Optical Networking for Distributed Computing Environment

1
Optical Networking for Distributed Computing
Environment

• Hiroaki Harai (harai_at_nict.go.jp)
• National Institute of Information and
  Communications Technology (NICT), Japan
• May 17, 2005
• CEF Workshop 2005

2
Research Background

• Distributed computing environment via Internet
  (GRID)
• Difficult to assure delay and bandwidth
• Best effort, TCP
• Difficulty in tractability for application
  engineer
• Distributed computing environment on
  wavelength-routed network (Optical GRID)
• On request from an end host, a lightpath is
  established between two end hosts
• Multi Gbit/s bandwidth is assured by directly
  connecting end host resources
• Request may be blocked
• Focus on Optical GRID

3
Research Goal and Contents

• Contribute to advance in distributed computing
  technology and distributed computing application
  development by
• Using optical networking technology (optical
  component and control system)
• Developing distributed computing environment that
  assures data transmission bandwidth between hosts
• Develop distributed computing environment over
  l-path networks
• 1st step Dynamic establishment of optical ring
  (multiple paths connecting multiple computing
  sites)
• Optical networking for distributed computing
  environment todays title

4
Remaining Contents of This Talk

• Make dynamic optical ring for distributed
  computing
• Cooperating routing for point-to-point
  communications
• Establishing an optical ring with minimal-cost by
  using a heuristic for traveling salesman problem
• Mapping a ring composition to signaling
• Evaluate performance
• Apply ring composition concept to CEF network
• Conclusion and future direction

5
Making Dynamic Optical Ring for Distributed
Computing Environment
6
Optical GRID ltImagegt

• Develop distributed computing environment over
  WDM networks
• Multi-site communications
• High-speed channel (path)
• guaranteed bandwidth

7
Optical Ring for Multi-Site Communications

• Establishing a set of fully-meshed lightpaths is
  difficult
• Tree or ring?
• Light tree (bi-directional multicast tree)
• Optimization Find the least-cost tree by solving
  minimum Steiner tree problem
• Required Multicast capability at internal node
• Optical ring (set of unidirectional lightpaths)
• Optimization Find the least-cost set of
  lightpaths by solving traveling salesman problem
• Required Data duplication at hosts
• Making optical rings
• Can reduce number of required wavelengths
  (channels)
• Can use wavelength resources effectively

8
Making Optical Ring Dynamically

• Make multiple lightpaths dynamically
• Hosts establish an optical ring with
• Hosts receive routing information from network
• Considering routing constraint (routing for
  point-to-point communication)
• Each host decides a host closest to the host as a
  destination of a lightpath
• Solving traveling salesman problem under limited
  route information of hosts
• Each host establish a lightpath
  to the destination host
• Map optical ring composition
  to wavelength reservation
  signaling

9
Considering Routing Constraint

• It is practically difficult for nodes (routers)
  to have all the optimal routes for the
  unidirectional rings
• For number of hosts N, the optimal number of
  routes for
• Optical lightpaths (1to1) is N (N -1)
• Optical rings (k-to-k) is
• Optimal route for optical ring is different
  depending on number of hosts and set of hosts
  
• Network does not exist only for multi-site
  communication
• Nodes have point-to-point routing information
  only
• Nodes advertise routing information to hosts
• Each host has routing information from itself and
  does not have routing information that is not
  related to itself

10
Optical Ring Composition (TSP Heuristic)

• Parent For set of hosts Shs, h1, h2, ,
  hG-1, Generate list of hosts assigned lightpath
  (L ), list of hosts not assigned lightpath (U
  hs,h1, h2, , hG-1)
• Parent Remove hs from U, add it to L (Lhs,
  U h1, h2, , hG-1). Establish a lightpath
  from hs to host h that is included in U and is
  the least-cost
• Child Remove destination host h from list U and
  add it to list L (e.g., Lhs, hG-1, U h1,
  h2, , hG-2 if hhG-1 ). If not U , establish
  a lightpath from h to host that is included in U
  and is the least cost. Repeat this until U .
• Child Establish a lightpath from host that is
  included in list L most recently to host hs

11
Ring Composition to Wavelength Reservation
Signaling (FORWARD direction)

• If a lightpath is reserved
• Destination host sends ACK to the source host
• Destination starts reservation to the next
  lightpath
• Source host sends P-ACK to the parent host
• Otherwise, destination sends NACK to the source
  host
• The parent host regards receiving all P-ACKs as
  success of ring establishment

12
Evaluating Performance
13
Performance Evaluation (Number of Used
Wavelengths)

• Evaluate number of links required for
  communication in a group
• 16-node (4x4 mesh) network
• Tree setting up bi-directional lightpaths from
  parent hosts to other hosts
• Ring Non-Shortest setting up unidirectional
  lightpaths in random order
• Ring Shortest setting up unidirectional
  lightpaths for optimal resource usage

14
Performance Evaluation (FORWARD)

• Ring request Poisson, holding time exponential
  (mean 1)
• G hosts in a group is selected randomly
• At each source host, a wavelength is randomly
  selected for reservation
• Larger G (G10) slightly increases performance
  difference
• Sustaining the number of links required gives
  good influences to the performance

15
Applying Ring Composition Conceptto
Customer-Empowered-Fiber Network
16
Peer Model, Overlay Model

• Peer model is assumed in previous discussion
• End host is a unit of a network
• Collect routing information from the network
• Request a wavelength path

• Overlay model is possible in previous discussion
• End host estimates network topology in logical
  level
• End host estimates available wavelengths

- Lightpath request - Estimate virtual topology
and available wavelengths
Peer model
17
Customer-Empowered-Fiber Network Model

• End host can
• Send control packets to the network (peer model)
• Obtain routing information (peer model)
• Set up internal connection of internal node (
  peer model)
• Communication between customer to every node
• May consume time

18
Conclusion and Future Direction
19
Conclusion

• Optical networking for distributed computing
  environment
• Dynamic ring composition method for effective
  wavelength utilization
• TSP heuristic
• Signaling (FORWARD reservation)
• Cooperation between nodes and hosts for lightpath
  establishment
• Nodes only do routing for point-to-point
  communications and advertise the routing
  information to neighbor hosts
• Hosts collect routing information and decide
  destination for lightpath establishment
• Future direction
• BACKWARD reservation (compatibility to GMPLS
  RSVP-TE)
• Implementation
• Deployment to Peer/CEF network

20
Future Direction (On Planning)

• JGN2 (RD testbed by NICT, http//www.jgn.nict.go.
  jp/e/) has 1/10Gbps L2/L3 lines, and dark fibers
  (G.655 NZDSF, G.652 SMF)

NICT
21
Acknowledgment

• Professor M. Murata (Osaka University) for
  valuable comments for optical networking
• Dr. F. Kubota, and Dr. T. Miyazaki (NICT) for
  experiment planning discussion

22
Thank You for Your Attention