Transcript CSE331-13
CSE331: Introduction to Networks and Security Lecture 13 Fall 2002 Announcements • Reminder: – Project 1 due on Monday, Oct. 7th – In-class midterm Wednesday, Oct. 9th • Monday’s Class – Further Topics in Networking – Review / Question & Answer CSE331 Fall 2002 2 Recap • Application Level Protocols – SMTP – HTTP – SNMP Today • Congestion control • Resource Management • Quality of Service CSE331 Fall 2002 3 Sharing Resources • How do we effectively & fairly share resources on the net? – Bandwidth of the links – Buffers space in routers in switches • Many competing users – What does “fairly” mean? CSE331 Fall 2002 4 Contention and Congestion • Packets contend at a router for use of a link – Multiple packets are enqueued at the router • Congestion is when packets are dropped because the queue is full – Wasted resources – Can lead to timeouts/retransmission • Problem is resource allocation CSE331 Fall 2002 5 Congestion In Packet-switched Networks Sources Router CSE331 Fall 2002 Destination 6 Network Resource Allocation • Challenges – – – – – Distributed resources are hard to coordinate Only way to coordinate is through the network itself! Not isolated to single level of the protocol hierarchy Not always possible to “route around” congestion Bottleneck not always visible from the source • Resource allocation – Attempt to meet competing demands of applications – Not always possible! CSE331 Fall 2002 7 Flows • A flow is a sequence of packets sent along the same route between a source and dest. • Connectionless Flows – No per-flow state at the routers – Example: Pure datagram model • Connection Oriented Flows – – – – Necessary per-flow state at the routers Explicitly created/removed by signalling Example: Virtual Circuit Switching Potentially does not scale • Soft-state Flows – Some (not strictly necessary) per-flow state – Example: Routing information in Learning Bridges CSE331 Fall 2002 8 Multiple flows Source 1 Dest. 1 Router Router Source 2 Router Dest. 2 Source 3 CSE331 Fall 2002 9 Router- vs. Host-Centric • Router-centric – Each router selects packets to forward & packets to drop – Routers inform hosts about network conditions • Host-centric – Hosts observe network behavior by watching ACKs, Timeouts, ICMP messages, etc. – Adjust behavior accordingly • Not mutually exclusive approaches CSE331 Fall 2002 10 Reservation vs. Feedback • Reservation – – – – End hosts ask network for certain amount of capacity If request can’t be satisfied, router rejects the flow Examples: measure MTU or link capacities Router-centric approach • Feedback – – – – End hosts send data without reserving capacity Adjust behavior based on feedback Explicit feedback: TCP flow control Implicit feedback: Packet losses CSE331 Fall 2002 11 Throughput, Delay and Load • Network load is a measure of total link utilization • Ideally we would – Maximize throughput – Minimize delay • Increasing #packets in network lengthens queues, which increases delay. • Power = Throughput/Delay CSE331 Fall 2002 12 Throughput/Delay Power vs. Load Optimal Load Load • Ratio of Throughput/Delay as a function of network load • Difficult to control load in fine-grained ways • Need stable mechanism: avoid thrashing CSE331 Fall 2002 13 Fair Resource Allocation • What does “Fair” mean? – Equal share of resources for all flows? – Proportional to how much you pay for service? – Should we take route length into account? Router Router Router CSE331 Fall 2002 Router 14 FIFO Queuing • First-in First-out – Scheduling discipline: determines order • Tail Drop – If queue is full, most recent packet to arrive is dropped – Drop policy: which packets are dropped • Most widely used in Internet routers – Pushes congestion control & resource allocation to end hosts (TCP) – Does not discriminate between flows – Trusts end hosts to “share” – but no one is forced to use TCP, for example. CSE331 Fall 2002 15 Priority Queuing • Simple variant on FIFO – Use the IP Type of Service header field as a priority – Send all higher priority packets in the queue before sending lower priority packets • Problems – Starvation of low-priority flows – Who sets priorities? (Not end user!) CSE331 Fall 2002 16 Fair Queing Flow 1 Round Robin Flow 2 Flow 3 • Strategy – Maintain a separate queue for each flow being handled by the router – Individual queues are treated FIFO with tail-drop – Queues are handled round-robin CSE331 Fall 2002 17 Fair Queuing Continued • Designed to be used with end-to-end congestion control – Doesn’t restrict transmission rates of end hosts – Badly-behaved end hosts only hurt themselves • Details – Different packet sizes complicates “fairness” – Link is never idle (as long as there is data to send) – If N flows are transmitting, each gets maximum of 1/N bandwidth CSE331 Fall 2002 18 Congestion Avoidance Mechanisms • Try to prevent congestion before it occurs – Unlike TCP, which reacts to existing congestion • Strategy 1: Routers watch their queues – Routers set a bit in outgoing packets if avg. queue length > 1 – Receiver copies bit into its ACK – Sender increases/decreases send window based on # of packets that report congestion – Called the DECbit algorithm CSE331 Fall 2002 19 Congestion Avoidance Continued • Strategy 2: Random Early Detection – Router monitors queue length – If length > dropLevel then drop packet with certain probability – Source times out on dropped packets – TCP causes send window to decrease – Much tuning of parameters to optimize performance CSE331 Fall 2002 20 Quality of Service Issues • Sometimes best effort is not enough • Application requirements – Real time: data must arrive within certain time constraints to be useful • Telephony, video conferencing – Jitter (variation in arrival times of packets) is bad • Audio/visual data need low jitter – Packet loss:can it be tolerated or not? • Mpeg can interpolate missing frames • Remote robot surgeon cannot tolerate packet loss CSE331 Fall 2002 21 Playback Buffer Example Packet Arrival Sequence # Packet Generation buffer delay Playback Time CSE331 Fall 2002 22 Integrated Services (RSVP) • Proposed in 1995-1997 • Service Classes – Guaranteed arrival service • For delay intolerant applications • Guarantee a maximum delay – Controlled Load • For loss tolerant, adaptive applications • Emulate lightly loaded network CSE331 Fall 2002 23 Implementation Mechanisms • Flowspecs – Describe the kind of service needed • “I need maximum delay of 100ms” • “I need to use controlled load service” • Admission Control – Network decides whether it can provide the desired service • Resource Reservation – Mechanism to exchange info about requests • Packet Scheduling – Manage queuing and scheduling. CSE331 Fall 2002 24