Discovering Network and Grid topology

1
Discovering Network and Grid topology

• - By Tarun Saxena

2
Design
AppLeS etc.
Estimator
Communicator
Statistics
Prober
Network
3
Implementation - Prober

• Parameter Detected
• Bandwith
• Probing frequency
• 5 mins.
• Machines used
• Intel machines in GARL

4
Algorithm used

• Flooding.
• Target machine is flooded with UDP datagrams at
  the highest possible rate by the sender.
  Datagrams carry timestamp of their generation.
  Those that make through are summed and divided by
  time taken to estimate bandwidth.
• Done independently and simultaneously to identify
  colliding routes.

5
Algorithm.

• Hosts come up in sequential fashion. Every host
  should connect to ones executing before it.
• The host that comes up first generates a token.
• The host that holds the token conducts all its
  experiments and then passes it to the next host.
  This is to avoid collisions.
• Colliding routes are discovered by explicitly
  commanding other systems to probe. These results
  are compared with the those without any
  interference and if the interference exceeds a
  threshold, routes are considered as colliding.

6
Avoiding collision

• By token passing. This approach of total mutual
  exclusion is used in NWS.
• Same is used here.
• In a better version. Sub tokens are implemented
  that signal the independent nodes to start
  probing. These sub tokens are time lived and
  unpassable.

7
Execution Results

• Packets sized 1K were used.
• Machines inside the GARL domain all seem to be on
  independent paths as insignificant variation was
  detected on simultaneous probing. This is
  probably because the underlying network works on
  switches.

8
(No Transcript)
9
Comparison with NWS
10
Comparison..

• Results are parallel.
• Uses UDP instead of TCP.

11
Simultaneous probing on colliding routes

• On one side of both routes garl-intel1 is used
  and other ends are garl-intel2 and garl-intel5.
• Routes collide and the observed transfer rate is
  nearly half.
• This can be judged from the fact that probing on
  local host gives gt900mbps range gt that packet
  processing on local machines is fast and any
  bottleneck is due to network only. Also only
  about 10 of packets are able to make through in
  any experiment.

12
Collision graph
13
Implementation - Delay

• Ping and record the RTT
• ½ RTT is the estimated delay.

14
(No Transcript)
15
References

• www.cc.gatech.edu/people/home/gantsich/Projects/CS
  7001_Project1/Report.html
• www.cs.cornell.edu/boom/1999sp/projects/Network20
  Topology/topology.html
• SNMP RFC 1157
• Using Effective Network Views to Promote
  Distributed Application - Gary Shao, Fran Berman
  Rich Wolski
• A Framework Based to Approach to the Development
  of Network-Aware Applications Bolliger J. and
  Gross T.