W. Feng, D. Kandlur, D. Saha, and K. Shin, "Techniques for eliminating packet loss in congested TCP/IP networks," University of Michigan, Tech. Rep. CSE-TR-349-97, November 1997.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....enhancement to VQ, called MCR scheduling [SIU97] proposes the emulation of a weighted scheduler to provide Minimum Cell Rate (MCR) guarantees to ATM connections. In this scheme, a per VC, weighted variable (W i ) is updated in proportion to the VCs MCR, and compared with a global threshold. [FENG] proposes a combination of a Packet Marking Engine (PME) and an Enhanced RED scheme based on per connection accounting and multiple thresholds (MA MT) PME ERED is designed for the IETF s differentiated services architecture, and can provide loose rate guarantees to connections. The PME measures ....

Wu-chang Feng, Dilip Kandlur, Debanjan Saha, Kang G. Shin, "Techniques for Eliminating Packet Loss in Congested TCP/IP Networks," University of Michigan, CSE-TR-34997, November 1997.

....the router, so that it can control its average queue length regardless of the number of active TCP connections. However, SRED still has many control parameters, and configuration difficulty of these control parameters is not addressed. Moreover, another approach called Adaptive RED is proposed in [14]. Adaptive RED dynamically increases or decreases the maximum packet marking probability max p , which is one of the RED control parameters, according to its average queue length. More specifically, when the average queue length is smaller than min th , Adaptive RED decreases max p by (1 # 1) ....

W.-C. Feng, D. D. Kandlur, D. Saha, and K. G. Shin, "Techniques for eliminating packet loss in congested TCP/IP networks," Tech. Rep. CSE-TR-349-97, Oct. 1997.

....to congestion notification. RED also must ensure that congestion notification is given at a rate which sufficiently suppresses the transmitting sources without underutilizing the link. Unfortunately, when a large number of TCP sources are active, the aggregate traffic generated is extremely bursty [7, 10]. Bursty traffic often defeats the active queue management techniques used by RED since queue lengths grow and shrink rapidly, well before RED can react. Figure 1 shows a simplified pictorial example of how RED functions under this congestion scenario. The congestion scenario presented in Figure ....

....and adjust their transmission rates. Finally, at t = 7, a decrease in offered load at the bottleneck link is observed. Note that it has taken from t = 1 until t = 7 before the offered load becomes less than the link s capacity. Depending upon the aggressiveness of the aggregate TCP sources [7, 10] and the amount of buffer space available in the bottleneck link, a large amount of packet loss and or deterministic ECN marking may occur. Such behavior leads to eventual underutilization of the bottleneck link. One way to solve this problem is to use a large amount of buffer space at the RED ....

W. Feng, D. Kandlur, D. Saha, and K. Shin. Techniques for Eliminating Packet Loss in Congested TCP/IP Networks. In UM CSE-TR-349-97, October 1997.

....reaches max p at average queue size max th ; subsequently all packets are dropped w.p. 1. While optimum values of RED parameters as well as enhancements to the preliminary This work was supported in part by AFOSR Grant F49620 1 0472 RED algorithm [4] are still subject to ongoing research (e.g. [7], 8] the key objective of RED can be stated as follows: packet drop (i.e. congestion notification) must be generated at a rate sufficient to prevent packet loss due to buffer overflow (otherwise, RED will act as a Tail Drop) while sustaining high link utilization. In view of the Internet ....

W. Feng, D. Kandlur, D. Saha, and K. Shin, "Techniques for eliminating packet loss in congested TCP/IP networks," Technical Report CSE-TR349 -97, University of Michigan, 1997.

....ECN into ERED can potentially alleviate many of the problems in using TCP over differentiated services networks. We are also experimenting with modifications to TCP s linear increase mechanism, in order to alleviate the quadratic dependence of congestion notification on the number of connections [3]. Without such modifications, a large percentage of packets must be marked in order to prevent packet loss. Finally, we are experimenting with a new active queue management algorithm in which marking is not triggered by ei15 ther the instantaneous or average queue lengths. One of the problems with ....

W. Feng, D. Kandlur, D. Saha, and K. Shin. Techniques for Eliminating Packet Loss in Congested TCP/IP Networks. UM-CSE-TR-349-97, November 1997.

