Upload
louise-briggs
View
214
Download
0
Embed Size (px)
Citation preview
Shivkumar KalyanaramanRensselaer Polytechnic Institute
1
A TCP Friendly Traffic Marker for IP Differentiated Services
Feroz Azeem, Shiv Kalyanaraman, Amit Rao
Rensselaer Polytechnic Institute
http://www.ecse.rpi.edu/Homepages/shivkuma
Shivkumar KalyanaramanRensselaer Polytechnic Institute
2
TCP performance over Differentiated Services What are “TCP-friendly” building blocks ? TCP-friendly traffic conditioners Sample performance results
Overview
Shivkumar KalyanaramanRensselaer Polytechnic Institute
3
Diff-serv Model
Congestion = traffic jams of the Internet Congestion control (CC) requires end-system support
Traffic management = providing bandwidth services TM requires some support from all network components
Problem: providing better-than-best-effort services for TCP Lies at the intersection of the TM and CC problems
Ingress Edge
Router
Egress Edge Router
Interior Router
End system
End system
Shivkumar KalyanaramanRensselaer Polytechnic Institute
4
TCP service performance
User metrics: Average per-flow goodput Coefficient of variation (spread) of per-flow goodput
Timeouts Provider metrics:
Bottleneck Utilization Queue length (avg/max) Packet loss rate
Parameters: a) Workloadb) Configuration components and protocols
System
Metrics:a) User metricsb) Provider Metrics
Shivkumar KalyanaramanRensselaer Polytechnic Institute
5
TCP performance issues User-perceived performance can be bad even if provider-perceived
performance is good. Key problem: second-order effects of packet-loss
Per-connection burst loss of packets is bad for TCP Reno Even isolated loss of packets is bad for a TCP connection with a
small window. Happens with: high degrees of muxing or, slower bottleneck rates or buffer sizes
Probability of burst loss high when a number of TCP connections in slow start phase (eg: WWW)
Hypothesis: problems can be alleviated by using “TCP-friendly” building blocks (possible violation of layering)
Shivkumar KalyanaramanRensselaer Polytechnic Institute
6
TCP-friendly building blocks Goals: Reduce probability of burst loss and TCP synchronization Convert aggregate burst loss into per-connection isolated packet
loss. Also minimize per-connection loss instances. Protect small window connections from packet loss
Sample Methods: Random early dropping: packet drop scheme like RED Control TCP rate or dampen TCP rate increases:
Control left, right edges of TCP window (rcv_wnd and ack #) and/or rate of release of acks
Reduce TCP burstiness Interleave packets of multiple flows Protect small window flows from loss
Shivkumar KalyanaramanRensselaer Polytechnic Institute
7
TCP-friendly Marker Problem: Allocate a limited aggregate pool of tokens among the
active TCP flows to maximize performance as measured by user and provider best-effort metrics.
Method: Estimate window sizes of flows W(i).
Assume: maximum token requirement for flow i = W(i) First allocate and satisfy IN token requirements for all small-
window flows, I.e., W(i) < 4. Divide the remaining tokens in a max-min fair way among the
rest of the flows. Provide maximum equal spacing between packets marked OUT. Carry over of tokens across intervals limited by policy
Shivkumar KalyanaramanRensselaer Polytechnic Institute
8
O U T P acket
IN P acket
W ith Token B ucket M arker
W ith TC P -friend ly M arker
Illustration of the TC P-friendly M arking Schem e
Shivkumar KalyanaramanRensselaer Polytechnic Institute
9
Configuration1
5
2
.
.
.
Marker0
6
10
7
.
.
.
Marker1
R1 R2
1
5
2
.
.
.
6
10
7
.
.
.
S ourcesD estina tions
A ll L inks 1 .5 M bps, 1000 K m un less o therw isem entioned
10 M bps
10 M bps
Shivkumar KalyanaramanRensselaer Polytechnic Institute
10
Sample Results In all cases, the performance improvement is consistent, as
measured by provider and user metrics. Performance improvement is marked with higher speed links
and larger number of flows.
Eg:With 100 flows, 150 Mbps bottleneck: TCPFM: 1261 timeouts, 240 losses/s, 194kbps TBM: 4254 timeouts, 311 losses/s, 134kbps No marker:2954 timeouts, 322 losses/s, 151kbps
Validates Ibanez et al’s observations in the IETF (March 1999) about the unpredictability of TBM
Shivkumar KalyanaramanRensselaer Polytechnic Institute
11
Sample Results (contd)
Results also extend to TCP SACK. But the total number of timeouts in all cases is smaller.
Better-than-best-effort service: The token and capacity over-provisioning required to
provide assured service also reduces with the TCP-friendly marker
But the second-order effects of loss on TCP are not totally compensated for.
Later experimental work (submitted to ICNP’2000) also protects retransmissions which leads to order-of-magnitude performance gains, validating the approach.
Shivkumar KalyanaramanRensselaer Polytechnic Institute
12
Summary
TCP-friendly refers to components which promote better TCP performance, esp timeout and service-variance reduction.
A simple TCP-friendly marker which leverages the dimension of network-based packet marking in conjunction with the assured service. Effects: Reduce user-perceived service variance (user benefit) Reduce token provisioning requirements (provider benefit)