethernet frame to switch 2, it must communicate that the frame is associated with VLAN ID 10.

If two switches are to be connected together, and the switches are configured for VLANs, then the switchports used for cross-connecting the switches must be configured to allow Ethernet frames from any VLAN to be forwarded to the other switch. In addition, the sending switch must tag each Ethernet frame with the VLAN ID so that the receiving switch can ensure that only hosts on the matching VLAN are eligible to receive the frame.

When a switchport is configured to pass frames from all VLANs and tag them with the VLAN IDs it is called a trunk port. IEEE 802.1Q is the network standard that describes how VLAN tags are encoded in Ethernet frames when trunking is being used.

Note that if you are using VLANs on your physical switches to implement tenant isolation in your OpenStack cloud, you must ensure that all of your switchports are configured as trunk ports.

It is important that you select a VLAN range that your current network infrastructure is not using. For example, if you estimate that your cloud must support a maximum of 100 projects, pick a VLAN range outside of that value, such as VLAN 200–299. OpenStack and all physical network infrastructure that handles tenant networks must then support this VLAN range.

Trunking is used to connect between different switches. Each trunk uses a tag to identify which VLAN is in use. This ensures that switches on the same VLAN can communicate.

Subnets and ARP¶

While NICs use MAC addresses to address network hosts, TCP/IP applications use IP addresses. The Address Resolution Protocol (ARP) bridges the gap between Ethernet and IP by translating IP addresses into MAC addresses.

IP addresses are broken up into two parts: a network number and a host identifier. Two hosts are on the same subnet if they have the same network number. Recall that two hosts can only communicate directly over Ethernet if they are on the same local network. ARP assumes that all machines that are in the same subnet are on the same local network. Network administrators must take care when assigning IP addresses and netmasks to hosts so that any two hosts that are in the same subnet are on the same local network, otherwise ARP does not work properly.

To calculate the network number of an IP address, you must know the netmask associated with the address. A netmask indicates how many of the bits in the 32-bit IP address make up the network number.

There are two syntaxes for expressing a netmask:

• dotted quad
• classless inter-domain routing (CIDR)

Consider an IP address of 192.168.1.5, where the first 24 bits of the address are the network number. In dotted quad notation, the netmask would be written as 255.255.255.0. CIDR notation includes both the IP address and netmask, and this example would be written as 192.168.1.5/24.

Sometimes we want to refer to a subnet, but not any particular IP address on the subnet. A common convention is to set the host identifier to all zeros to make reference to a subnet. For example, if a host’s IP address is 10.10.53.24/16, then we would say the subnet is 10.10.0.0/16.

To understand how ARP translates IP addresses to MAC addresses, consider the following example. Assume host A has an IP address of 192.168.1.5/24 and a MAC address of fc:99:47:49:d4:a0, and wants to send a packet to host B with an IP address of 192.168.1.7. Note that the network number is the same for both hosts, so host A is able to send frames directly to host B.

The first time host A attempts to communicate with host B, the destination MAC address is not known. Host A makes an ARP request to the local network. The request is a broadcast with a message like this:

To: everybody (ff:ff:ff:ff:ff:ff). I am looking for the computer who has IP address 192.168.1.7. Signed: MAC address fc:99:47:49:d4:a0.

Host B responds with a response like this:

To: fc:99:47:49:d4:a0. I have IP address 192.168.1.7. Signed: MAC address 54:78:1a:86:00:a5.

Host A then sends Ethernet frames to host B.

You can initiate an ARP request manually using the arping command. For example, to send an ARP request to IP address 10.30.0.132:

$ arping 10.30.0.132
ARPING 10.30.0.132 from 10.30.0.131 eth0
Unicast reply from 10.30.0.132 [54:78:1A:86:1C:0B]  0.670ms
Unicast reply from 10.30.0.132 [54:78:1A:86:1C:0B]  0.722ms
Unicast reply from 10.30.0.132 [54:78:1A:86:1C:0B]  0.723ms
Sent 3 probes (1 broadcast(s))
Received 3 response(s)

To reduce the number of ARP requests, operating systems maintain an ARP cache that contains the mappings of IP addresses to MAC address. On a Linux machine, you can view the contents of the ARP cache by using the arp command:

$ arp -n
Address                  HWtype  HWaddress           Flags Mask            Iface
10.0.2.3                 ether   52:54:00:12:35:03   C                     eth0
10.0.2.2                 ether   52:54:00:12:35:02   C                     eth0

DHCP¶

Hosts connected to a network use the Dynamic Host Configuration Protocol (DHCP) to dynamically obtain IP addresses. A DHCP server hands out the IP addresses to network hosts, which are the DHCP clients.

