Transcript A Simulation model for e-VLBI traffic on network links in the

A Simulation model for e-VLBI traffic
on network links in the Netherlands
Julianne Sansa*
* With Arpad Szomoru & Thijs van der Hulst
5th e-VLBI Workshop, 17-20
September 2006, Haystack
1
Outline
•
•
•
•
•
•
•
Background
Motivation
Related Work
Setup
Results
The model
Conclusion & future work
2
Background
• TCP Congestion Control algorithm (AIMD) on
LFN
Cwnd = max. # packets that TCP sender injects into
network before receiving ACK.
ACK:Cwnd  Cwnd + 1/Cwnd
DROP: Cwnd  Cwnd -1/2*Cwnd
• Cwndoptimal = Bandwidth *RTT
• CA
• Evaluation of proposed TCP algorithms that
address the challenge and specifically in e-VLBI
setting.
3
Motivation
• Need for a model that can be used to
test & relate suggested improvements
of the underlying transport protocols to
the e-VLBI data in the ns-2
environment.
• ns-2 is a publicly available network
simulator
Breslau et.al.(2000), Nicol D.M.(2003), www.isi.edu/nsnam/ns
4
Related Work
• General TCP/IP data generation models: Danzig
et.al.(1992) and Paxson & Floyd (1994)
• Application specific data generation models:
Crovella et.al.(1998) - web , Hernandez-Campos
F. et.al. (2001) - FTP & SMTP
Various methods used to trace the data:
– Embedding instrumentation software in the client
– Installing specialised software and hardware in the network
– Installing publicly available packet capture tools on off-theshelf hardware
5
Setup
• TCPdump used to gather network statistics.
• ns-2 simulator used to simulate various scenarios,
each simulation is run for a period of 80 s and
repeated five times.
• High performance options set and also simualated:
MTU-8192 Bytes, TCP Buffers-4 MB,
txqueuelen-20,000
6
CWND & RWND for real and
simulated flows
Real
Simulated
7
Throughput for real and
simulated flows
Real
Simulated
8
The e-VLBI data generation model
The three factors
Large idle times
More
background
traffic
Low throughput
Low throughput
maxCWND < 256 packets
Increasing maxCWND
High throughput
maxCWND > 256 packets
Increasing maxCWND
Constant throughput
9
The e-VLBI data generation model
The combined effect
• ”on/off” bursty data generation, initially with
data bursts of 500 ms and idle times of 500
ms.
• Receiver limitation simulated with the
maximum CWND to 64 packets (0.06 Mbytes)
and RWND to the 50 packets (0.05 Mbytes).
• background traffic composed of
– 10 normal sized TCP flows from the reverse
direction
– 25 small TCP flows in the same direction
– 5 small TCP flows flowing in the opposite
direction,
– 110 web sessions starting randomly during the
10
Conclusions
• By comparing results of a real flow against
those of a simulation, the best approximation
for the e-VLBI data generation follows a
bursty pattern i.e. large bursts separated by
idle periods.
• The 3 factors seen to affect the flow’s
throughput are idle periods (most significant),
receiver limitation & background traffic.
11
Future work
• Future work will include designing data
generation models for the other commonly used
Mark5 transfer modes such as In2Net-Net2Out,
In2Net-Net2Disk,etc.
• Validating of data generation model by
conducting experiments elsewhere to guard
against biases due to local network conditions
such as hardware and local usage patterns
• Explore models that eliminate or shorten the idle
time between data bursts by using these models
in evaluation of transport protocols through
simulation
12
Questions
13