Some Unsolved Mathematical Problems in Systems Area Networking Mark Stewart M elbourne O perations Re search.

Slides:



Advertisements
Similar presentations
Cognitive Radio Communications and Networks: Principles and Practice By A. M. Wyglinski, M. Nekovee, Y. T. Hou (Elsevier, December 2009) 1 Chapter 9 Fundamentals.
Advertisements

February 20, Spatio-Temporal Bandwidth Reuse: A Centralized Scheduling Mechanism for Wireless Mesh Networks Mahbub Alam Prof. Choong Seon Hong.
Data and Computer Communications
CCNA3: Switching Basics and Intermediate Routing v3.0 CISCO NETWORKING ACADEMY PROGRAM Switching Concepts Introduction to Ethernet/802.3 LANs Introduction.
Introduction to Computer Networks Spanning Tree 1.
COS 461 Fall 1997 Routing COS 461 Fall 1997 Typical Structure.
1 Traffic Engineering (TE). 2 Network Congestion Causes of congestion –Lack of network resources –Uneven distribution of traffic caused by current dynamic.
Advanced Networking Wickus Nienaber Daniel Beech.
The Structure of Networks with emphasis on information and social networks T-214-SINE Summer 2011 Chapter 8 Ýmir Vigfússon.
What's inside a router? We have yet to consider the switching function of a router - the actual transfer of datagrams from a router's incoming links to.
Multiple constraints QoS Routing Given: - a (real time) connection request with specified QoS requirements (e.g., Bdw, Delay, Jitter, packet loss, path.
December 20, 2004MPLS: TE and Restoration1 MPLS: Traffic Engineering and Restoration Routing Zartash Afzal Uzmi Computer Science and Engineering Lahore.
Improving TCP Performance over Ad-hoc Network 11/28/2000 Xuanming Dong, Duke Lee, and Jin Wang Course Project for EE228A --- Fall 2000 (Professor Jean.
A General approach to MPLS Path Protection using Segments Ashish Gupta Ashish Gupta.
Dynamic routing – QoS routing Load sensitive routing QoS routing.
CSE 461: Distance Vector Routing. Next Topic  Focus  How do we calculate routes for packets?  Routing is a network layer function  Routing Algorithms.
1 Lecture 25: Interconnection Networks Topics: communication latency, centralized and decentralized switches, routing, deadlocks (Appendix E) Review session,
A General approach to MPLS Path Protection using Segments Ashish Gupta Ashish Gupta.
1 25\10\2010 Unit-V Connecting LANs Unit – 5 Connecting DevicesConnecting Devices Backbone NetworksBackbone Networks Virtual LANsVirtual LANs.
Spanning Tree and Multicast. The Story So Far Switched ethernet is good – Besides switching needed to join even multiple classical ethernet networks Routing.
Layer-3 Routing Natawut Nupairoj, Ph.D. Department of Computer Engineering Chulalongkorn University.
1 Algorithms for Bandwidth Efficient Multicast Routing in Multi-channel Multi-radio Wireless Mesh Networks Hoang Lan Nguyen and Uyen Trang Nguyen Presenter:
Ch. 28 Q and A IS 333 Spring Q1 Q: What is network latency? 1.Changes in delay and duration of the changes 2.time required to transfer data across.
Dragonfly Topology and Routing
Switching, routing, and flow control in interconnection networks.
Operations Research Models
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS 2007 (TPDS 2007)
High-Performance Networks for Dataflow Architectures Pravin Bhat Andrew Putnam.
Introduction to Operations Research
1 Optical Burst Switching (OBS). 2 Optical Internet IP runs over an all-optical WDM layer –OXCs interconnected by fiber links –IP routers attached to.
Multiplexing FDM & TDM. Multiplexing When two communicating nodes are connected through a media, it generally happens that bandwidth of media is several.
15.1 Chapter 15 Connecting LANs, Backbone Networks, and Virtual LANs Copyright © The McGraw-Hill Companies, Inc. Permission required for reproduction or.
Networking Fundamentals. Basics Network – collection of nodes and links that cooperate for communication Nodes – computer systems –Internal (routers,
1 Data Link Layer Lecture 23 Imran Ahmed University of Management & Technology.
1 Presented by Sarbagya Buddhacharya. 2 Increasing bandwidth demand in telecommunication networks is satisfied by WDM networks. Dimensioning of WDM networks.
Supporting DiffServ with Per-Class Traffic Engineering in MPLS.
Interconnect Networks Basics. Generic parallel/distributed system architecture On-chip interconnects (manycore processor) Off-chip interconnects (clusters.
1 Buffering Strategies in ATM Switches Carey Williamson Department of Computer Science University of Calgary.
CCNA3 Module 4 Brierley Module 4. CCNA3 Module 4 Brierley Topics LAN congestion and its effect on network performance Advantages of LAN segmentation in.
© Janice Regan, CMPT 128, CMPT 371 Data Communications and Networking Congestion Control 0.
1 Low Latency Multimedia Broadcast in Multi-Rate Wireless Meshes Chun Tung Chou, Archan Misra Proc. 1st IEEE Workshop on Wireless Mesh Networks (WIMESH),
Computer Communication and Networking Lecture # 4 by Zainab Malik 1.
Routing in Delay Tolerant Network Qing Ye EDIFY Group of Lehigh University.
© Sudhakar Yalamanchili, Georgia Institute of Technology (except as indicated) Deadlock: Part II - Recovery.
4. Interconnecting Networks: Routers. © Tallal Elshabrawy 2 Bridges Vs Routers BRIDGES DO WELL IN SMALL (FEW HUNDRED HOSTS) WHILE ROUTERS USED IN LARGE.
VADD: Vehicle-Assisted Data Delivery in Vehicular Ad Hoc Networks Zhao, J.; Cao, G. IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, 鄭宇辰
Chapter 10 Congestion Control in Data Networks and Internets 1 Chapter 10 Congestion Control in Data Networks and Internets.
1 Lecture 22: Interconnection Networks Topics: Routing, deadlock, flow control, virtual channels.
William Stallings Data and Computer Communications
Network Layer COMPUTER NETWORKS Networking Standards (Network LAYER)
Delay-Tolerant Networks (DTNs)
Routing and Switching Fabrics
Author: Daniel Guija Alcaraz
ITEC 202 Operating Systems
Managing the performance of multiple radio Multihop ESS Mesh Networks.
ISP and Egress Path Selection for Multihomed Networks
COS 561: Advanced Computer Networks
Data and Computer Communications
Congestion Control (from Chapter 05)
Congestion Control (from Chapter 05)
CEG 4131 Computer Architecture III Miodrag Bolic
Lecture: Interconnection Networks
Congestion Control (from Chapter 05)
CS 6290 Many-core & Interconnect
Congestion Control (from Chapter 05)
Congestion Control (from Chapter 05)
Congestion Control (from Chapter 05)
Routing and Switching Fabrics
Congestion Control (from Chapter 05)
Congestion Control (from Chapter 05)
Presentation transcript:

