Lambda TCP Tests after Host Upgrade to Linux 2.4

• TCP Tests
  • VLAN WCW 1 & 2 <=> VLAN EVL
  • VLAN WCW 1 <=> VLAN WCW 2
  • Bi-Directional Tests VLAN WCW 1 & 2 <=> VLAN EVL
• More Result Links:

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

  • UDPmon Large
  • Index of All Results

In this document some results are presented obtained with TCP 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.
• the memory of the hosts located at SARA, Amsterdam, have been upgraded from 128 MB to 512 MB.
• the Linux kernel have been upgraded from V. 2.2.15 to V. 2.4.16-web100.

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

The throughput tests, described in this subsection, were executed from the SARA hosts gwgsara3 in VLAN WCW 1 and gwgsara5 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. The tests were executed with one and 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, taken over all streams, is presented this setup as a function of the sum of the TCP window size, also taken over all streams, and of the total # streams. Source and destination TCP window sizes are chosen identical.

displays these data for the streams from source host gwgsara5 to destination host reynolds. In these data are given for the reverse direction. shows these 3D throughput data for the two host pairs in the direction VLAN WCW 1 & 2 => VLAN EVL, while in these data are presented for the reverse direction.

.

Sum throughput for one host pair in the direction gwgsara5 => reynolds as a function of the total window size and the # streams.

.

Sum throughput for one host pair in the direction reynolds => gwgsara5 as a function of the total window size and the # streams.

.

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.

In the figures below the throughput per stream as a function of the TCP window size is displayed. The data for each # streams are represented in this plot by a separate trace.

displays these data for the streams from source host gwgsara5 to destination host reynolds. In these data are given for the reverse direction. shows the throughput per stream data for the two host pairs in the direction VLAN's WCW 1 & 2 => VLAN EVL, while in these data are presented for the reverse direction.

.

Throughput per stream for one host pair in the direction gwgsara5 => reynolds as function of the window size per stream.

.

Throughput per stream for one host pair in the direction reynolds => gwgsara5 as function of the window size per stream.

.

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

.

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

Conclusions

The results in the , ...,  can be compared with the corresponding results before the Linux upgrade, presented there by the figures 1, 2, 5, 6, 9, ..., 12. Comparison of both figures series leads to the following conclusions:

• The behaviour of the Linux 2.4 leads to a more regular behaviour of the throughput as function of the TCP window size.
• However, the performance with the Linux 2.4 kernel seems to be lower than with the 2.2 kernel. But the tests were performed at different days, so external influences cannot be excluded.

Setup

In this section the results are presented from throughput tests between two SARA hosts, i.e. gwgsara3, located in VLAN WCW 1 and gwgsara5, located in VLAN WCW 2. This implies that the traffic between these two hosts and VLAN's is following the route Amsterdam - Chicago - Amsterdam. Also in these tests 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

The results in the , ...,  can be compared with the corresponding results before the Linux upgrade, presented there by the figures 13, ..., 16. From both figures series the following conclusions can be drawn:

• The sum throughput plots after the Linux upgrade ( and ) show a better and more consistent performance as in the equivalent figures before the upgrade (figures 13 and 14).
• But the plots with the throughput per stream after the Linux upgrade ( and ) shows a throughput per stream which is in general less than the maximum values in the corresponding figures before the upgrade (figures 15 and 16), but the throughput values before are more consistent, explaining the overall better throughput performance mentioned above.
• Also in these results the tests before and after the Linux upgrade were executed at different days, so daily variations also might have played a role here.

Setup

This subsection describes the results of bi-directional tests between the VLAN's WCW 1 & 2 and the EVL VLAN. In these tests the streams  prusin => gwgsara3  and  gwgsara5 => reynolds  were used. Again the maximum window size per host was 16 Mbyte and the test duration for each stream was 60 s.

Results

displays the sum of the TCP throughput as a function of the sum over the TCP window size and the total # streams. In the average throughput per stream has been given as a function of the TCP window size.

.

Sum throughput from the bi-directional streams  prusin => gwgsara3  and  gwgsara5 => reynolds  as a function of the total window size and of the # streams.

.

Throughput per stream for the bi-directional streams  prusin => gwgsara3  and  gwgsara5 => reynolds  as a function of the window size per stream.

Conclusions

When the bi-direction results from are compared with the corresponding results for two host pairs from section "VLAN WCW 1 & 2 <=> VLAN EVL" ( and ) it follows that the performances of both tests are about the same, reflecting the full duplex character of the connection. But note that also daily differences may play a role here.