Transcript Efficient Policies for Carrying Traffic Over Flow

Efficient Policies for Carrying
Traffic Over Flow-Switched Networks
Anja Feldmann, Jenifer Rexford, and Ramon Caceres
Presenters: Tauhid, Reji and Murshed
Advanced Computer Networks
1
Objective
• Increase the efficiency of the switching technique
• Reduce the overhead for establishing and
maintaining the dedicated route or shortcut for
flow
Advanced Computer Networks
2
Introduction
A possible solution could be grouping series of related
data packets into flows and sending the flows through a
shortcut path
Benefits: improve performance and capitalizing on
QoS of the switches
Drawbacks: consume network resources to create
and maintain dedicated path for the flow
Different definition and criteria are necessary for flow
construction and shortcut creation and maintenance
Advanced Computer Networks
3
Introduction (Contd.)
Three parameters determine the flow and
shortcut decisions:
• aggregation: level for traffic combination
• timeout: time interval of any flow
• trigger: traffic quantity or number of packets for
constituting a flow
Advanced Computer Networks
4
Introduction (Contd.)
The effects of varying these parameters were
explored on three metrics of interest:
• percentage of traffic following the shortcut
• shortcut setup rate
• number of simultaneous shortcut
Advanced Computer Networks
5
Trace Collection of Packets
The Extensive packet-level trace collection was
carried out on T1 line of AT&T Research Lab.
Several aspects were considered as:
-the traffic trace reflects the dominance of WWW
-long continuous traffic traces were observed
-the endpoint machines were taken for consideration
-all the flow and shortcut parameters were evaluated
on various parameters
Advanced Computer Networks
6
Trace Collection of Packets (Contd.)
•The T1 line connects the AT&T Lab to the external
internet. The Ethernet segment carries all traffic to and
from the T1 line at a speed of 10-Mbits/s
The Sniffer was equipped with tcdump to collect
information of 100 00 packets in raw binary format.
Advanced Computer Networks
7
Trace Collection of Packets (Contd.)
• Tcdump read each packet and generated an ASCII log
• Perl script classified each packet into flows
• Splus function processed all the log files on different
performance metrics
Advanced Computer Networks
8
Trace Collection of Packets (Contd.)
From the optional field of the web request messages, the
operating system of the client machine was identified.
Linux and/or Windows systems were classified as
personal computers
IP address associated with multiple types of machines
such as Windows and SunOS are classifies as Proxy
Servers
Advanced Computer Networks
9
Hourly average throughput of the
Ethernet segment
Advanced Computer Networks
10
The packet size distribution of the network traffic
Advanced Computer Networks
11
Trace Collection of Packets (Contd.)
• Half of the outgoing traffic consists of 40
bytes TCP acknowledgement packets, so the
average packet size became 123 bytes
• More than 60% of the incoming packets have
552, 567, or 1500 bytes, that correspond to
common maximum transfer unit size on the
Internet
• Using tcpreduce tool, the FTP, HTTP,
SMTP, DNS, etc. data transfer were measured
Advanced Computer Networks
12
Flow Size Distributions
• Mixture of flows between 100 and 100,000 bytes
• Majority of flows within range of 50 to 1000 bytes
• 1/3 of flows between 1000 – 10,000 bytes. 1/3
between 100 – 1000 bytes
Advanced Computer Networks
13
Cumulative Distribution
Advanced Computer Networks
14
Derivative of Cumulative Distribution
Advanced Computer Networks
15
Density of the logarithm of the flow sizes
Advanced Computer Networks
16
Flow sizes of Web Traffic
• SMTP (email) – large concentration of 1000 byte
flows, remaining flows around 3000 bytes
• Telnet – Bytes per flow is typically less than 100
bytes
• HTTP has the largest bytes per flow. Flows fall
into 3 main categories
– 1. Less than 150 bytes, from failed TCP sessions and
error messages.
– 2. 150 – 300 bytes, from cache hit messages.
– 3. Greater than 300 bytes, actual web page transfers.
Advanced Computer Networks
17
Flow sizes of Traffic
Advanced Computer Networks
18
Flow duration of Web Traffic
• Duration of SMTP traffic falls between 0.01
seconds and 100 seconds, with high concentration
between .01 and 1 second
• Telnet flows typically have the longest duration.
