Table 1 (Page 2) is used as a motivational experiment to show that the degree to which various amounts of network sharing affect the performance of ping traffic across the network.

Software Required

Experimental Configurations

In this experiment, we deviated from our the standard network configuration and instead used the configurations shown in the diagrams below. However we still used the same hosts as described here.

One host, idle network

Shared sourcehost, shared egress port

Shared destination host, shared ingress port

Shared host, shared ingress and egress

Two hosts, shared switch queue

Software Configuration

In experiments involving ping, we used the following command on the sending host:

sudo ping X.X.X.X -i 0.001 -c 5000 > output

In experiments involving iperf, we used the following command on the sending host:

iperf -s -i 1

And the following command on the receiving host:

iperf -c 10.10.0.9 -t 9999 -i 1

Raw Data

Our original unprocessed datasets can be found here (63kB). There are 5 files in the archive. The file names are:

Each file contains a 1 line header followed by 5000 rows of the form:

64 bytes from 10.10.0.85: icmp_req=1 ttl=64 time=0.347 ms

Followed by a few lines of footer.

Processing the results

The processing scripts can be found here here. To produce the results in the table, run:

./process_pings.sh

which will produce a result something like

Parse errors found: 0
Range errors found: 0
Total samples 5000
0% - 0.022000ms
1% - 0.025000ms
10% - 0.036000ms
25% - 0.038000ms
50% - 0.085000ms
75% - 0.112000ms
90% - 0.118000ms
99% - 0.126000ms
100% - 0.662000ms
Range = 0.640000ms

From wich the 50th and 99th percentiles can easily be found (85us and 126us in this case).