DHCP clients locate the DHCP server by sending a UDP packet from port 68 to address 255.255.255.255 on port 67. Address 255.255.255.255 is the local network broadcast address: all hosts on the local network see the UDP packets sent to this address. However, such packets are not forwarded to other networks. Consequently, the DHCP server must be on the same local network as the client, or the server will not receive the broadcast. The DHCP server responds by sending a UDP packet from port 67 to port 68 on the client. The exchange looks like this:

1. The client sends a discover (“I’m a client at MAC address 08:00:27:b9:88:74, I need an IP address”)
2. The server sends an offer (“OK 08:00:27:b9:88:74, I’m offering IP address 10.10.0.112”)
3. The client sends a request (“Server 10.10.0.131, I would like to have IP 10.10.0.112”)
4. The server sends an acknowledgement (“OK 08:00:27:b9:88:74, IP 10.10.0.112 is yours”)

OpenStack uses a third-party program called dnsmasq to implement the DHCP server. Dnsmasq writes to the syslog, where you can observe the DHCP request and replies:

Apr 23 15:53:46 c100-1 dhcpd: DHCPDISCOVER from 08:00:27:b9:88:74 via eth2
Apr 23 15:53:46 c100-1 dhcpd: DHCPOFFER on 10.10.0.112 to 08:00:27:b9:88:74 via eth2
Apr 23 15:53:48 c100-1 dhcpd: DHCPREQUEST for 10.10.0.112 (10.10.0.131) from 08:00:27:b9:88:74 via eth2
Apr 23 15:53:48 c100-1 dhcpd: DHCPACK on 10.10.0.112 to 08:00:27:b9:88:74 via eth2

When troubleshooting an instance that is not reachable over the network, it can be helpful to examine this log to verify that all four steps of the DHCP protocol were carried out for the instance in question.

IP¶

The Internet Protocol (IP) specifies how to route packets between hosts that are connected to different local networks. IP relies on special network hosts called routers or gateways. A router is a host that is connected to at least two local networks and can forward IP packets from one local network to another. A router has multiple IP addresses: one for each of the networks it is connected to.

In the OSI model of networking protocols, IP occupies the third layer, which is known as the network layer. When discussing IP, you will often hear terms such as layer 3, L3, and network layer.

A host sending a packet to an IP address consults its routing table to determine which machine on the local network(s) the packet should be sent to. The routing table maintains a list of the subnets associated with each local network that the host is directly connected to, as well as a list of routers that are on these local networks.

On a Linux machine, any of the following commands displays the routing table:

$ ip route show
$ route -n
$ netstat -rn

Here is an example of output from ip route show:

$ ip route show
default via 10.0.2.2 dev eth0
10.0.2.0/24 dev eth0  proto kernel  scope link  src 10.0.2.15
192.168.27.0/24 dev eth1  proto kernel  scope link  src 192.168.27.100
192.168.122.0/24 dev virbr0  proto kernel  scope link  src 192.168.122.1

Line 1 of the output specifies the location of the default route, which is the effective routing rule if none of the other rules match. The router associated with the default route (10.0.2.2 in the example above) is sometimes referred to as the default gateway. A DHCP server typically transmits the IP address of the default gateway to the DHCP client along with the client’s IP address and a netmask.

Line 2 of the output specifies that IPs in the 10.0.2.0/24 subnet are on the local network associated with the network interface eth0.

Line 3 of the output specifies that IPs in the 192.168.27.0/24 subnet are on the local network associated with the network interface eth1.

Line 4 of the output specifies that IPs in the 192.168.122/24 subnet are on the local network associated with the network interface virbr0.

The output of the route -n and netstat -rn commands are formatted in a slightly different way. This example shows how the same routes would be formatted using these commands:

$ route -n
Kernel IP routing table
Destination     Gateway         Genmask         Flags   MSS Window  irtt Iface
0.0.0.0         10.0.2.2        0.0.0.0         UG        0 0          0 eth0
10.0.2.0        0.0.0.0         255.255.255.0   U         0 0          0 eth0
192.168.27.0    0.0.0.0         255.255.255.0   U         0 0          0 eth1
192.168.122.0   0.0.0.0         255.255.255.0   U         0 0          0 virbr0

The ip route get command outputs the route for a destination IP address. From the below example, destination IP address 10.0.2.14 is on the local network of eth0 and would be sent directly:

$ ip route get 10.0.2.14
10.0.2.14 dev eth0  src 10.0.2.15

The destination IP address 93.184.216.34 is not on any of the connected local networks and would be forwarded to the default gateway at 10.0.2.2:

$ ip route get 93.184.216.34
93.184.216.34 via 10.0.2.2 dev eth0  src 10.0.2.15

It is common for a packet to hop across multiple routers to reach its final destination. On a Linux machine, the traceroute and more recent mtr programs prints out the IP address of each router that an IP packet traverses along its path to its destination.

