Bandwidth measurement & Multicast

Transcription

1 Capacity vs. Available Bandwidth Capacity: Maximum throughput without cross-traffic Bandwidth measurement & Multicast Available bandwidth: Maximum throughput given crosstraffic A A A A 4 EECS 49 Computer Networks Z. Morley Mao Tuesday Nov, 004 Source Destination C C C C 4 Acknowledgement: Some slides taken from Kurose&Ross and Katz& Stoica Applications Existing Methods Efficiency of application Choose the best server Congestion control Multicast routing Etc Pathchar based (Jacobson) Packet pair based (Bolot) Nettimer (Lai and Baker) AMP (NLANR) OCXmon (NLANR) MRTG (Oetiker) Pathload (Dovrolis and Jain) 4 Current Techniques Pathchar pathchar Measures bandwidth of every link accurately Requires special software on only one host Slow Can consume significant amounts of network bandwidth Lots of UDP probes with different sizes and TTLs Estimates latency and bandwidth q q q 4 q n- n rtt from (n-)th to n-th *lat + ip_size / bw Dynamic behavior hard as queues neglected and various other assumptions Link-by -link measurement 5

2 Bandwidth Measurement Algorithms pathchar Packet-pair s = MTU Number of different packet sizes: h Total time: p s l i where h: number of i = hops, l i : round trip latency from sender to hop i, p: number of packets sent per size () 0-hop Ethernet network,0ms avglatency 44s s Avg bandwidth + to avg packet size probe a hop = roundtriplatency l i Total data transferred: s ( i) i = p h Send packets back-to-back and estimate the narrow link capacity from the packet dispersion C L/C Only measures end-to-end capacity while neglecting cross-traffic C L/C L/C C Packet Pair Bandwidth Measurement Algorithms Packet Pair Not statistically robust kernel density estimator Not scalable passive implementation Slow gradual packet pair implementation Not robust on all traffic Potential Bandwidth Filtering (PBF) Not flexible to bandwidth changes window Difficult to deploy Receiver Only Packet Pair Not studied under controlled conditions Simulation Two packets queued next to each other at bottleneck link exit the link t seconds apart: Assuming constant bottleneck separation: s t = s : size of second packet b bnl b bnl : bottleneck bandwidth If a packet queues in between: s = t b bnl s + s = t b bnl Bandwidth Measurement Algorithms Packet Pair Filtering (MBF) Bandwidth Measurement Algorithms Receiver and Sender Based PP How to filter noise caused by time compressed and extended packets? - Use mean or median of samples No! - Use histogram to find point of greatest density - Kernel density estimator algorithm: gives greater weight to samples closer to the point at which we want to estimate the density Simple and fast to compute Makes no assumption about distribution it operates on RBPP t is measured at the receiver: s = a a a i : arrival time of packet i SBPP uses round trip time: s b bnl = r i : arrival time of ACK of packet i r r Filtering techniques can be used to reject incorrect estimates RBPP is more accurate but harder to deploy ROPP sacrifices a little accuracy for ease of deployment b bnl

3 Bandwidth Measurement Algorithms Timeliness versus Accuracy Potential Bandwidth Filtering PP usually implemented to run over a fixed number of packets before providing estimate Solution: Calculate bandwidth gradually - Converges to correct bandwidth within packets - Problem: slow to detect bandwidth change - Solution: use packet window (size: w); BUT, may reduce stability Potential Bandwidth Problem: PP cannot measure a higher bandwidth than that at which the sender sends PBF correlate the potential bandwidth and measured bandwidth of a sample in deciding how to filter 4 Nettimer AMP by NLANR Packet pair technique can not measure more than the rate of transmission (potential bandwidth problem ) Solves the above problem using potential bandwidth filtering Active Measurement Project (AMP) monitors installed at critical points in the network. Measure RTT, Loss, Topology IPMP to remove problems due to ICMP Measured bandwidth Costly to install at each router Can become highly intrusive Potential bandwidth Only measures end-to-end capacity while neglecting cross-traffic Link-by -link measurement 5 OCXmon MRTG Passive network measurement tool by NLANR MRTG to monitor the traffic load on network links Deployed at strategic network locations Highly portable SNMP based tool A PC connected to a router via an optical splitter Specific hardware needs to be installed at the measurement points Available only for optical, DS, FDDI interfaces Provides only 5 min averages of link utilization Used by the network operators only as router SNMP community string information required Link-by -link measurement Link-by -link measurement