Some Unsolved Mathematical Problems in Systems Area Networking Mark Stewart M elbourne O perations Re search

Abstract Cluster computing represents the only feasible way of addressing many significant and computationally challenging problems today. This in turn has created a demand for networking technologies with higher bandwidths and lower latencies than contemporary Local Area Networks. The Systems Area Network (SAN) is the answer. This talk presents a number of open problems which I encountered whilst working at a start-up company developing InfiniBand Switches for the SAN market. The problems are mathematical (and/or computer scientific) in nature and in my opinion have an enduring nature to them.

Routing Algorithms: Deadlock Freedom and Load Balancing Most SAN technologies use a class of flow control algorithms known as “credit based flow control,” in which packet loss is avoided by a downstream node holding onto a packet until it’s upstream neighbor has indicated that it has adequate resources to accept more packets. This can lead to a phenomena known as Deadlock in which a cycle of nodes are each waiting on the others to free up resources before they can make progress. Think of Grid Lock in a road network and you have the same concept

Routing Algorithms: Deadlock Freedom and Load Balancing Digression: the dynamics of how a network enters a Deadlocked state is interesting. Most (all?) of the literature focuses on how to detect the potential for deadlock, but does not deal with the expected time before a network enters such a state. The instances of deadlock I have witnessed give a expected time until deadlock well below human thresholds for perception. Far faster than any simplistic modeling would predict. A better understanding of why the time to deadlock can be so fast, might lead to a far better deadlock avoidance and recovery schemes than are currently used.