TCP/UDP/ICMP¶

For networked software applications to communicate over an IP network, they must use a protocol layered atop IP. These protocols occupy the fourth layer of the OSI model known as the transport layer or layer 4. See the Protocol Numbers web page maintained by the Internet Assigned Numbers Authority (IANA) for a list of protocols that layer atop IP and their associated numbers.

The Transmission Control Protocol (TCP) is the most commonly used layer 4 protocol in networked applications. TCP is a connection-oriented protocol: it uses a client-server model where a client connects to a server, where server refers to the application that receives connections. The typical interaction in a TCP-based application proceeds as follows:

1. Client connects to server.
2. Client and server exchange data.
3. Client or server disconnects.

Because a network host may have multiple TCP-based applications running, TCP uses an addressing scheme called ports to uniquely identify TCP-based applications. A TCP port is associated with a number in the range 1-65535, and only one application on a host can be associated with a TCP port at a time, a restriction that is enforced by the operating system.

A TCP server is said to listen on a port. For example, an SSH server typically listens on port 22. For a client to connect to a server using TCP, the client must know both the IP address of a server’s host and the server’s TCP port.

The operating system of the TCP client application automatically assigns a port number to the client. The client owns this port number until the TCP connection is terminated, after which time the operating system reclaims the port number. These types of ports are referred to as ephemeral ports.

IANA maintains a registry of port numbers for many TCP-based services, as well as services that use other layer 4 protocols that employ ports. Registering a TCP port number is not required, but registering a port number is helpful to avoid collisions with other services. See Appendix B. Firewalls and default ports of the OpenStack Configuration Reference for the default TCP ports used by various services involved in an OpenStack deployment.

The most common application programming interface (API) for writing TCP-based applications is called Berkeley sockets, also known as BSD sockets or, simply, sockets. The sockets API exposes a stream oriented interface for writing TCP applications: from the perspective of a programmer, sending data over a TCP connection is similar to writing a stream of bytes to a file. It is the responsibility of the operating system’s TCP/IP implementation to break up the stream of data into IP packets. The operating system is also responsible for automatically retransmitting dropped packets, and for handling flow control to ensure that transmitted data does not overrun the sender’s data buffers, receiver’s data buffers, and network capacity. Finally, the operating system is responsible for re-assembling the packets in the correct order into a stream of data on the receiver’s side. Because TCP detects and retransmits lost packets, it is said to be a reliable protocol.

The User Datagram Protocol (UDP) is another layer 4 protocol that is the basis of several well-known networking protocols. UDP is a connectionless protocol: two applications that communicate over UDP do not need to establish a connection before exchanging data. UDP is also an unreliable protocol. The operating system does not attempt to retransmit or even detect lost UDP packets. The operating system also does not provide any guarantee that the receiving application sees the UDP packets in the same order that they were sent in.

UDP, like TCP, uses the notion of ports to distinguish between different applications running on the same system. Note, however, that operating systems treat UDP ports separately from TCP ports. For example, it is possible for one application to be associated with TCP port 16543 and a separate application to be associated with UDP port 16543.

Like TCP, the sockets API is the most common API for writing UDP-based applications. The sockets API provides a message-oriented interface for writing UDP applications: a programmer sends data over UDP by transmitting a fixed-sized message. If an application requires retransmissions of lost packets or a well-defined ordering of received packets, the programmer is responsible for implementing this functionality in the application code.

DHCP, the Domain Name System (DNS), the Network Time Protocol (NTP), and Virtual extensible local area network (VXLAN) are examples of UDP-based protocols used in OpenStack deployments.

UDP has support for one-to-many communication: sending a single packet to multiple hosts. An application can broadcast a UDP packet to all of the network hosts on a local network by setting the receiver IP address as the special IP broadcast address 255.255.255.255. An application can also send a UDP packet to a set of receivers using IP multicast. The intended receiver applications join a multicast group by binding a UDP socket to a special IP address that is one of the valid multicast group addresses. The receiving hosts do not have to be on the same local network as the sender, but the intervening routers must be configured to support IP multicast routing. VXLAN is an example of a UDP-based protocol that uses IP multicast.

The Internet Control Message Protocol (ICMP) is a protocol used for sending control messages over an IP network. For example, a router that receives an IP packet may send an ICMP packet back to the source if there is no route in the router’s routing table that corresponds to the destination address (ICMP code 1, destination host unreachable) or if the IP packet is too large for the router to handle (ICMP code 4, fragmentation required and “don’t fragment” flag is set).

The ping and mtr Linux command-line tools are two examples of network utilities that use ICMP.