4 Pathload Motivation Sends Self-Loading Periodic Streams at increasing rates till the rate is larger than the tight link available bandwidth and the relative one way delays of packets show an increasing trend. This scheme is highly intrusive even though the scheme measures the available bandwidth of the tight link End-to-end available bandwidth measurement Combine active and passive approaches Most tools estimate narrow link capacity Accuracy Scalability Statistical robustness Not intrusive 9 0 Tight Link Identification Tight Link Identification (Contd.) Measurement packets Data record Version Type Length Checksum IP address Data Record (optional) Counter Data Record (optional) Timestamp 0 measurement packets sent in a second, to make the tool non-intrusive Speed Inserted/modified by the hops of the path Counter information from MIB-II in router Tight Link Identification (Contd.) Tight Link Identification (Contd.) 0 0 checksum S checksum B A... B... C... D checksum B D D 0 packets in one second Utilization of I-th interface at time t k U Ik c = t c t I ( k ) ( k ) Available bandwidth AIk = S I U Ik At least agree link of the estimates should concur about the tight link identity. Otherwise the next batch of 0 packets is sent. Ik k 4 4

5 Resource Route Available Bandwidth Estimation Method: ABEst ABEst (Contd.) More accurate, reliable estimate of the available bandwidth of the identified tight link Based on MRTG++ Reliably predicts the utilization of the link for a future interval, that varies in size k-p+ k k+h We use the past p samples to predict the utilization for the next h samples Utilize the covariance method for prediction Values of p and h varied according to the estimation error 5 Performance Evaluation Performance Evaluation (Contd.) Tight link identification ABEst Framework The Definition Bottleneck Traffic Engineering Automated Manager Management Plane Simulation Tool DiffServ/ MPLS Domain a_bw bottleneck LSP Routing Traffic Routing LSP Setup/ Dimensioning LSP Capacity Allocation LSP Preemption Traffic Engineering Tool (TET) Measurement/ Performance Evaluation Tool TEAM Network Planning 9 0 hop choke points 0 5

6 measurement packets Recursive Packet Train (RPT) 0 55 Load packets measurement packets 0 pkts, 0 B 0 pkts, 500 B 0 pkts, 0 B TTL S g R g R g R Transmission of RPT g g g Properties of Pathneck Issues of Pathneck Single end control - Relies on ICMP responses Low overhead - Load packets: 0 x 500B = 0KB - Measurement packets: x 0 x 0B =.KB - Instantaneous load (.4 TCP pkts) == TCP slow start Directly comparable - All the gap measurements are from the same packet train Firewall ICMP packet generation time Reverse path congestion Packet loss Route change Last hop 4 gap Choke Point choke points bottleneck point gap Confidence sv i- sv i 0 Choke point is a hop where is the gap value changes significantly hop 0 hop conf: The percentage of available bandwidth change - Assume node i is a choke point, it s confidence is Bottleneck point is the top ranked choke point We use dynamic algorithm to detect the choke points - Threshold is empirically selected as 0. 5

7 Ranking Patheneck: Algorithm Available bandwidth based ranking - The choke point that has the smallest available bandwidth is the bottleneck Hop position based ranking - The choke point that is closest to the destination is the bottleneck They are not equivalent - Because of the packet train squeezing. Probe the same dst N times (N=0). For each probing, label the choke points, and compute their conf values. If a hop appears in over N/ probings as a choke point candidate with conf > 0., it is a real choke point 4. The top ranked choke point is the bottleneck point Motivation Example: Internet Radio This approach does not scale (techno station) - Sends out Kb/s MP music streams - Peak usage ~9000 simultaneous streams only 5 unique streams (trance, hard trance, hard house, eurodance, classical) - Consumes ~.Gb/s bandwidth costs are large fraction of their expenditures (maybe 50%?) - If 000 people are getting their groove on in Berkeley, 000 unicast streams are sent from NYC to Berkeley Broadcast Center Backbone ISP 9 40 Instead build trees Multicast Routing Approaches Broadcast Center Copy data at routers At most one copy of a data packet per link Backbone ISP Kinds of Trees - Source Specific Trees - Shared Tree Tree Computation Methods - Link state - Distance vector Routers keep track of groups in real-time Routers compute trees and forward packets along them LANs implement link layer multicast by broadcasting 4 4

8 Source Specific Trees Source Specific Trees 4 5 Each source is the root of its own tree One tree per source Tree can consists of shortest paths to each receiver 4 5 Each source is the root of its own tree One tree per source Tree can consists of shortest paths to each receiver 0 0 Members of the multicast tree Sender 4 Very good performance but expensive to construct/maintain; routers need to manage a tree per source 44 Shared Tree Shared Tree Easier to construct/maintain but hard to pick good trees for everyone! 4 5 One tree used by all members in a group Ideally, find asteiner tree - the minimum-weighted tree connecting onlythe multicast members 5 0 The Steiner tree problem, succintly, is a minimum interconnection problem. The most basic version is in a graph: given a weighted graph in which a subset of vertices are identified as terminals, find a minimumweight connected subgraph that includes all the terminals. If the edge weights are all positive, then the resulting subgraph is obviously a tree. 4 Shared Tree Shared Tree Ideally, find a Steiner tree minimum-weighted tree connecting only the multicast members Finding Steiner Tree is NP hard Heuristics are known Alternatively, find a minimum-spanning tree minimum-weighted tree connecting all nodes in the network Finding a minimum spanning tree is much easier 0 4 4