Routing Algorithms: Deadlock Freedom and Load Balancing Current Solutions: Timeouts – In the event that a switch has held a packet to long drop the packet to free up resources. Only really effective if the expected time to deadlock is large compared with timeout values. (see previous digression) Deadlock Free Routing – use a routing algorithm which guarantees that the associated buffer dependency graph is acyclic. Adaptive Routing – mitigates the problem but doesn’t solve it – also a series of talks in its own right.

Routing Algorithms: Deadlock Freedom and Load Balancing Deadlock Free Routing: Spanning Tree algorithm –Simplest of the known algorithms –Choose a spanning tree –Don’t use links that aren’t in the spanning tree –Makes poor use of network resources

Routing Algorithms: Deadlock Freedom and Load Balancing Deadlock Free Routing: Up*/Down* –Select a node to be the Hub –Order the nodes by “distance” from the Hub Any Tie Breaking rule is fine –Label the directed arcs as Up arcs if they go to a node “closer” to the hub. –Label the directed arcs as down arcs if they go to a node “further” from the hub. –Choose the shortest paths which do not route from a down arc to an up arc. –In principal better than the spanning tree algorithm but in practice …

Routing Algorithms: Deadlock Freedom and Load Balancing Deadlock Free Routing: A common network topology hub The orange switches carry no traffic. Degrades sustainable network throughput by an order of magnitude

Routing Algorithms: Deadlock Freedom and Load Balancing The Challenge Develop a Generic Deadlock Free Routing algorithm that makes better use of existing network resources Note: A related problem would be to design networks for which the existing algorithms are more appropriate. I’d lack the courage to try and sell that idea to a customer again.

Inverse Multiplexing The impact of Cut Through There are essentially two ways to build faster links –Use higher frequencies –Use more “wires” and inverse multiplexing Multiple frequencies is effectively more wires and very hard in this context –Typically a combination of both is used. Cut through is a mechanism by which switching latency is reduced through forwarding a packet before the switch has finished receiving a packet. –Any SAN switch must do this.

Inverse Multiplexing The impact of Cut Through If we have multiple “wires” between two switches should we: Use inverse multiplexing to form a faster link Use some form of load balancing across the slower wires? Conventional queuing analysis says we should inverse multiplex. –And that this is always better –A server of rate M is better than M servers of rate 1. Not so (at least not always)

Inverse Multiplexing The impact of Cut Through Cut Through has a problem when the outgoing link of a switch is faster than the incoming link. –If the packet is forwarded too soon the switch will run out of data and the packet will corrupted. A switch must delay packets when there is a rate mismatch. In lightly loaded networks this is the dominant source of network latency! For some special applications even under heavy load this is the dominant source of network latency.

Inverse Multiplexing The impact of Cut Through Question: –Does there exist a traffic pattern for which adaptive routing would not outperform the faster link? –What can we say about average performance etc.

Network Design and Network Load Many SANs are designed around Non-Blocking Topologies. –But lack the signaling infrastructure to exploit the to make use of the networks non-blocking potential. The principal advantages derived from such topologies is a modest maximum hop count and that the expected offered load to any link is less than one. –Assuming good load balancing, and a few other things

Network Design and Network Load Given that it is rare to exploit the non-blocking potential of the networks, is there a better choice of network topology? Absolutely! –As an existence proof there are other Non-Blocking Topologies of identical cost and lower average hop count. –They are not suitable for a self routing network, but that has not been a design consideration for years (decades?)

Network Design and Network Load What if we drop the Non-Blocking bit altogether? Question: What is the Topology with fewest switches? That has –N single ported end stations –Switches of valency V –A maximum offered load L on any link Under the assumption that each end station talks to every other end station at rate 1/(N-1). Shortest Path Routing is used, Ties are resolved by some load balancing heuristic. –A maximum hop count H between any two end stations.

Network Design and Network Load There is a lot of work in the literature on finding networks of fixed maximum valency, and of proscribed maximum diameter. These networks may be used to construct good but sub-optimal solutions to the above question. Note: this sub-optimality is evidenced as pockets of sub-optimality.

Network Design and Network Load For N=30, V=6, L=1, H=2 K3*C5 yields the following graph But doesn’t extend to larger N (unless we increase V,H or L)

Network Design and Network Load For N=42, V=6, L=1, H=2 The following graph is viable. Family works for any even V. Is it optimal?

Network Design and Network Load For N=84, V=6, L=1, H=3 The following graph is viable. Is it optimal?