Kernel Korner - Analysis of the HTB Queuing Discipline

In order to test the accuracy and performance of the HTB, we used the following pieces of network equipment: one Ixia 400 traffic generator with a 10/100 Mbps Ethernet load module (LM100TX3) and one Pentium 4 PC (1GB RAM, 70GB hard drive) running a 2.6.5 Linux kernel. Two testing models were designed, one to test policing accuracy and one to test bandwidth sharing.

The first model (Figure 3) was used for testing the policing accuracy of a specific defined class. Port 1 in the Ixia machine generated traffic sent to IP 192.168.10.200 from one or more streams. The Linux machine routed the packets to interface eth0 (static route) and then sent them back to the Ixia machine on Port 2. All of the traffic control attributes were defined on the eth0 interface. All of the analysis was completed based on traffic results captured on Port 2 (the Ixia machine).

Figure 3. Test Model #1 Configuration

The second model (Figure 4) was used to test the way the bandwidth of two streams from the same class is shared. In this case, another two Ixia ports for transmitting data were used.

Figure 4. Test Model #2 Configuration

Port 1, Port 2 and Port 3 in the Ixia machine generated traffic sent to IP 192.168.10.200, each using one stream. The Linux machine routed those packets to interface eth0 based on a static route and then sent them back to the Ixia machine on Port 2. Traffic control attributes were defined on the eth0 interface. All analysis was done based on the traffic result captures on Port 2 (Ixia machine).

In all of the tests, the sending ports transmitted continuous bursts of packets on a specified bandwidth. The Ixia 10/100 Mbps Ethernet load module (model LM100TX3) has four separate ports, and each port can send up to 100Mbit/s. The Ixia load module provided support for generating multiple streams in one port but with one limitation: it couldn't mix the streams together and served only one stream at a time. This limitation exists because the scheduler works in a round-robin fashion. It sends a burst of bytes from stream X, moves to the next stream and then sends a burst of bytes from stream Y.

In order to generate a specific bandwidth from a stream, which is part of a group of streams defined in one port, specific attributes of the Ixia machine's configuration had to be fine-tuned. The attributes that required fine-tuning and their definitions are as follows:

See Table 2 for Ixia configuration details.

The goal was to determine the appropriate burst size that would achieve the requested generated bandwidth for each stream. Because all three streams used the same physical line, the way the data was sent on the line resembles the illustration in Figure 5.

Figure 5. Data as sent on the line from the Ixia machine to the Linux system being tested.

The following equations define the relationship between the attributes:

Table 3 explains the variables used in the equation.

Assuming that the packet size is the same for all streams, as in the example, the remaining calculation is that of the burst size.

Because all the streams share the same bandwidth, the requested burst values can be found by examining the ratios between the requested bandwidths, using the equation B^s-i = B^n-i. This number could be unusually large, though, so it can be divided until a reasonable value is obtained. In order to have different packet sizes defined for each stream, the burst size values can be altered until the required bandwidth is obtained for each stream. A spreadsheet program simplifies the calculation of multiple bandwidths.