NOTE

Phase 1 Performance Test on Static Routing

In my experiments, if not otherwise specified, the network infrastructure is always:

Experiment 1 always refers to unrouted connection:

Madrid:eth2( 192.168.3.1 ) ======> Prague:eth2( 192.168.3.2 ) Two NIC card connected with CrossOver cable directly

Experiment 2 always refers to:

Madrid:eth1( 192.168.1.2 ) ====>Dublin:eth2 ====>Dublin:eth1 ====>Prague:eth1( 192.168.2.2)

Dublin is working as router between network 192.168.1.0 and 192.168.2.0

Experiment 3 always refers to:

Madrid:eth0( 132.177.8.27 ) ===>Switch .... ===>Prague:eth0( 132.177.8.29 )

Setting up the Environment

Performance Test

Test on Kernel 2.2.16-22
3Com nic card is faster than tulip cards, See experiments HERE
1. Kernel 2.4.10 is much faster than Kernel 2.2.16-22 on IP Routing, See experiments HERE
2. Kernel 2.4.10 with Tulip NIC driver compiled in directly is much faster on ip routing than Kernel with NIC driver as modules. See Experiments here
Very weird!!! I did not change any setting on the 3 computers. But this afternoon's experiments dispute the above conclusion 2, See experiment here
Very weird!!! The experiments today also show that Kernel 2.2.16 is almost the same as Kernel 2.4.10 on IP routing, See Experiment here
NOW I CAN NOT RETURN TO those slow routing environment:(((
The relation between throughput and socket buffer size chart. TCP Experiment here
The relation between throughput and socket buffer size chart. UDP Experiment here
Round Trip Time and Throughput test by nettest. Experiment here
The max throughput when request size is optimal is about 1% less than ideal theoretical throughput 100Mb/sec. See Experiment here
The fluctuation of round trip time when requset size is 65536 Byte is NOT because of "TCP packet retransmitting" if the assumption that If some IP fragment in a given packet is loss or corrupted, the retransmission mechanism will transmit the whole packet again. And TCP retransmission can be captured by "tcpdump" is right. See experiments here