Master Dissertation
Master Dissertation
Scenario: ARP as server, COP e SOLDIER as clients. 100Mbps network, messages of 500 bytes.
| Simultaneous | Client | Protocol | Throughput (kbps) | Latency (µs) |
|---|---|---|---|---|
| No | SOLDIER | TCPM | 16584 | 421 |
| SCTP | 15520 | 516 | ||
| COP | TCPM | *9176 | 425 | |
| SCTP | 15448 | 523 | ||
| yes | SOLDIER | TCPM | 4816 | 490 |
| SCTP | 7736 | 681 | ||
| COP | TCPM | 4792 | 490 | |
| SCTP | 7784 | 681 |
Server is a lot slower than either client, so the results will not be congruent with former tests. So we remade one-client tests for comparision purposes.
In an ideal world, total throughput should be always the same regardless of client count, and it should be evenly and fairly divided among all. Also, throughput in both directions should be equal (since network is full-duplex). SCTP was more successful in this than TCPM.
The TCPM result marked with an asterisk (which is fairly bad) was caused by unbalanced throughput between the two directions; one direction finished well before the other (i.e. it could not use the full-duplex network efficiently), and since the final average is based on "wall clock" execution time, it resulted low.
Also, TCPM presented lower throughputs when two clients run simultaneously.
In latency tests, TCPM latency increased only 15% with two clients instead of one. SCTP latency increased 30%, so it is more sensible to client count, possibly because it uses more CPU, and ARP server is quite slow.