9 Shared Tree Shared Tree Alternatively, find a minimum-spanning tree minimum-weighted tree connecting all nodes in the network Finding a minimum spanning tree is much easier. How? Alternatively, find a minimum-spanning tree minimum-weighted tree connecting all nodes in the network Finding a minimum spanning tree is easier. How? Prune back to get multicast tree Multicast Service Model Multicast Service Model (cont d) S [G, data] Net [G, data] R joins G [G, data] R 0 joins G [G, data] R n- joins G Receivers join a multicast group which is identified by a multic ast address (e.g. G) Sender(s) send data to address G Network routes data to each of the receivers R 0 R... R n- Membership access control - open group: anyone can join - closed group: restrictions on joining Sender access control - anyone can send to group - anyone in group can send to group - restrictions one which host can send to group Note: multicast vs. broadcast - Broadcast: packets are delivered to all end-hosts in the network - Multicast: packets are delivered only to end- hosts that are in (have joined) the multicast group 5 5 Multicast and Layering Multicast Implementation Issues Multicast can be implemented at different layers - data link layer e.g. Ethernet multicast - network layer e.g. IP multicast - application layer e.g. End system multicast Which layer is best? How are multicast packets addressed? How is join implemented? How is send implemented? How much state is kept and who keeps it?

10 Data Link Layer Multicast Problems with Data Link Layer Multicast Recall: end-hosts in the same local area network (LAN) can hear from each other at the data link layer (e.g., Ethernet) Reserve some data link layer addresses for multicast Join group at multicast address G - Network interface card (NIC) normally only listens for packets s ent to unicast address A and broadcast address B - To join group G, NIC also listens for packets sent to multicast address G (NIC limits number of groups joined) - Implemented in hardware, thus efficient Send to group G - Packet is flooded on all LAN segments, like broadcast - Can waste bandwidth, but LANs should not be very large Only host NICs keep state about who has joined scalable to large number of receivers, groups Single data link technology Single LAN - limited to small number of hosts - limited to low diameter latency - essentially all the limitations of LANs compared to internetworks 55 5 Network Layer (IP) Multicast IP Multicast Routing Overcomes limitations of data link layer multicast Performs inter-network multicast routing - relies on data link layer multicast for intra-network routing Portion of IP address space defined as multicast addresses - addresses for entire Internet through Open group membership Anyone can send to group - flexible, but leads to problems Intra-domain - Distance-vector multicast - Link-state multicast Inter-domain - Protocol Independent Multicast - Single Source Multicast 5 5 Distance Vector Multicast Routing Protocol (DVRMP) Reverse Path Flooding (RPF) An elegant extension to DV routing Use shortest path DV routes to determine if link is on the source-rooted spanning tree Three steps in developing DVRMP - Reverse Path Flooding - Reverse Path Broadcasting - Truncated Reverse Path Broadcasting Extension to DV unicast routing Packet forwarding - If incoming link is shortest path to source - Send on all links except incoming - Packets always take shortest path assuming delay is symmetric Issues - Some links (LANs) may receive multiple copies - Every link receives each multicast packet, even if no interested hosts s: s: s: s: s: s: s: s: s: s: s s r r

11 Example Reverse Path Broadcasting (RPB) Flooding can cause a given packet to be sent multiple times over the same link SS a b x z z y duplicate packet Chose parent of each link along reverse shortest path to source Only parent forward to a link (child link) Identify Child Links. Routing updates identify parent. Since distances are known, each router can easily figure out if it's the parent for a given link. In case of tie, lower address wins forward only to child link child link of x for S a b x SS 5 z z y Solution: Reverse Path Broadcasting Don t Really Want to Flood! Truncated Reverse Path Broadcasting (TRPB) This is still a broadcast algorithm the traffic goes everywhere Need to Prune the tree when there are subtrees with no group members Solution: Truncated Reverse Path Broadcasting Extend DV/RPB to eliminate unneeded forwarding Identify leaves - Routers announce that a link is their next link to source S - Parent router can determine that it is not a leaf Explicit group joining on LAN - Members periodically (with random offset) multicast report locally - Hear an report, then suppress own Packet forwarding - If not a leaf router or have members - Out all links except incoming r r NL L r r NL SS NL L 4 Pruning Details Pruning Details Prune (Source,Group) at leaf if no members - Send Non-Membership Report (NMR) up tree If all children of router R send NRM, prune (S,G) - Propagate prune for (S,G) to parent R On timeout: - Prune dropped - Flow is reinstated - Down stream routers re-prune Note: a soft-state approach How to pick prune timers? - Too long large join time - Too short highcontrol overhead What do you do when a member of a group (re)joins? - Issue prune-cancellation message (grafts) 5

