View
215
Download
0
Category
Tags:
Preview:
Citation preview
588 Section 4
Neil Spring
April 27, 1999
Schedule
• Notes
• Project 2 description
• Fair Queueing (Demers et.al.)
Notes
• Graded Homework Assignment 1 soon– first seven problems have been graded.
• Homework 2 due date– was May 3, now…
• Programming Assignment 1 “solution” available as part of PA2
Project 2, Routing & Congestion
• Routing: – topology discovery, (make table)– routing packets (lookup in table)– failure management (change table)
• no partitioning
• Congestion control– adaptive window sizing – Drop packets or ?– Loss rate won’t be as bad this time.
Fair Queueing
• What’s the problem?
Fair Queueing
• What’s the problem?– Fair bandwidth allocation?
• What’s wrong with FCFS?
Fair Queueing
• What’s the problem?– Fair bandwidth allocation?
• What’s wrong with FCFS?– FTP vs. Telnet– Mean users break end-to-end congestion
control
Nagel’s algorithms
• Silly Window Syndrome– Allow only one unacknowledged small packet
into the network.
• Packetwise fair queueing– Compared to Bitwise Round-robin (BR)
described in this paper.
FTP vs. Telnet
• FTP shoves a lot of packets into the network.
• Why?
FTP vs. Telnet
• FTP shoves a lot of packets into the network.
• Why?– It gets a bigger share of the bandwidth– It makes sure it gets what bandwidth is
available
• What are the consequences of full queues?
FTP vs. Telnet
• FTP shoves a lot of packets into the network.
• Why?– It gets a bigger share of the bandwidth– It makes sure it gets what bandwidth is available
• What are the consequences of full queues?– Packets get dropped– Packets get delayed
What is fairness
• Plenty of rhetorical questions:
• Equal allocation?– By source? Destination? Connection?
• Some sources really need bandwidth.– NFS service
• What about users with many processes?
• Sources with many outgoing connections?
Experimental Setup
• Simulation of FCFS & FQ using three flow control algorithms:
• Generic flow control– 2rtt, fixed window size
• Jacobson & Karels’ (JK)– timeouts signals congestion: modify cwnd
• DECbit– header bit when passing congested gateways
Measurements
• Throughput– fairness evident– more telnet packets are good
• Average Roundtrip– delay
• Retransmissions– suggest variability in delay, since timeouts fire
• Dropped Packets– imply congestion was not resolved.
Results: Underloaded Gateway
• Experimental setup looks like a modem link (56Kbit)
• fairness
• low delay
• DECbit already had decent delay properties– has strange Roundtrip times for FTP in FQbit
Results: Overloaded Gateway
• Small buffer size
• Notice: several ways to achieve fairness
• FQ affects telnet delay
• FQ doesn’t reduce the number of retransmits/drops
Results: Ill-behaved source
• Mean source fills the queue
• FQ charges for dropped packets
• Effective at shutting it down.
• Roundtrip for good apps preserved
Results: Mixed Protocols
• Explain the 12 for one Generic source?
• Motivation for sources to implement JK.
Results: Multi-hop networks
• Key is the fourth column
• DECbit doesn’t work as well
• Why roundtrip time is the same across all routes is mysterious…
Results: Complex networks
• Column 4 & 8
• Senders aided by timely acks.
Summary
• what’s wrong with fair queueing?
Recommended