Force10 Loop-Back -- UDT & TCP -- Bandwidth Interval Plots

Bandwidth Interval Plots

In the following sub-sections results are displayed in the form of UDT and TCP bandwidth plots as a function of the absolute interval time of the network test which implies that also the delay times are included at the interval times.

Shaped Bandwidth Variation

In the plots of the bandwidth interval reports from the UDT and TCP flows are presented as a function of the absolute interval time of the network test, also with delay times included. In the left plots the UDT flows are delayed and in the right plots the TCP flows. In the TCP plot trace the markers are denoting the overall sum of the TCP flows. The standard deviation in the TCP sum is denoted by the error-bars, representing the variation in the TCP bandwidths. These tests were executed with one parallel flow per node pair. There were 11 node pairs used in each VLAN. Each node had been connected with all nodes in the other VLAN, so 11 × 11 different TCP flows were used. Each figure shows the value of a different shaped TCP bandwidth varying from 2 Mbits/s () to 10 Mbits/s ().

UDT, Sleep TCP; P = 1, N = 121, B = 2 Mbits/s

. Single flow UDT and multiple flows TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 11 × 11 node pairs, using one parallel flow per node pair. The shaped TCP bandwidth was 2 Mbits/s.

UDT, Sleep TCP; P = 1, N = 121, B = 4 Mbits/s

. Single flow UDT and multiple flow TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 11 × 11 node pairs, using one parallel flow per node pair. The shaped TCP bandwidth was 4 Mbits/s.

UDT, Sleep TCP; P = 1, N = 121, B = 6 Mbits/s

UDT, Sleep TCP; P = 1, N = 121, B = 8 Mbits/s

UDT, Sleep TCP; P = 1, N = 121, B = 10 Mbits/s

. Single flow UDT and multiple flow TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP tests were defined by all possible 11 × 11 node pairs, using one parallel flow per node pair. The shaped TCP bandwidth was 10 Mbits/s.

In the equivalent results are displayed as in the with the exception that there are now five threaded parallel TCP flows per node pair used while in the previous situation only a single TCP flow per node pair had been used with the constraint that the sum over all parallel flows per node pair had been kept the same in both situations.

UDT, Sleep TCP; P = 5, N = 121, Sum B = 2 Mbits/s

. Single flow UDT and multiple flow TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 11 × 11 node pairs, using five parallel flows per node pair. The sum of the shaped TCP bandwidth over the parallel flows per node pair was 2 Mbits/s.

UDT, Sleep TCP; P = 5, N = 121, Sum B = 4 Mbits/s

. Single flow UDT and multiple flow TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 11 × 11 node pairs, using five parallel flows per node pair. The sum of the shaped TCP bandwidth over the parallel flows per node pair was 4 Mbits/s.

UDT, Sleep TCP; P = 5, N = 121, Sum B = 6 Mbits/s

. Single flow UDT and multiple flow TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 11 × 11 node pairs, using five parallel flows per node pair. The sum of the shaped TCP bandwidth over the parallel flows per node pair was 6 Mbits/s.

UDT, Sleep TCP; P = 5, N = 121, Sum B = 8 Mbits/s

. Single flow UDT and multiple flow TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 11 × 11 node pairs, using five parallel flows per node pair. The sum of the shaped TCP bandwidth over the parallel flows per node pair was 8 Mbits/s.

UDT, Sleep TCP; P = 5, N = 121, Sum B = 10 Mbits/s

. Single flow UDT and multiple flow TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 11 × 11 node pairs, using five parallel flows per node pair. The sum of the shaped TCP bandwidth over the parallel flows per node pair was 10 Mbits/s.

Comparing the single flow per node pair results () with the tests were five threaded parallel flows per node pair were used () shows that there are not much differences between both test types. In the regions with combined UDT and TCP traffic, there is a small decrease in the sum of the TCP bandwidth for shaped bandwidths ≥ 8 Mbits/s, but when five parallel flows are used hardly any influence from the UDT flow could be noticed. Besides this difference the results of both tests are well comparable. The observed bandwidths are in good agreement what could be expected from the shaped bandwidth values.