....predictable average delays with RED would require constant tuning of the parameters to adjust to current traffic conditions. Our goal in this paper is to solve this problem with minimal changes to the overall RED algorithm. To do so, we revisit the Adaptive RED proposal of Feng et al. from 1997 [6, 7]. We make several algorithmic modifications to this proposal, while leaving the basic idea intact, and then evaluate its performance using simulation. We find that this revised version of Adaptive RED, which can be implemented as a simple extension within RED routers, removes the sensitivity to ....

....basic RED design, and our purpose here is to look for a more minimal change to RED that can alleviate the problems of variable delay and parameter sensitivity mentioned above. We think that the basic insight for such a solution lies in the original Adaptive RED proposal of Feng et al. from 1997 [6, 7]. This proposal retains RED s basic structure and merely adjusts the parameter max p to keep the average queue size between min th and max th . In this paper, we describe a new implementation of Adaptive RED which incorporates several substantial algorithmic changes to the original Adaptive RED ....

[Article contains additional citation context not shown here]

W. Feng, D. Kandlur, D. Saha, and K. Shin. Techniques for Eliminating Packet Loss in Congested TCP/IP Network. U. Michigan CSE-TR-349-97, November 1997.

....(SACK) 14] The Explicit Congestion Notification (ECN, one bit explicit feedback) scheme [19] requires both end to end and network support. Network based enhancements are aimed at improving fairness and throughput. 4 These include mechanisms like scheduling [5] improved packet discard policies [7, 6, 13, 12] and explicit rate control of TCP [20, 15] The end to end proposals aim to provide better filtering to detect congestion events from loss events, improved retransmission, and attempt to reduce the occurrence of timeouts. Providing fairness and queueing delay control have required some form of ....

W. Feng, D. Kandlur, D. Saha, K. Shin, "Techniques for Eliminating Packet Loss in Congested TCP/IP Networks," U. Michigan CSE-TR-349-97, November 1997.

....in per connection goodput) Ibanez et al. [11] observe similar problems when TCP runs over the assured service. Kim et al. [15] use an improved version of TCP and see fewer problems. Feng et al. propose an adaptive marker, TCP modifications and an improved PHB implementation using the ERED algorithm [8]. In contrast, we postulate that one important reason for these performance problems is the lack of TCP friendly building blocks (droppers, markers, shapers, PHBs) in the network. By TCP friendly we mean the capability of these building blocks to: ffl provide for packet interleaving, ffl ....

....of the ack rate as suggested here. 3. 4 Drop policies and Edge to edge feedback control Drop policies are part of the per hop behavior (PHB) RED is the first example of a TCP friendly building block (as compared to drop tail) Examples of even more TCP friendly drop policies are FRED and ERED [7, 8] which can allocate loss probabilities during a congestion event in a more controlled manner. While we have work in progress in this area, we do not present results about this dimension in this paper. Edge to edge feedback control introduced in an earlier work by our group [13] involves a simple ....

W. Feng, D. Kandlur, D. Saha, K. Shin, "Techniques for Eliminating Packet Loss in Congested TCP/IP Networks," U. Michigan CSE-TR-349-97, November 1997.

....between 30 and 55 seconds (causing congestion at the link CR3 to CR2) and approaches zero until approximately 80 seconds of simulation time. Permanent queueing happens at the congested output port of CR1 served by the bottleneck link in direction CR2. Fig. 3. Simulated network As shown in [27], RED s aggressiveness regarding packet dropping has to be adjusted dynamically to the number of flows traversing the output port of a congested router in order to avoid convergence of the average queue size to the minimum or maximum queue size threshold. RED s aggressiveness can be adjusted by ....

....threshold. RED s aggressiveness can be adjusted by increasing the maxp parameter if avg becomes smaller than minth and decreasing maxp in case avg becomes greater than maxth. For simulations in this paper a modified version of WRED with adoption of maxp f has been implemented. Similar to [27], we have set the decrease increase factor for maxp f to 2. Other WRED parameters are set as follows: byte mode is activated, w q = 0.002, minth uf = 5000 bytes, maxth uf = 9000 bytes, maxp uf = 1, minth f = 10000 bytes, maxth f = 17500 bytes, maxp f = 1 50, mean pktsize = 500 bytes, total ....

