1 / 48

Feng, Kandlur, Saha, and Shin Transactions on Networking Oct. 1999

Adaptive Packet Marking for Maintaining End-to-End Throughput in a Differentiated-Services Internet. Feng, Kandlur, Saha, and Shin Transactions on Networking Oct. 1999. Abstract. Adaptive Priority Marking provides soft bandwidth guarantee in a differentiated-service Internet.

page
Download Presentation

Feng, Kandlur, Saha, and Shin Transactions on Networking Oct. 1999

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Adaptive Packet Marking for Maintaining End-to-End Throughput in a Differentiated-Services Internet Feng, Kandlur, Saha, and Shin Transactions on Networking Oct. 1999

  2. Abstract • Adaptive Priority Marking provides soft bandwidth guarantee in a differentiated-service Internet. • It does not requiredresource reservation. • It can be supported with minimum changes to the network in the form of priority handling for marked packets at the edges of the network.

  3. I. Introduction • Best-Effort Service • used in the current Internet • Integrated Service (INTSERV) /RSVP (Resource Reservation Protocol) • ▲ QoS guaranteed▼ Significant changes to the Internet infrastructure▼ Complexity • Differentiated Service (DIFFSERV) • Keep the network core as simple as possibleand leave the complexity (intelligence) to the network edges • At the edges • Packet classification and Type-of-Service (TOS) marking • At the core • Priority handling of packets according to TOS • ▲ Simplicity▼ QoS not guaranteed

  4. This paper • Offers • a modest enhancement to the best-effort service • Assumes • an one-bit priority scheme with lower loss rates for higher priority traffic • Assumes • TCP as the transport layer protocol that makes use of the feedback mechanism for measuring the throughput • Can be adapted for any transport protocol that is responsive to network congestion • Assumes • some incentives that encourage users from continually requesting the highest priority, such as usage-based pricing • ▲ Simple mechanisms for calculating near-optimal pricing based on congestion costs ?

  5. Adaptive packet marking • The user or network administrator • specifies a desired minimum service rate for a connection or connection group • and communicates this to a control engine (Packet-Marking Engine, PME) • By default, all packets are generated as low-priority packets • The PME monitors and sustains the requested level of service by setting the ToS bits in the packet headers appropriately: • If the observed service rate at the low-priority level either meets or exceeds the requested service rate, the PME simply monitors. • If the observed throughput falls below the minimum target rate, the PME starts prioritizing packets until the desired target rate is reached. • Once the target is reached, it strives to reduce the number of priority packets without falling below the minimum requested rate.

  6. II. TOS Architecture • Assumes • Two traffic types: • Priority • Best-effort • The traffic types are carried by the ToS bit in the IP header:

  7. Handling of multi-priority trafficat the core routers / gateways: • Separate queues • for different classes • with different scheduling priority • A common queue ☆ • for all traffic • with different packet drop preferences▲ simplify scheduling▲ maintain packet ordering → help TCP☆Used by this paper

  8. This paper takes • the common queue approach • and an enhanced RED (Random Early Detection) algorithm • Classical RED • Packets are dropped randomly with a given probability when the queue length exceeds a certain threshold. • The drop probability depends on the queue length and the time elapsed since the last packet was dropped. • Enhanced RED ☆ • The drop probabilities of marked packets are lower than that of unmarked packets

  9. The goal of this paper is • to develop packet marking schemesdeployed at the host network interfacethat will allow an individual connection or a connection groupto achieve a target throughputspecified by the user or the network administrator. • The objective of the packet marking scheme is • to monitor the throughput and to adjust the packet marking so that the sustained rate is maintained satisfying all the policy constraints. • Packet remarking is allowed at the provider boundary, but this paper considers where packets are marked only once.

  10. Two marking flavors: • Source-Transparent Marking • The PME is transparent and external to the host • It can be integrated into the infrastructure without affecting the hosts and routers • Source-Integrated Marking ☆ • The PME is integrated with the host • It can adapt better with the flow and congestion control used at the transport layer

  11. III. Source-Transparent Marking • The PME measures the local throughputinstead of end-to-end goodput • Simplicity: • It does not have to know the semantics of the transport layer in order to know whether or not the application data was actually delivered. • Even if it is aware of the transport layer semantics, it may not have access the stream of ACKs from the receiver to computer goodput, especially when the forward and return paths are different. • Protection from malicious or non-adaptive sources • The throughput is counted against itself • Measurement • Average bandwidth = amount of data packets / time window • Small window: biased to recent observation ☆ • Large window: biased to long-term observation

  12. TCP-Independent Algorithm • mprob = marking probability • obw = observed bandwidth • tbw = target bandwidth • increment • scale = difference between obw and tbw (Updated Periodically)

  13. Simulation by the “ns simulator” • Six nodes: n0 ~n5 • Bandwidth: as shown • Delay: 10ms • ERED: minth = 10 packets, maxth = 80 packets initial drop probability = 0.05 (unmarked) • Three connections: • C1: Infinite TCP best-effort • C2: Infinite TCP | tbw = 4Mbps • C3: 50s off/on TCP | tbw = 4Mbps • Update period = 100ms • Measured at n1 (edge router) ?

  14. C3=0,C2: SS+CA+FR C3=0,C2marked↓ C3↑,C2 marked↑slowly • (a) Step = 0.01 • Step = (scale*increment) • Marking rate of C2 lags behind the changes in the network load (C1+C2+C3)

  15. C3=0,C2: SS+CA+FR C3=0,C2marked↓ C3↑,C2 marked↑quickly • (b) Step = 1 • mprob =1 (all marked) or 0 (none marked) • Marking rate adapts quickly • Significant burstiness in both marked and unmarked streams • Achieving target bandwidth even at increased load

  16. (a) Packet trace of packet marking • The target rate is reached • Marked packets are cut down • Unmarked packets are increased • The target rate is not maintained • Marked packets are increased • Marked packets are cut down • Problem: Burstiness of both marked and unmarked streams

  17. TCP-Like marking probability update algorithm • Estimated number of packets in flight = obw × rtt • Estimated number of marked packets in flight (i.e. priority window)= pwnd = mprob × obw × rtt • For each ACK: update pwnd as shown in Fig.5 • Mprob = pwnd / (obw × rtt) ◆TCP-like: 1/cwndincrease no more than one per RTT (round trip time)

  18. (b) Packet trace of TCP-Like packet marking • Increase and decrease slowly

  19. TCP-like: (a)Transient • Very reactive to network load • C2 remains at or above its tbw most of the time • Changes in mprob is more network friendly (i.e. stable)

  20. (1) (2) (3) (4) • TCP-like: (b) Aggregate • BW = 10Mbps • One group of three TCP connections with total tbw = 6Mbps • One group of four best-effort TCP connections • (1)No BE traffic,∴Group 1 gets all 10Mbps • (2) One BEs, ∴ 10/4*3=7.5Mbps • (3) Four Bes, ∴10/7*3=4.3 < 6 ∴marking begins • (4) Group 1 stops, ∴ Group 2 gets all 10Mbps

  21. IV. Source-Integrated Approach • Source-Transparent Approach has little control on the flow and congestion control at the source, it often marks more packets than required. • TCP source fails to compete fairly with best-effort connections for its share of best-effort bandwidth. • Thus this paper experiments with a PME that is integrated with TCP sender in order to minimize the amount of marked packets.

  22. C1 total C1 marked C1 unmarked • Another Source Transparent Marking Exp. (Bandwidth) • C1:TCP tbw = 3M • C2~C6: BEs • C1 = 10/6 = 1.67M∴marking begins∴C1 stays at 3M • C2 = (10-3)/5 = 1.4M • C1 must mark a larger portion of its packets than it should(Why? See the next graph.)

  23. Another Source Transparent Marking Exp. (Window) • C1 cannot compete with a large amount of BEs,∴its window gets to zero?∴need marking more packets? • Thus a Source-Integrated Marking is needed!

  24. Customized TCP congestion control • Two windows: • pwnd (Priority window) • the number of marked packets that are in the network • bwnd (best-effort window) • the number of unmarked packets that are outstanding • When a loss • If marked, reduce bothpwnd and bwnd (∵severe congestion) • If unmarked, reduce bwnd only (∵minor congestion) • Two additional threshold values: • pssthresh (priority slow start threshold) • bssthresh (best-effort slow start threshold) • Like two fairly independentconnections • Slow start: when (pwnd < pssthresh) | (bwnd < bssthresh) • Congestion Avoidance: when … > ...

