SL - ASP Itanium Tests -- Rate Limit -- TCP Profile Tests

TCP Profile Tests

Description

In this section the results of TCP profile tests will be presented. The TCP bandwidth will be shaped according to the triangle profile that has been shown in figure 1 from the "Rate Limit" main section. The shaped TCP profile has been generated with a modified version of the Iperf V. 1.6.5 traffic generator. Also in this version the pthread library will be used to generate multiple parallel flows.

Results

In the following figures the resulting TCP bandwidth values, induced by the provided shaping profile bandwidth, have been displayed as a function of the profile time. Both the client and server bandwidth results are presented in these figures with separate plot traces. The results for various values of the top triangle bandwidth have been shown by individual plots. In these tests a TCP window size of 2 MBytes have been used. Only results are presented for the direction where the rate limiting is active, also for: Amsterdam Science Park -> StarLight.

In the these results are presented for top triangle bandwidths of 50, 100, ..., 500 Mbits/s with a single TCP flow from source to destination.

. Single flow TCP profile throughput as a function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size is 2 MBytes/s, and the top triangle bandwidth is 50 Mbits/s. The client and server throughput values have been presented with separate plot traces.

. Single flow TCP profile throughput as a function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size is 2 MBytes/s, and the top triangle bandwidth is 100 Mbits/s. The client and server throughput values have been presented with separate plot traces.

. Single flow TCP profile throughput as a function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size is 2 MBytes/s, and the top triangle bandwidth is 150 Mbits/s. The client and server throughput values have been presented with separate plot traces.

. Single flow TCP profile throughput as a function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size is 2 MBytes/s, and the top triangle bandwidth is 200 Mbits/s. The client and server throughput values have been presented with separate plot traces.

. Single flow TCP profile throughput as a function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size is 2 MBytes/s, and the top triangle bandwidth is 250 Mbits/s. The client and server throughput values have been presented with separate plot traces.

. Single flow TCP profile throughput as a function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size is 2 MBytes/s, and the top triangle bandwidth is 300 Mbits/s. The client and server throughput values have been presented with separate plot traces.

. Single flow TCP profile throughput as a function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size is 2 MBytes/s, and the top triangle bandwidth is 350 Mbits/s. The client and server throughput values have been presented with separate plot traces.

. Single flow TCP profile throughput as a function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size is 2 MBytes/s, and the top triangle bandwidth is 400 Mbits/s. The client and server throughput values have been presented with separate plot traces.

. Single flow TCP profile throughput as a function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size is 2 MBytes/s, and the top triangle bandwidth is 450 Mbits/s. The client and server throughput values have been presented with separate plot traces.

. Single flow TCP profile throughput as a function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size is 2 MBytes/s, and the top triangle bandwidth is 500 Mbits/s. The client and server throughput values have been presented with separate plot traces.

Show also the plots from the as in a new browser window.

In the corresponding results have been displayed as in the previous figures, but here two parallel flows were used. Both for the top triangle bandwidths and for the presented throughput values the sums have been taken over the parallel flows. To be able to take this sum, the throughput interval reports have been re-sampled according to the original profile intervals.

. TCP throughput sum, taken over two parallel flows, as function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size sum is 2 MBytes/s, and the top triangle bandwidth sum is 50 Mbits/s. Both sums are also taken over the parallel flows. The client and server throughput values have been presented with separate plot traces.

. TCP throughput sum, taken over two parallel flows, as function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size sum is 2 MBytes/s, and the top triangle bandwidth sum is 100 Mbits/s. Both sums are also taken over the parallel flows. The client and server throughput values have been presented with separate plot traces.

. TCP throughput sum, taken over two parallel flows, as function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size sum is 2 MBytes/s, and the top triangle bandwidth sum is 150 Mbits/s. Both sums are also taken over the parallel flows. The client and server throughput values have been presented with separate plot traces.

. TCP throughput sum, taken over two parallel flows, as function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size sum is 2 MBytes/s, and the top triangle bandwidth sum is 200 Mbits/s. Both sums are also taken over the parallel flows. The client and server throughput values have been presented with separate plot traces.

. TCP throughput sum, taken over two parallel flows, as function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size sum is 2 MBytes/s, and the top triangle bandwidth sum is 250 Mbits/s. Both sums are also taken over the parallel flows. The client and server throughput values have been presented with separate plot traces.

. TCP throughput sum, taken over two parallel flows, as function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size sum is 2 MBytes/s, and the top triangle bandwidth sum is 300 Mbits/s. Both sums are also taken over the parallel flows. The client and server throughput values have been presented with separate plot traces.

. TCP throughput sum, taken over two parallel flows, as function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size sum is 2 MBytes/s, and the top triangle bandwidth sum is 350 Mbits/s. Both sums are also taken over the parallel flows. The client and server throughput values have been presented with separate plot traces.

. TCP throughput sum, taken over two parallel flows, as function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size sum is 2 MBytes/s, and the top triangle bandwidth sum is 400 Mbits/s. Both sums are also taken over the parallel flows. The client and server throughput values have been presented with separate plot traces.

. TCP throughput sum, taken over two parallel flows, as function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size sum is 2 MBytes/s, and the top triangle bandwidth sum is 450 Mbits/s. Both sums are also taken over the parallel flows. The client and server throughput values have been presented with separate plot traces.

. TCP throughput sum, taken over two parallel flows, as function of the profile time for the test direction with rate limiting Amsterdam Science Park -> StarLight. The used TCP window size sum is 2 MBytes/s, and the top triangle bandwidth sum is 500 Mbits/s. Both sums are also taken over the parallel flows. The client and server throughput values have been presented with separate plot traces.

Show also the plots from the as in a new browser window.

Conclusions

From the TCP profile tests presented in the the following conclusions can be drawn:

In the single flow results () the following patterns can be seen:
- For the top triangle bandwidth values of 50 and 100 Mbits/s, ( ) the expected theoretical profile is also be found. The reason for that could be:
  - The achieved maximum bandwidth is below the theoretical limit induced by the used TCP window size of 2 MBytes/s which should lead with a round-trip time of about 100 ms to a maximum bandwidth of about 0.15 Gbits/s.
  - Also the rate limiting obviously does not play a role yet.
- For the top triangle bandwidth values of 150, 200 and 250 Mbits/s (, ) a saturation level of about 0.11 Gbits/s could be found. This is in well agreement with the bandwidth limit induced my the TCP window size mentioned above.
- For the results with a top triangle bandwidth ≥ 300 Mbits/s () the maximum achieved throughput falls back below 0.02 Gbits/s. This presumably has been caused by the rate limit settings. This value is in good agreement with the results for a single flow in the TCP bulk results. See figure 1 in the "TCP Bulk Tests" section.
For the results with two parallel flows () about the same conclusions can be drawn as for the results obtained with a single flow. The most important difference here is that the fall back caused by the rate limiting effects is starting at larger values of the profile times. Because the top triangle bandwidths per flow are here the half from the corresponding single flow values, there could be less packets lost caused by busts that are related with the shaped bandwidth height.
In general the values for the client and server bandwidths are identical as could be detected. When there are differences they probably are related with differences in the domain size of the sample interval, eventually combined with our re-sampling to the original sample intervals. Please note that the domain sizes of the Iperf sampling intervals are in general becoming irregular at the areas with packets lost.