Please note that the unrealistic high bandwidth values from the highest interval report is an artifact from the Iperf tool. These values were also found in the raw Iperf data. Please note also that the Iperf and UDT boundaries of the bandwidth reports do not exactly match with the used interval time. To be able to sum the TCP bandwidths, the intervals are re-sampled back to the time stamps following from the used interval time. To be consistent the same re-sampling algorithm has been used for all results, so also for the single flow results.

# TCP Flows Variation

As before the are presenting bandwidth interval report plots as a function of the absolute interval time, but in these plots the # TCP flows, defined by the # source - destination node pairs, will be varied. These results are displayed for single parallel flows and with a shaped TCP bandwidth of 10 Mbits/s.

UDT, Sleep TCP; P = 1, N = 1, B = 10 Mbits/s

. Single flow UDT and (here also) single flow TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flow. The TCP flow was defined by the only possible 1 × 1 node pair, using one parallel flow. The shaped TCP bandwidth was 10 Mbits/s.

UDT, Sleep TCP; P = 1, N = 4, B = 10 Mbits/s

. Single flow UDT and multiple flows TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 2 × 2 node pairs, using one parallel flow per node pair. The shaped TCP bandwidth was 10 Mbits/s.

UDT, Sleep TCP; P = 1, N = 9, B = 10 Mbits/s

. Single flow UDT and multiple flows TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 3 × 3 node pairs, using one parallel flow per node pair. The shaped TCP bandwidth was 10 Mbits/s.

UDT, Sleep TCP; P = 1, N = 16, B = 10 Mbits/s

. Single flow UDT and multiple flows TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 4 × 4 node pairs, using one parallel flow per node pair. The shaped TCP bandwidth was 10 Mbits/s.

UDT, Sleep TCP; P = 1, N = 25, B = 10 Mbits/s

. Single flow UDT and multiple flows TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 5 × 5 node pairs, using one parallel flow per node pair. The shaped TCP bandwidth was 10 Mbits/s.

UDT, Sleep TCP; P = 1, N = 36, B = 10 Mbits/s

. Single flow UDT and multiple flows TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 6 × 6 node pairs, using one parallel flow per node pair. The shaped TCP bandwidth was 10 Mbits/s.

UDT, Sleep TCP; P = 1, N = 49, B = 10 Mbits/s

. Single flow UDT and multiple flows TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 7 × 7 node pairs, using one parallel flow per node pair. The shaped TCP bandwidth was 10 Mbits/s.

UDT, Sleep TCP; P = 1, N = 64, B = 10 Mbits/s

. Single flow UDT and multiple flows TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 8 × 8 node pairs, using one parallel flow per node pair. The shaped TCP bandwidth was 10 Mbits/s.

UDT, Sleep TCP; P = 1, N = 81, B = 10 Mbits/s

. Single flow UDT and multiple flows TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 9 × 9 node pairs, using one parallel flow per node pair. The shaped TCP bandwidth was 10 Mbits/s.

UDT, Sleep TCP; P = 1, N = 100, B = 10 Mbits/s

. Single flow UDT and multiple flows TCP bandwidth as a function of the absolute test time. In the left plot the UDT flow has been delayed and in the right plot the TCP flows. The TCP flows were defined by all possible 10 × 10 node pairs, using one parallel flow per node pair. The shaped TCP bandwidth was 10 Mbits/s.

UDT, Sleep TCP; P = 1, N = 121, B = 10 Mbits/s

Conclusions

From the results in the following conclusions can be drawn:

With zero roundtrip time the UDT flow is almost completely adjusting itself to the multiple Iperf flows which are almost not influenced by it.
There are no significant differences between the tests where the UDT flow is started first and the reverse situation where the Iperf flows are started first. The (almost) zero round-trip times are probably the reason for that.
There is almost no variation in the Iperf bandwidth values. Presumably that will also be connected with the zero round-trip time values.
As stated before: in the shaped bandwidth variation tests there is not much difference between the Iperf tests with a single flow per node pair () compared with the tests with five parallel flows per node pair ().
The unrealistic high bandwidth values from the Iperf bandwidth reports in the largest interval is an artifact from that tool. These values were also found in the raw Iperf data and are especially true for parallel flows. May be the time values listed in the last are underestimating the end time. This would imply that some flows are ending later. The result would be that a too high sum bandwidth will be suggested at the end of the interval reports.

UDT & TCP Flows

Table of Contents

Shape Bandwidth Plots

RBUDP & TCP Flows