Lambda TCP / UDP Tests after Host Upgrade to 512 MB

• TCP Tests
  • VLAN WCW 1 <=> VLAN EVL
  • VLAN's WCW 1 & 2 <=> VLAN EVL
  • VLAN WCW 1 <=> VLAN WCW 2
• UDP Tests
  • VLAN's WCW 1 & 2 <=> VLAN EVL
• More Result Links:

  • ^^		Before the ONS Upgrade
    ^		Before the Memory Upgrade to 512 MB
    v		After the Linux Upgrade to V. 2.4       v   Loop-back at SARA

  • UDPmon Large
  • Index of All Results

In this document some results are presented obtained with TCP and UDP measurements at the Lambda between SURFnet, Amsterdam, and StarLight, Chicago, as part of the Netherlight project. The results displayed here are obtained after the upgrade of the ONS and after the memory of the hosts located at SARA, Amsterdam, have been upgraded from 128 MB to 512 MB. See also the results before the ONS upgrade and the results before the memory upgrade.

The tests were executed between the following hosts:

For the TCP throughput tests the usual topology has been used with two VLAN's with two hosts each, ranging from the SSR 8000 at SARA, Amsterdam, as far as the LSR 6509 at Chicago. The EVL hosts were located in a third VLAN. The TCP streams at these tests were generated by Iperf, because this tool can easily handle multiple streams. A special script has been used to start the streams at multiple hosts more or less simultaneously using the remote secure shell mechanism. The details of these tests are described in the following sections.

Setup

In this setup throughput tests were performed between the hosts gwgsara2 and gwgsara3, both located in VLAN WCW 1 at SARA and the hosts prusin and reynolds, located at the EVL VLAN. The tests were performed with one or two host pairs. The maximum TCP window size per host was 16 Mbyte. The duration of each test was 60 s.

Results

In the 3D figures, displayed below, the sum of the TCP throughput values, taken over all streams, is presented as a function of the sum of the TCP window size, also taken over all streams, and as a function of the total # streams. Source and destination TCP window sizes are chosen identical.

displays these data for the streams from source host gwgsara3 to destination host reynolds. In these data are presented for the reverse direction. shows the 3D throughput data for the two host pairs in the direction VLAN WCW 1 => VLAN EVL, while in the reverse direction is presented.

.

Sum throughput for one host pair in the direction gwgsara3 => reynolds as a function of the total window size and the # streams.

.

Sum throughput for one host pair in the direction reynolds => gwgsara3 as a function of the total window size and the # streams.

.

Sum throughput for two host pairs in the direction WCW 1 => EVL as a function of the total window size and the # streams.

.

Sum throughput for two host pairs in the direction EVL => WCW 1 as a function of the total window size and the # streams.

In the mean throughput per stream has been given as function of the TCP window size from source host gwgsara3 to destination host reynolds. The data for each # streams are represented in this plot by a trace. supplies these data for the reverse direction. gives the mean throughput data for two host pairs in the VLAN direction WCW 1 => EVL, while in these data are shown for the reverse direction.

.

Throughput per stream for one host pair in the direction gwgsara3 => reynolds as function of the window size per stream.

.

Throughput per stream for one host pair in the direction reynolds => gwgsara3 as function of the window size per stream.

.

Throughput per stream for two host pairs in the direction WCW 1 => EVL as function of the window size per stream.

.

Throughput per stream for two host pairs in the direction EVL => WCW 1 as function of the window size per stream.

From , ...,  the following conclusions can be drawn:

• The performance in the direction EVL => WCW is better than in the reverse direction. This phenomenon is also seen in the tests before the ONS upgrade and before the memory upgrade. It is probably due to host effects.
• There is no improvement in performance when TCP window and socket buffer sizes larger than 4 Mbyte are selected. With the observed round-trip time of nearly 100 ms, an optimal TCP window of about 10 Mbyte might be expected.

Setup

The throughput tests, described in this subsection, were executed from the SARA hosts gwgsara2 in VLAN WCW 1 and gwgsara4 in VLAN WCW 2 to the EVL hosts prusin and reynolds, both situated in the same VLAN. Also measurements in the reverse direction were performed. All tests were executed with two host pairs. The maximum window size per host was 16 Mbyte. The test duration per stream was 60 s.

Results

In the 3D figures below the sum of the TCP throughput has been given for this setup as a function of the TCP window sizes sum and of the total # streams. displays these data in the direction WCW => EVL. In the results of the reverse direction are presented.

.