12 Distance Vector Multicast Scaling Core Based Trees (CBT) State requirements: - O(Sources Groups) active state How to get better scaling? - Hierarchical Multicast - Core-based Trees Pick a rendevouz point for the group called the core. - Shared tree Unicast packet to core and bounce it back to multicast group Tree construction is receiver-based - Joins can be tunneled if required - Only nodes on One tree per group tree involved Reduce routing table state from O(S xg) to O(G) Core-Based Tree A tree rooted at the core is constructed for the group. State is maintained on a per-group basis. That is, each router keeps (G, list of interfaces to send). The notion of upstream/downstream interfaces is defined w.r.t. the core. Packets are forwarded by a router to its parent and children except the one on which the packets arrive. Core-Based Tree Core On tree relay router On tree router Sender Router with directly attached group member 9 0 Example Disadvantages Group members: M, M, M M sends data M root Sub-optimal delay Single point of failure - Core goes out and everything lost until error recovery elects a new core Small, local groups with non-local core - Need good core selection - Optimal choice (computing topological center) is NP hard M M control (join) messages data

13 Problems with Network Layer Multicast (NLM) NLM Reliability Scales poorly with number of groups - A router must maintain state for every group that traverses it - Many groups traverse core routers Supporting higher level functionality is difficult - NLM: best-effort multi-point deliveryservice - Reliability and congestion control for NLM complicated Deployment is difficult and slow - ISP s reluctant to turn on NLM Assume reliability through retransmission Sender can not keep state about each receiver - E.g., what receivers have received - Number of receivers unknown and possibly very large Sender can not retransmit every lost packet - Even if only one receiver misses packet, sender must retransmit, lowering throughput N(ACK) implosion - Described next 4 (N)ACK Implosion NACK Implosion (Positive) acknowledgements - Ack every n received packets - What happens for multicast? Negative acknowledgements - Only ack when data is lost - Assume packet is lost S R R R When a packet is lost all receivers in the sub-tree originated at the link where the packet is lost send NACKs S??? R R R 5 Barriers to Multicast Application Layer Multicast (ALM) Hard to change IP - Multicast means change to IP - Details of multicast were very hard to get right Not always consistent with ISP economic model - Charging done at edge, but single packet from edge can explode into millions of packets within network Let the hosts do all the special work - Only require unicast from infrastructure Basic idea: - Hosts do the copying of packets - Set up tree between hosts Troublesome security model - Anyone can send to a group - Denial-of-service attacks on known groups Example: Narada [Yang-hua et al, 000] - Small group sizes <= hundreds of nodes - Typical application: chat

14 Narada: End System Multicast Algorithmic Challenge Gatech Stanford Stan CMU Stan Berk Choosing replication/forwarding points among hosts - how do the hosts know about each other - and know which hosts should forward to other hosts Berkeley Berk Gatech Overlay Tree Stan Stan CMU Berk Berk 9 0 Advantages of ALM Performance Concerns No need for changes to IP or routers No need for ISP cooperation End hosts can prevent other hosts from sending Easy to implement reliability - use hop-by-hop retransmissions Stretch - ratio of latency in the overlay to latency in the underlying network Stress - number of duplicate packets sent over the same physical link Performance Concerns Single Sender Multicast Gatech Delay from CMU to Berk increases Stan Stan Many problems with IP multicast disappear if each group is associated with a single source Duplicate Packets: Bandwidth Wastage CMU CMU Gatech Berk Stanford Berkeley Berk Stan Stan Berk Berk Hosts joining multicast group can send join messages to source - this sets up delivery tree - no worry about root being in wrong place This solves several problems: - better security and charging model - simple algorithm 4 4

15 Example What s Wrong with SSM? Group members: M, M, M source Multiple sources? - Can set up group per source, or... - Source can serve as relay for other senders M Algorithm? - Trivial M M So, why isn t SSM the answer? - Multicast no longer serves as rendezvous - Ok for broadcast apps, not good for meeting apps control (join) messages data 5 What Do You Need to Know? DVRMP CBT SSM How they compare 5