In TCP, Packet loss = congestion, so lossy links like wireless make TCP perform badly.
The paper does not introduce new ideas how to improve this, but does a great job in reviewing, systematizing and comparing the approaches which are listed in Table 1.
Two approaches:
- Hide losses from TCP sender, local problem solved locally, higher quality link with reduced effective bandwidth
- make sender aware of lossy links
Also that there might be timers in the TCP layer and the link layer which might be redundant and issue unnecessart retransmissions.
The conclusion
- reliable link-layer protocol with TCP awareness (e.g. in order packet delivery) improve performance greatly
- Can avoid split-connection without performance loss
- TCP mechanisms like SACK help to recover quicker from multiple packet losses
- explicit loss notification improves performance
1 comment:
Doesn't the link-layer retransmission mentioned in this paper operate at a far smaller timescale and thus not produce the overhead of a complete end-to-end transport-layer retransmission? Otherwise, there may definitely exist issues of redundancy - layering violations should be handled with care in this regard.
Post a Comment