CPU Twinhead Subnote(486 33MHz)
OS BSD/OS V2.1
Buffer Size 8196byte
Message Size 2048byte
Number of Send(DBS,ttcp) 2048
Time of Send(Netperf) 10s
Throughput
Result
Mean Throughput
DBS Throughput=100
DBS
6.22Mbps
100
Netperf
6.36Mbps
102
ttcp
6.17Mbps
99
Changes of the throughput between two hosts in this benchmark are depicted as a
graph shown in Figure 5. This graph is a plot of the mean throughput at the Host2
(TCP segment receiver) in every 50ms. This can be calculated easily from the log files
generated by the DBS. As a reference, the mean throughput obtained by both the
Netperf and the ttcp are also plotted in the same graph. As shown in the figure, the
throughput is 0Kbps from its start time (0ms) to about 200ms. This 0Kbps period is
caused by both the sender's Slow-Start algorithm and the receiver's Delayed ACK
technique. At the beginning of the benchmark, the Host1 immediately sends only a
single TCP segment to the Host2 because of its Slow-Start mechanism. After a very
short period (about 1ms), the Host2 completely receives the segment. Because of the
TCP's Delayed ACK mechanism, the Host2 does not reply an ACK back to the Host1
instantly, and it waits for the next TCP segment. However, by the sender's Slow-Start
mechanism, the Host1 is also waiting an ACK from the Host2. Thus, these two hosts
fall into the deadlock situation. However, this deadlock is resolved by the Delayed
ACK timer mechanism. When the first segment was received by the Host2, it enabled
the Delayed ACK timer. When the Delayed ACK timer (UNIX fast timer) is expired,
the Host2 replied the ACK back to the Host1. This is the mechanism where the 0Kbps
period was generated. It is obvious that only the DBS can help this kind of analysis,
because only the DBS has a function to evaluate changes of the throughput.