W. Feng et al., "Techniques for eliminating Packet Loss in congested TCP/IP Networks", University of Michigan, CSE-TR-349-97, November 1997

....in turn depends on its arrival rate, which is higher for unresponsive traffic due to the lack of end to end congestion control. Additionally, a small drop probability persists even for responsive low demand traffic, preventing these flows from reaching their fair share. 2. 2 ARED: Adaptive RED In [22] it has been shown that the rate at which RED signals congestion indications (i.e. packet drops or ECN signals) to TCP sources has to be directly proportional to the number of TCP flows traversing the output port. RED s congestion indication rate can be adapted by varying the ....

.... min th : max p = max p beta The parameters alpha and beta are set to 3 and 2, respectively. It is shown by simulation that these changes make the average queue size converge between the minimum and maximum threshold, independently of the number of flows traversing the output port. As shown in [22], steady state analysis of TCP in [23] 24] gives some insight why RED s aggressiveness should depend on the number of TCP connection traversing the output port. The Rate (R) of a TCP flow in the congestion avoidance phase can be modeled as a function of the flow s RTT, Maximum Segment Size (MSS) ....

[Article contains additional citation context not shown here]

W. Feng et al., "Techniques for eliminating Packet Loss in congested TCP/IP Networks"

No context found.

W. Feng, D. D. Kandlur, D. Saha, K. S. Shin, Techniques for eliminating packet loss in congested TCP/IP networks, Tech. Rep. CSE-TR-349-97, U. Michigan (4 1997).

....to congestion notification. RED also must ensure that congestion notification is given at a rate which sufficiently suppresses the transmitting sources without underutilizing the link. Unfortunately, when a large number of TCP sources are active, the aggregate traffic generated is extremely bursty [4, 6]. Bursty traffic often defeats the active queue management techniques used by RED since queue lengths grow and shrink rapidly, well before RED can react. Figure 1 shows a simplified pictorial example of how RED functions under this congestion scenario. The congestion scenario presented in Figure ....

....and adjust their transmission rates. Finally, at t = 7, a decrease in offered load at the bottleneck link is observed. Note that it has taken from t = 1 until t = 7 before the offered load becomes less than the link s capacity. Depending upon the aggressiveness of the aggregate TCP sources [4, 6] and the amount of buffer space available in the bottleneck link, a large amount of packet loss and or deterministic ECN marking may occur. Such behavior leads to eventual underutilization of the bottleneck link. One way to solve this problem is to use a large amount of buffer space at the RED ....

W. Feng, D. Kandlur, D. Saha, and K. Shin. Techniques for Eliminating Packet Loss in Congested TCP/IP Networks. In UM CSE-TR-349-97, October 1997.

....to congestion notification. RED must also ensure that congestion notification is given at a rate which sufficiently suppresses the transmitting sources without underutilizing the link. Unfortunately, when a large number of TCP sources are active, the aggregate traffic generated is extremely bursty [8, 9]. Bursty traffic often defeats the active queue management techniques used by RED since queue lengths grow and shrink rapidly, well before RED can react. Figure 1(a) shows a simplified pictorial example of how RED functions under this congestion scenario. The congestion scenario presented in ....

....and adjust their transmission rates. Finally, at t = 7, a decrease in offered load at the bottleneck link is observed. Note that it has taken from t = 1 until t = 7 before the offered load becomes less than the link s capacity. Depending upon the aggressiveness of the aggregate TCP sources [8, 9] and the amount of buffer space available in the bottleneck link, a large amount of packet loss and or deterministic ECN marking may occur. Such behavior leads to eventual underutilization of the bottleneck link. One way to solve this problem is to use a large amount of buffer space at the RED ....

W. Feng, D. Kandlur, D. Saha, and K. G. Shin. Techniques for Eliminating Packet Loss in Congested TCP/IP Networks. In UM CSE-TR-349-97, October 1997.

....early random drop mechanism to correct this problem, but this requires a per flow accounting. The RED algorithm works well in times of moderate congestion. In case of heavy congestion, the RED fails to provide benefit to the network. This issue and enhancements to the RED algorithm are studied in [22, 23, 24]. These papers propose adaptive RED algorithms that estimate the number of active flows based on different information. 2.2.1 The RED Algorithm The RED algorithm tries to keep the average queue length between two thresholds. No packet is dropped while the average queue size is below the lower ....