  25. p (SS) (CA) If < targetthen + b (SS) (CA) If > targetthen - • Source-Integrated Marking (Window Opening)

  26. If priority loss Fast ReTx If BEloss Fast ReTx • Source-Integrated Marking(Window Closing)

  27. C1 total C1 marked C1 unmarked • Source Integrated Marking Exp. vs. Fig. 7 (Bandwidth) • C1 total:same as Fig. 7 • C2same as Fig. 7 • C1 marked< that of Fig. 7C1 unmarked> that of Fig. 7≒ C2

  28. Source Integrated Marking Exp. vs. Fig. 7 (Window) • bwnd ≒ pwnd • i.e. Fair Bandwidth Sharing

  29. Assume • B = bandwidth of the bottleneck link • n = number of connections • Ri = target rate of connection i • ri = optimal marking rate with of Ri • b = best-effort bandwidth / “excess” for priority • If Ri < b, then no need to mark, i.e. using BE • The system constrains • ri + b = Ri(marked + unmarked = target) • Σri + nb = B(total marked + total unmarked = bottleneck BW)

  30. Integrated Marking Exp. (Bandwidth) • B = 10Mbps • C1 tbw = 3MC2 tbw = 2M6 BEs • t=0: C1, C2t=100: 2 BEst=200: +2 BEst=300: +2 Bes • 0-100: C1=C2=10/2=5 • 100-200:10/4=2.5∴C1 marks to 3∴Other (10-3)/3=2.33 ∴C1 marks3-2.33=0.67

  31. Integrated Marking Exp. (Bandwidth) • 200-300:10/6=1.67∴C1 marks to 3∴C2 marks to 2∴other (10-3-2)/4=1.25∴C1 marks 3-1.25= 1.75∴C2 marks 2-1.25= 0.75 • 300-400:10/8=1.25∴C1 marks to 3∴C2 marks to 2∴other (10-3-2)/6=0.83∴C1 marks 3-0.83= 2.17∴C2 marks 2-0.83= 1.17

  32. 2.17 1.75 1.17 0.75 0.67 • Integrated Marking Exp. (Marking Rate) • Ideal marking rates for C1 and C2

  33. Transparent Marking Exp. (Bandwidth) • Compared with Fig. 11(a) • More marked packets generated by the PME • More fluctuation

  34. Transparent Marking Exp. (Marking Rates) • Compared with Fig. 11(b) • More marked packets generated by the PME • More fluctuation

  35. V. Handling Over-Subscription • Adaptive Packet Marking relieves from the use of reservation signaling protocol, e.g. RSVP, and admission control. • When aggregate demand exceeds capacity, all connections with nonzero target rates carry only marked packets. ERED degrades to RED since there are only marked packets in the queue. • For transparent marking, TCP still behaves well. • For integrated marking, Modified TCP simply works the same as normal TCP. • Weighted-bandwidth sharing:additional priority bits, different ERED queue (weighted fair queuing or class-based queuing)

  36. Integrated Marking Exp. For Over-Subscription • C1,C1: tbw = 5C3,C4: tbw=10

  37. Integrated Marking Exp. For Over-Subscription • Fair share:10/4=2.5

  38. Integrated Marking Exp. For Over-Subscription • C1,C1: tbw = 5C3,C4: tbw=10

  39. Integrated Marking Exp. For Over-Subscription • C1,C1: tbw = 5C3,C4: tbw=10

  40. VI. Dealing with Non-Responsive • Non-responsive applications: • Applications which do not adapt to network dynamics • Can lead to severe network performance degradation and even congestion collapse

  41. T1~T4 gets all T1~T4 down to 9∵M1=1 T1~T4 down to 7∵M1=3 T1~T4 stay at 7even M1>3 M1 stays at 3 M1 flow • Non-responsive Flow (1) • B = 10Mbps • Four TCP: T1 ~ T4aggregate target rate = 7Mbps • A non-responsive flow: M1target rate = 3Mbps • When M1 > 3, M1 stays at 3?∵target rate = 3∴same marking∴ drop rate ↑ • Note: M3 marked = 0 & M3 unmarked = 3Solution: adopting FRED queue (Fair)

  42. Non-responsive Flow (2) - FRED • M1 marked =0 • M1 unmarked= 10/5 = 2(fair share) • T1~T4 unmarked= 10-2=8 > 7∴T1~T4 marked = 0

  43. VII. Deployment Issues • The Internet is heterogeneous and slow-evolving, thus most routers do not support service differentiation. • What if the network does not support end-to-end service differentiation? • For transparent marking, TCP still behaves well, • For integrated marking, it may be twice as aggressive?[Solution] This paper turns off the marking and window modifying.

  44. VIII. Conclusion • Adaptive Priority Marking provides soft bandwidth guarantee in a differentiated-service Internet. • Both the Transparent and Integrated Marking algorithms have advantages and disadvantages from the stand point of performance and deployment issues. • Future works: • Marking packets at multiple places in the network • Interaction and interoperability of other QoS supporting schemes • Two-priority to multiple-priority ToS schemes

More Related