Sum throughput for two host pairs in the direction WCW 1 & 2 => EVL as a function of the total window size and the # streams.

.

Sum throughput for two host pairs in the direction EVL => WCW 1 & 2 as a function of the total window size and the # streams.

shows the mean throughput per stream as function of the TCP window size for the direction WCW => EVL. In these data are presented for the reverse direction.

.

Throughput per stream for two host pairs in the direction WCW 1 & 2 => EVL as a function of the window size per stream.

.

Throughput per stream for two host pairs in the direction EVL => WCW 1 & 2 as a function of the window size per stream.

Conclusions

Due to the analogous topology the results from the , ...,  can be compared with the results before the memory upgrade, that are presented in the figures 1, 2, 5 and 6. This comparison leads to the following conclusions:

• The maximum throughput in the direction WCW => EVL is after the memory upgrade lower than before, while in the reverse direction the maximum throughput is higher than before, while in the reverse direction the maximum throughput after the upgrade is higher. This behaviour might be caused by an increased difference in host performances after the memory upgrade.
• However, when the maximum throughput values are compared for single host streams, the performance after memory upgrading is better for the direction WCW => EVL and worse for the reverse direction. Different window sampling in both situations may play a role: the line slope in the graphs of the throughput per stream as function of the TCP window size per stream is considerable.
• Comparable conclusion as in the "Conclusions" subsection of the "VLAN WCW 1 <=> VLAN EVL" section can be drawn.

Setup

In this section the results are presented from throughput tests between two SARA hosts, i.e. gwgsara2, located in VLAN WCW 1 and gwgsara4, located in VLAN WCW 2. This implies that the traffic between these two hosts and VLAN's is following the path Amsterdam - Chicago - Amsterdam. Again the maximum TCP window and socket sizes per host were 16 Mbyte and the test duration was 60  seconds.

Results

In the sum of the TCP throughput has been presented as a function of the sum over the TCP window size and the total # streams. displays these data for the reverse direction.

.

Sum throughput for one host pair in the direction WCW 1 => WCW 2 as a function of the total window size and the # streams.

.

Sum throughput for one host pair in the direction WCW 2 => WCW 1 as a function of the total window size and the # streams.

gives the mean throughput per stream as function of the TCP window size for the direction WCW 1 => WCW 2. In these data are displayed for the reverse direction.

.

Throughput per stream for one host pair in the direction WCW 1 => WCW 2 as a function of the window size per stream.

.

Throughput per stream for one host pair in the direction WCW 2 => WCW 1 as a function of the window size per stream.

Conclusions

Because hosts were being upgrade to Linux 2.4 at that moment, only hosts gwgsara2 and gwgsara4 were available for these tests, so that only one host pair could be used. Unfortunately that makes the comparison with the results before the update difficult, because from those tests only results with two host pairs were available.

But also in these tests presented by the , ...,  the bad performance at the route Amsterdam - Chicago - Amsterdam has been observed again. We do not have a good explanation for this poor performance.

For the UDP bandwidth tests the same topology and tools has been used as described in the "TCP Tests" section. The details of the UDP tests are described in the following sections. Also for these tests Iperf was used.

Setup

In this setup UDP bandwidth tests were performed from the SARA hosts gwgsara2 in VLAN WCW 1 and gwgsara4 in VLAN WCW 2 to the EVL hosts prusin and reynolds, both situated in the same VLAN. Also measurements in the reverse direction were performed. All tests were executed with two host pairs. The sum of the bandwidth to be send over the streams was maximal 1500 Mbit/s. The duration of each test was 60 s.

Results

displays the percentage total packets lost as a function of the total bandwidth send in the direction WCW 1 & 2 => EVL. In the results for the reverse direction are presented. In both figures the data valid for a particular # streams are represented by a separate plot trace.

.

Total packets lost for two host pairs in the direction WCW 1 & 2 => EVL as a function of the sum of the bandwidths.

.

Total packets lost for two host pairs in the direction EVL => WCW 1 & 2 as a function of the sum of the bandwidths.

Conclusions

Comparison with the UDP bandwidth results before the memory results, described there in the corresponding "Setup" section is not well possible, because at the time the UDP tests were executed no three hosts with 512 Mbyte memory were not yet available to do the tests as described in that document.

But from the there can be concluded:

• The observed bandwidths without packets lost seem to be limited by the hosts, not the Cisco ONS.
• The better performance in the direction EVL => WCW 1 & 2 can be understood by the better performance of the WCW hosts.