Data shows duration up to 17 minutes. Usually
around 100 seconds
• HTTP Flow duration typically around 1 second,
but can last up to 100 seconds
Advanced Computer Networks
19
Flow duration of Traffic
Advanced Computer Networks
20
HTTP Flows by Machine Type
• End point machines have unique flow
characteristics
• Flow duration for modem connection is longer
• Traffic to proxy servers tend to be a mixture of
shorter and longer flows
• Larger transfers have higher throughputs since
TCP window grows
Advanced Computer Networks
21
Flow duration for different machines
Advanced Computer Networks
22
Combining Multiple Web Responses
• 3 different levels of end-point aggregation
– Port-to-port: Combines packets with the same
IP address and port numbers at both end-points.
– Host-level: Combines different TCP session
from the same Web server to the same Web
client.
– Net-to-Net: Aggregating hosts that share the
first 16 bits of the IP address.
Advanced Computer Networks
23
Flow sizes for web responses
Advanced Computer Networks
24
Flow sizes for web responses
Advanced Computer Networks
25
Proportion of Shortcut Traffic
• With a trigger of 10 – 20 packets, network can
avoid establishing shortcuts for short-lived flows.
• X-packet trigger less effective for host-to-host
using packet trigger since aggregation increases
number of packets thus triggering too early
Advanced Computer Networks
26
Percent of shortcut traffic (port-to-port)
Advanced Computer Networks
27
Percent of shortcut traffic (host-to-host)
Advanced Computer Networks
28
Shortcut Setup Rate
Advanced Computer Networks
29
Shortcut Setup Rate (Contd.)
Advanced Computer Networks
30
Shortcut Setup Rate (Contd.)
• Setup rate is lower for larger packet size
• Setup rate is lower for higher timeout
Advanced Computer Networks
31
Simultaneous Shortcut Connections
Advanced Computer Networks
32
Simultaneous Shortcut Connections (Contd.)
• 0-packet trigger 75.8 shortcuts in average
• 10-packet trigger reduces the average no of
simultaneous shortcuts by a factor of 3.3 (75.8 to
23.6)
– Reason: higher packet trigger causes less
shortcut connections
• For stable operation during heavy load, coarser
aggregation and larger triggers may be necessary
Advanced Computer Networks
33
Simultaneous Shortcut Connections (Contd.)
Advanced Computer Networks
34
Simultaneous Shortcut Connections (Contd.)
• 12 sec timeout causes 4.3 shortcuts
• 300 sec timeout causes 44.5 shortcuts
• Larger timeout increases no of shortcuts, as a
result it reduces signaling load
• Establishment of shortcuts is limited by no of
connections, or the signaling capacity of the
switch, or both
• Substantial network overhead reduction is possible
by timeouts, triggers, end-point aggregation
Advanced Computer Networks
35
Traffic Flows Along Partial Routes
• As the hop count increases, the numbers
increases exponentially. ( EX: first seven
hops have 26, 71, 137, 267, 409, 916, 1508
different outcomes, respectively.)
Advanced Computer Networks
36
Partial Route Aggregation
Advanced Computer Networks
37
Partial Route Aggregation (Contd.)
• Aggregating traffic along a portion of the routes
decreases both the setup rate and the number of
simultaneous connection
– Seven hop: Setup rate = 20%, No. of Shortcuts = 11%
– Three hop: Setup rate = 37%, No of Shortcuts = 16%
• Percent of Byte in Shortcuts
– End-to-End:
– Seven hop:
– Three hop:
92.2%
94.0%
96.0%
Advanced Computer Networks
38
Partial Route Aggregation (Contd.)
• Partial-route aggregation reduces network
overheads while increasing the proportion of
shortcut traffic
• Partial-route aggregation combines transforms
from replicas of the same web site, as well as
related servers at the same institution
• Partial route flows benefit more from larger
timeout values
– As seen in Table III for 60 sec and 300 sec.
Advanced Computer Networks
39
Summary
• Flow characteristics vary with endpoint type
(modem, personal computer, compute
servers, proxy servers)
• Aggregating consecutive and concurrent
transfers from the same web server yields
substantial benefits
Advanced Computer Networks
40
Summary (Contd.)
• Aggregation and triggers both reduce overhead,
but the reductions are not multiplicative
• Variability in traffic load changes with time scale
• Aggregating traffic along portions of the route
between the source and destination yields
additional reduction in network overhead
Advanced Computer Networks
41
Future Work
• To evaluate specific new policies that balance the
short-term tradeoffs between processor and
network load
• To investigate the policy and performance
implications of combining traffic along a portion
of the route
• To study more detailed breakdown of web traffic
by content type, as well as the implications of
push technology and the new features in the
emerging HTTP standards
Advanced Computer Networks
42