Bandwidth

Bandwidth

Bandwidth is the amount of data that can be passed along a communications channel in a given period of time. Specifically, each message that is exchanged between two nodes communicating on a network requires that some network capacity, or bandwidth, be available to allow the movement of information between the two nodes.

In a LAN environment where nodes are connected to the same switch or in a campus network where nodes are connected within proximity to one another, available bandwidth is generally much higher than that required by the two communicating nodes. However, as nodes become spread across a larger, more complex internetwork, network oversubscription might be encountered, or points of aggregation within the network might be encountered. These two factors have a significant impact on application performance, because the point of oversubscription or point of aggregation must effectively throttle transmission from attached networks and negotiate access from the higher-capacity network to the lower-speed, oversubscribed network.

Figure 2-3 shows multiple layers of oversubscription. Assuming each node is connected with the same data-link interface type and speed (for this example, assume 100-Mbps Fast Ethernet), the four client workstations contend for the available capacity between the access switch and the distribution switch. The three access switches must contend for available capacity between the distribution switch and the server. All 12 of the client workstations must contend through these oversubscribed links for access to the server itself.

Figure 2-3. Oversubscription at Various Points Within the Network

[View full size image]

Assuming fairness between equally prioritized connections, each client would receive only approximately 8 percent of the server's network capacity, and this does not take into consideration the overhead associated with application, transport, or network layer mechanics.

This problem is compounded in WAN environments where the oversubscription or aggregation might be dramatically higher. In the case of a WAN, not only do multiple network nodes contend for access to available bandwidth, but the available bandwidth is many orders smaller than that by which the node is communicated to the LAN. In this way, not only does the network element need to manage access to the smaller bandwidth link, but it must also handle the pacing of traffic onto the node through queuing.

Figure 2-4 shows multiple locations each with multiple client workstations. These client workstations are connected by the same data-link interface type and speed as the servers and switches (for this example, assume 100-Mbps Fast Ethernet). The routers in this example connect to the LAN in each location and also a point-to-point T1 (1.544 Mbps) circuit. Not only do the client workstations in each of the remote offices have to contend with 4:1 oversubscription to the router, but the bandwidth disparity in the network also introduces another layer of oversubscription: 67:1 (100 Mbps / 1.544 Mbps).

Figure 2-4. Oversubscription in a WAN Environment

[View full size image]

Assuming fairness between equally prioritized connections, each client would receive only approximately 25 percent of the WAN capacity, which equates to 384 Kbps, or roughly 1/260, of the link speed to which the client is attached to the LAN. This also assumes that there is no overhead associated with the application protocol, transport protocol, and network protocols being used.

The performance challenges created by bandwidth disparity, oversubscription, and aggregation can quickly degrade application performance. Not only are each of the communicating nodes strangled because their flows are negotiated onto a lower-speed link, but the flows themselves remain strangled even when returning to a higher-capacity link, as shown in Figure 2-5. From the perspective of the server, each of the clients is only sending small amounts of data at a time (due to network oversubscription and bandwidth disparity). The overall application performance is throttled because the server can then only respond to the requests as they are being received.

When the server responds to the client, the same challenges are encountered, as shown in Figure 2-6. Potentially large amounts of data can be sent from the server when it services a user request. This data, when set in flight, encounters the neighboring router, which is managing the bandwidth disparity, and the data is trickled over the network based on the capacity available to the flow. From the perspective of the client, the server is only sending small amounts of data at a time (due to network oversubscription and bandwidth disparity).

The prior examples assume a perfect world where everything works as advertised and overhead is unheard of. The reality is that application protocols add overhead to message exchanges. Transport protocols add overhead in the form of segmenting, window management, and acknowledgment for guaranteed delivery. Network and data link protocols add overhead due to packetization and framing. Each of these consumes a noticeable amount of network capacity and can be classified as control information that only serves the purpose of helping data reach the distant node, reassembling data in the correct order, and informing the application on the distant node of the process that is being attempted.

This process of exchanging application data using a transport protocol over an internetwork with potentially many segments between communicating nodes directly follows the Open System Interconnection (OSI) reference model, which is outlined in Table 2-1.

The application reads and writes blocks of data to a socket interface, which abstracts the transport protocol itself from the application. The blocks of data vary in size based on the application and the amount of memory allocated to the socket buffers. Application data blocks are generally 1 KB to 64 KB in size, and in some cases can be larger or smaller. Notice that an application data block is measured in bytes and not bits, as from the perspective of an application it is simply reading or writing block data to or from a buffer in memory.

The transport protocol is then responsible for draining data from the socket buffer into the network layer. When the data is written to the socket buffer by the application layer, the socket interacts with the transport protocol to segment data into datagrams. These datagrams are sized with the knowledge of the network transmission capacity (based on a search, discovery, or predefined parameter if using rate-based transmission protocols). For connection-oriented transmission protocols such as TCP, control information such as source and destination ports is attached along with other information, including cyclic redundancy check (CRC)/checksum information, segment length, offset, sequence number, acknowledgment number, and other flags. Most other transport protocols, reliable or not, provide additional data that helps to identify the application process on the local and distant nodes, along with checksum information to provide some means of guaranteeing data integrity and correctness, as well as other flow-control parameters.

The transport protocol manages both the drainage of data written to socket buffers into segments of data exchanged on the network and the extraction of data received from the network into socket buffers to be read by a local application. When data has been read from the socket buffer and processed by TCP (it is ready to be handled by the network layer), it is packetized by the network protocol (in this case, IP), where a header and trailer are added denoting the source and destination network address on the internetwork. Other information might be included such as version, header length, type of service, fragment identification, header checksums, protocol, options, and other flags.

The network protocol (generally IP) then sends the packet to be framed at the data link layer (generally Ethernet), where yet another header and trailer are added denoting the source and destination address of the next adjacent recipient on the subnet (router) and other flags. Each of these layers commonly adds checksum or CRC data to ensure correctness of data.

Figure 2-7 shows the overhead associated with application data traversing the OSI model.

Figure 2-7. Framing, Packetization, Segmentation, and Application Layer Overhead

[View full size image]

When the distant node receives the packets, the reverse process begins in order to reassemble the application data block. In many cases, as much as 20 percent or more of network bandwidth can be consumed by forwarding, routing, segmentation, and delivery guarantee-related overheads. To put it simply, the WAN, which introduces a massive bandwidth disparity, coupled with the overhead just described in this section, is not an environment conducive to high levels of application performance.