One of the main reasons of throughput degradation over wireless links is transmission errors. Only few packets are corrected by frame error correction codes but more packet errors are assumed as a corrupted packet and discarded. These packets are considered as a lost packet by TCP and congestion window size is reduced as TCP takes packet loss as a sign of network congestion, in reality it may not be due to the congestion. This unnecessary reduction of congestion window may lead to the throughput degradation.
Another reason of throughput degradation over wireless links is the Frame Error Rate (FER). It suffers from FER of 1.55% when transmitting 1400 byte frames over an 85 foot distance, with clustered losses. Reducing the frame size by 300 bytes halves the measured FER, but causes framing overhead to consume a larger fraction of the bandwidth. In shared medium WLANs, forward TCP traffic (data) contends with reverse traffic (acknowledgments). In the WaveLAN this can lead to undetected collisions that significantly increase the FER visible to higher layers. File transfer tests over a WaveLAN with a nominal bandwidth of 1.6 Mbps achieved a throughput of only 1.25 Mbps. This 22% throughput reduction due to a FER of only 1.55% is caused by the frequent invocations of congestion control mechanisms which repeatedly reduce TCP’s transmission rate. If errors were uniformly distributed rather than clustered, throughput would increase to 1.51 Mbps. This is consistent with other experiments showing that TCP performs worse with clustered losses.
Another reason of throughput degradation over wireless links is the Frame Error Rate (FER). It suffers from FER of 1.55% when transmitting 1400 byte frames over an 85 foot distance, with clustered losses. Reducing the frame size by 300 bytes halves the measured FER, but causes framing overhead to consume a larger fraction of the bandwidth. In shared medium WLANs, forward TCP traffic (data) contends with reverse traffic (acknowledgments). In the WaveLAN this can lead to undetected collisions that significantly increase the FER visible to higher layers. File transfer tests over a WaveLAN with a nominal bandwidth of 1.6 Mbps achieved a throughput of only 1.25 Mbps. This 22% throughput reduction due to a FER of only 1.55% is caused by the frequent invocations of congestion control mechanisms which repeatedly reduce TCP’s transmission rate. If errors were uniformly distributed rather than clustered, throughput would increase to 1.51 Mbps. This is consistent with other experiments showing that TCP performs worse with clustered losses.
No comments:
Post a Comment