....2, p max = 0:1 tions and the capacity of the bottleneck link. One solution to this problem could be an adaptive RED algorithm, which modifies the marking probability based on an estimate of the number of active connections as proposed in [24] or based on the average queue length as proposed in [22, 23]. In [23] it is shown that no single set of RED parameters will work well for different scenarios. If the value of p max is too small for a high load (many active flows) the average queue size of the RED gateway will be close to the upper threshold and often even exceed it. Therefore, most of the ....

W. Feng, D. Kandlur, D. Saha, and K. Shin. Techniques for Eliminating Packet Loss in Congested TCP/IP Networks. Technical report, CSE-TR-349-97, U. Michigan, 1997.

....to congestion notification. RED must also ensure that congestion notification is given at a rate which sufficiently suppresses the transmitting sources without underutilizing the link. Unfortunately, when a large number of TCP sources are active, the aggregate traffic generated is extremely bursty [18, 21]. Bursty traffic often defeats the active queue management techniques used by RED since queue lengths grow and shrink rapidly, well before RED can react. Figure 4.1 shows a simplified pictorial example of how RED functions under this congestion scenario. The congestion scenario presented in ....

....and adjust their transmission rates. Finally, at t = 7, a decrease in offered load at the bottleneck link is observed. Note that it has taken from t = 1 until t = 7 before the offered load becomes less than the link s capacity. Depending upon the aggressiveness of the aggregate TCP sources [18, 21] and the amount of buffer space available in the bottleneck link, a 44 L Mbs A B Sinks Sources Sending rate = L Mbs Queue drops and or ECN marks exactly the correct amount of packets to keep sending rate of sources at L Mbs Sinks generate DupAcks or ECN Figure 4.2: Ideal scenario large ....

W. Feng, D. Kandlur, D. Saha, and K. Shin. Techniques for Eliminating Packet Loss in Congested TCP/IP Networks. In UM CSE-TR-349-97, October 1997.

No context found.

W. Feng, D. Kandlur, D. Saha, and K. Shin, "Techniques for eliminating packet loss in congested TCP/IP networks," University of Michigan, Tech. Rep. CSE-TR-349-97, November 1997.

No context found.

W. Feng, D. Kandlur, D. Saha, and K. Shin. Techniques for eliminating packet loss in congested TCP/IP networks. Technical Report CSE-TR-349-97, University of Michigan, November 1997.

No context found.

W.-C. Feng, D. D. Kandlur, D. Saha, and K. G. Shin, "Techniques for eliminating packet loss in congested TCP/IP networks," Tech. Rep. CSE-TR-349-97, Oct. 1997.

No context found.

W. Feng, D. Kandlur, D. Saha, and K. Shin, "Techniques for Eliminating Packet Loss in Congested TCP/IP Networks," Tech. Rep. CSE-TR-349-97, Univ. of Michigan, 1997.

No context found.

W.-C. Feng, D. D. Kandlur, D. Saha, and K. S. Shin, "Techniques for eliminating packet loss in congested TCP/IP networks," Tech. Rep. CSE-TR-349-97, U. Michigan, Apr. 1997.

No context found.

Wu-chang Feng, Dilip Kandlur, Debanjan Saha, Kang G. Shin, "Techniques for Eliminating Packet Loss in Congested TCP/IP Networks," ____________.

No context found.

W.-C. Feng, D. D. Kandlur, D. Saha, and K. G. Shin, "Techniques for eliminating packet loss in congested TCP/IP networks," Tech. Rep. CSE-TR-349-97, Oct. 1997.

No context found.

W. Feng, D. Kandlur, D. Saha, K. Shin, "Techniques for Eliminating Packet Loss in Congested TCP/IP Networks", U. Michigan CSE-TR-349-97, November 1997.

No context found.

W. Feng, D. Kandlur, D. Saha, and K. Shin. Techniques for Eliminating Packet Loss in Congested TCP/IP Networks. Technical Report U. Michigan CSE-TR-349-97, November 1997.

No context found.

W. Feng, D. Kandlur, D. Saha, K. Shin, "Techniques for Eliminating Packet Loss in Congested TCP/IP Networks, " U. Michigan CSE-TR-349-97, November 1997.

First 50 documents

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC