Network Configuration Pros and Cons

In this topic we discuss the advantages and disadvantages of the three evaluated Network Configurations.

Which of the tested Networking Configurations is the appropriate choice for you? That depends on a number of factors, like highest throughputs, less CPU consumption, better CPU efficiency, easier configuration, usability, isolation and more.

To help compare and contrast characteristics and behavior of the tested network configurations, graphs have been produced comparing results from two separate workload runs. Recall that each run consists of 20 different tests (described in Workload tests in Network workload). Each graph compares the 20 workload test results from one run to the results of a separate run.

The graphs show the differences between the two compared runs. In the graphs below the 20 different workload tests are represented on the X-axis and the difference (in %) of 4 different performance characteristics are shown on the Y-axis. All of the performance characteristics have been normalized to show the difference of a second run (B) when compared to a first run (A). The titles of each graph take the form of “run B config” vs “run A config”.

If the result from run B is better than run A, the bar on the graph will be positive.
If run B's result is worse than run A's result, the bar will be negative. In all the graphs a positive bar means a better result.

For throughput, (the blue bar), a better results means a higher absolute value.

For transaction times (the orange bars), better equates to shorter time. Transaction time is equivalent to the overall latency of network transaction. The faster (or shorter) amount of time a transaction completes the better, so better is less overall time.

The yellow and green bars represent CPU efficiency for the KVM guest running the uperf client (yellow) and KVM guest running the uperf server (green). These bars compare the amount of CPU required to transfer a megabyte of data. For CPU efficiency, a better result or a positive bar means less CPU was required to transfer the same amount of data.

Note: All of the graphs that follow are from test results collected from runs using 4 uperf pairs (see uperf pairs in Network workload) which use 8 KVM guests running on 1 or 2 LPAR configurations.