Improving Internet Performance through Traffic Managers

Transcription

1 Improving Internet Performance through Traffic Managers Ibrahim Matta Computer Science Department Boston University Computer Science

2 A Glimpse of Current Internet b b b b Alice c TCP b (Transmission Control Protocol) b c c c Bob b bc c A road system without traffic lights Less efficient, less predictable c Catherine

3 The Internet Traffic Managers (ITM) Architecture planning Capacity Internet Backbone Access Network Proactively send congestion notification Adaptive even before control queue is full programs Core router p i RED (Random Aggregate Early control Detection) Control Per-class programs Active Differentiated control queue length q q Traffic Manager Queue Management avg Proxy control (Edge Router) dropping prob. p i control Traffic The TM framework Active Queue Management (AQM): Identification Traffic characterization Edge-to-edge control Edge-to-edge Traffic Engineering Access Internet Per-flow Management network q backbone Capacity Planning hi q low Packet Classification Intelligent Transmission

4 TCP Flow Size Statistics Majority of the bytes on the Internet (up to 95%) attributed to the Transmission Control Protocol (TCP) Length (lifetime and size) of TCP flows is heavy-tailed (HT) Many short transfers (mice) and few very long (elephants) Elephants carry most of the bytes e.g., 20% of the largest flows carry 85% of the traffic Log(P[X>x]) Exponential (non-ht) HT P[X>x] x -α as x, 0<α 2 Log(x) HT distribution vs. non-ht distribution

5 TCP Flow Statistics AIX MAE-West at NASA-Ames 1 day trace: July 15, 2001 (from 24% 1 %

6 TCP Bandwidth Sharing TCP needs a slow start period to probe available bandwidth Congestion window size depends on when packet loss is detected On average short flows transfer at a lower rate than long flows Rate Initial handshake Slow-start Congestion avoidance How Size-aware can we flow fix this problem? differentiation! Time

7 Size-Based Differentiated Control Traffic Manager Allocate appropriate resources to aggregates of packet flows with divergent characteristics class-based scheduling preferential dropping class-based routing

8 TCP over Wireless Satellite Fixed Hosts TCP Connections Wired Internet How Proxy can TCP we control! fix this problem? Congestion Loss Router Gateway Base Station Handoff Mobile Host Base Station TCP fails to distinguish wireless errors and react appropriately

9 Proxy Control Traffic Manager p = 0.1% p = 0.2%

10 n TCP Flows competing conlink BW (n=1) Bottleneck Bandwidth (n=1) conglink BW (n=16) Bottleneck Bandwidth (n=16) bits bits time BW (delay*bw=1 pkt, n=1) Bottleneck Bandwidth (n=1) How Aggregate can we TCP fix this problem? control! time Bottleneck Bandwidth (n=16) BW (n=16 delay*bw= 1 pkt) bits time bits time

11 Aggregate TCP Control Traffic Manager

12 Differentiated Control: Preferential Treatment to Short Flows

13 Motivation: Mice Vs. Elephants Flow arrivals Rate=2 Fair Sharing: (2+2+8)/3 = 4 Help Mice: (1+1+8)/3 = 3.3 Help Elephant: (6+4+8)/3 = 6 Shortest-Job First gives lowest average response time Fair Sharing does not always give best solution Let mice go first mice response time cut in half, elephants are not affected Let elephants go first elephants save some response time, mice significantly penalized

14 Apply SJF to TCP Approximate packet flow scheduling by job scheduling (fluid flow assumption) Existing discriminatory job algorithms: SJF,SRPT,... Difficult to apply them to flow scheduling: Packets of a flow get multiplexed with other packets Limited flow information Limited job (flow) classes Flow transmission subjected to end-system controls (flow control, congestion control, etc.)

15 Size-aware Control Through ITM Traffic Manager: identify long and short flows No advanced knowledge on flow size threshold-based classification Core Router: per-flow throughput differentiation No per-flow information proportional dropping TCP throughput Throughput C = RTT p Threshold=4

16 Numerical Analysis of TCP+Sizeaware AQM Step 1: derive TCP session response time as a function of the link loss rate p Step 2: Kleinrock s Conservation Law: T(x,p)[1-B(x)]dx = E[X 2 ] / 2(1-ρ) Step 3: Numerically solve for p get the response function For differentiated control: assume proportional dropping w i p for class i

17 Simulation Results Thresh = 50 pkts Thresh = 500 pkts The numerical tool accurately predicts the response time Medium sized flows receive the most benefits Large flows are not penalized much

18 Impact of Flow Size Distribution Exponential distribution weight = 4 threshold = 50 offered load = 0.99 Large flows may suffer high penalty when flow size distribution is not skewed

19 Adaptive Control Adjust parameters to external factors

20 Adaptive Control Size distribution Optimal size threshold Link load Adaptive Controller The adaptive controller dynamically adjusts classification policy according to pre-defined optimization strategy

21 Example of Adaptive Control NLANR Trace, Indiana University Optimization constraint: Penalty < 3% pm pm 5am 2am 11am Size Threshold Gain (%)

22 Implementation Linux Prototype

23 Implementation of ITM the Linux Netfilter+IPtables API Local-in Routing Prerouting Application Local-out Forward Postrouting qos_ctrl: if pkt_cnt > thresh_ Mark as low priority else do nothing iptables p tcp t mangle A POSTROUTING j qos_ctrl set-threshold 50 Linux netfilter -- packet capturing Linux iptables -- flow tracking and packet mangling

24 Prototype of ITM Architecture Edge Router (Traffic Manager) running Linux netfilter+iptable2 Dynamic Module If pkt cnt > thresh_ Mark as low priority Else Do nothing WAN Emulator running DummyNet Web Traffic Generator West-Net Total BW: 4 Mbps RIO-PS/RED/ DropTail queue Ethernet East-Net Ethernet Cloud Web Traffic Generator Host3-Net Web Client Pool Emulator Host2-Net

25 Experimental Results Varying weight Varying threshold In agreement with analysis and simulation Quantitative discrepancies due to Linux TCP optimizations: Large initial window, SACK

26 Control Theoretic Analysis

27 Model

28 Linearization + TCP Open loop Buffer RED Variables denote perturbations around operating point Different set of equations for each operating point

29 Efficiency Presence of exogenous losses prevents throughput from matching bottleneck capacity

30 Stability Unstable if output or magnitude of oscillations increases monotonically with time Stability deduced from linearization is dependent on the operating region TCP is in fact switching between two different regions

31 Conclusions Traffic managers can be placed in strategic places in the Internet to provide efficient QoS support - in front of clients/servers, or - at exchange/peering points between administrative domains Reduce cost and enhance commercial competitiveness of Internet service providers and carriers Basic research in the control of complex dynamical systems, that of the Internet Experimental research in the implementation of a programmable traffic manager - programming interface to soft services, i.e. capabilities can be turned on/off and control parameters dynamically adjusted