An efficient traffic control scheme for TCP over ATM GFR services

(1)

An eﬃcient traﬃc control scheme for TCP over ATM

GFR services

Chia-Tai Chan

a

, Pi-Chung Wang

a,b,*

, Yaw-Chung Chen

b

a

Telecommunication Laboratories, Chunghwa Telecom Co. Ltd., Taipei 106, Taiwan, ROC

b

Department of Computer Science and Information Engineering, National Chiao Tung University, Hsinchu 30050, Taiwan, ROC Received 8 February 2000; received in revised form 13 June 2000; accepted 5 September 2001

Responsible Editor: J. Sterbenz

Abstract

In ATM networks, the guaranteed frame rate (GFR) service category has been defined to support user applications which are neither able to specify the range of traffic parameter values, nor able to comply with the behavior rules. It provides a packet-level rate guarantee with a given maximum packet size. The service specifies that the excessive traffic from a user should share the available resource fairly. In this paper, we investigate TCP/IP packet transmission over ATM by using a selective packet-discard strategy with tracking of the available buffer space and a packet push-out buffering scheme to accommodate the GFR service. The simulation results showthat our proposed method fulfills the requirements of GFR service as well as improves the TCP throughput under the common FIFO scheduling. A feasible implementation approach is also addressed. Ó 2001 Elsevier Science B.V. All rights reserved.

Keywords: ATM networks; Guaranteed frame rate; Selective packet-discard; Packet push-out buﬀering

1. Introduction

The guaranteed frame rate (GFR, formerly called UBRþ) service category has been proposed as an enhancement to the ATM UBR (unspeci-ﬁed bit rate) service that guarantees a minimum throughput for non-real-time applications at the frame level. This service category is intended for user applications which are neither able to specify traﬃc parameter values, nor able to comply with

the behavior rules required by existing ATM ser-vices [1–5]. The GFR service needs only minimal interactions between users and ATM networks. In addition, the service also specifies that the exces-sive traffic of each user should fairly share the available resource. As a result, designing an effi-cient buffer management scheme for supporting the GFR service is a key issue towards the suc-cessful deployment of GFR.

The eﬀective throughput or goodput is deﬁned as the throughput that is valid in terms of high-layer protocols. The goodput in terms of the higher-layer protocol over ATM networks may be quite low due to the wasted bandwidth for transmitting cells of corrupted packets, which can be caused by the

www.elsevier.com/locate/comnet

*

Corresponding author. Tel.: 5731851; fax: +886-3-5727842.

E-mail address: [email protected] (P.-C. Wang).

(2)

packet fragmentation. Here we assume that the cell transmission is error free. It has been demonstrated that the selective packet-discard strategies, i.e., early packet discard (EPD) and partial packet discard (PPD), alleviate the packet fragmentation problem and restore goodput [8].

In this study, we propose a traffic control scheme by using a selective packet-discard strategy with tracking of the available buffer space (ABS), and a packet push-out buffering scheme to support the quality of GFR service using a FIFO queuing discipline. The upper-layer protocol discussed here is TCP/IP. The merit of the proposed GFR traffic control approach is the improvement in both transmission efficiency and fair sharing of network resources, with feasibility of implementation. The organization of this paper is as follows: Section 2 gives an overviewof TCP flow-control behavior, its related packet-discard strategies and several GFR implementation alternatives. In Section 3, we present the functional approach of our proposed scheme. The system model and simulation results are discussed in Section 4. Section 5 concludes the work.

2. Overview

2.1. TCP ﬂow-control behavior

TCP uses a window-based ﬂow-control mech-anism; its window-adjustment algorithm consists of two phases. A connection begins with the slow-start phase. When a newconnection is established, its congestion window (CWND) is initialized to one segment. Upon receiving an ACK packet, the CWND is increased by one segment; this process continues until it reaches a slow-start threshold (SSTHRESH, typically 65,535 bytes). The sender can transmit up to either CWND or receiver’s advertised window, whichever smaller will be cho-sen. It can be shown that CWND actually in-creases exponentially every round-trip time. When congestion occurs (indicated by a timeout or by reception of duplicate ACKs), one half of the current window-size value (the smaller value be-tween the CWND and the receiver’s advertised window, with a minimum of two segments) is

saved in SSTHRESH. Additionally, if the con-gestion is indicated by an expired timer, the CWND will be set to one segment. If CWND were no greater than SSTHRESH, TCP is in slow-start phase; otherwise it is in congestion avoidance phase. In the latter case, CWND is increased by ((segment size segment size)/CWND) each time an ACK is received, which results in a linear in-crease of CWND. TCP Reno implements the fast retransmit and recovery algorithms [6] that allow the connection to quickly recover from isolated segment losses.

It is known that fast retransmit and recovery cannot recover multiple packet losses, it only causes the exponential increment phase to last a very short time, and the linear increment phase to begin with a very small window. As a result, TCP operates at a very lowrate and loses a certain amount of throughput. TCP newReno is a mod-iﬁcation to fast retransmit and recovery. In TCP newReno, the sender can recover from multiple packet losses without having to timeout [7]. 2.2. Related packet-discard strategies

An AAL-5 encoded TCP/IP data packet, with 1 or more bytes of TCP payload, cannot fit in a single ATM cell. The destination cannot reassem-ble the corrupted packet with any cell missing. In order to provide reliable and transparent data transport service, the source must be requested to retransmit the entire corrupted packet. Thus, whenever one of the cells constituting a packet is lost, all followed cell transmissions of that packet are certainly unnecessary. Therefore, the packet fragmentation may result in lowgoodput. The packet-discard strategies EPD and PPD are able to accommodate the packet fragmentation problem and restore goodput. In PPD, if a cell is dropped from a switch buffer, the subsequent cells belong-ing to the same higher-layer protocol data unit (e.g., TCP/IP packets) are discarded. It can be shown that PPD improves performance to a cer-tain level, but its goodput still needs improvement. Therefore, EPD has been proposed, in which the entire higher-layer protocol data unit is dropped when the switch buffer occupancy reaches a pre-defined threshold. In Ref. [8], the authors have

(3)

shown that the goodput of EPD is better than PPD.

The setting of the EPD threshold determines howefficiently the buffer can be used and how often the cell dropping may occur. In an analysis based on the worst-case assumption [9], the EPD scheme that ensures 100% goodput under overload situation requires an extra buffer space of one maximum packet length beyond the EPD thresh-old per active virtual connection. In Ref. [10], the switch sets up EPD threshold according to the number of active virtual circuits, and 100% good-put can be achieved with substantially smaller buffer space than that predicted by the worst-case analysis. Obviously, it improves the buffer utiliza-tion and the TCP throughput.

2.3. A feasible realization of GFR

As described in Section 1, the GFR service has been defined to provide traffic streams with a minimum packet rate guarantee for packets that do not exceed a maximum packet size. In addition, the service also specifies that the excess traffic of each user should fairly share the available re-source. An efficient GFR service scheduling strat-egy should alloweach flowpassing through a network node to get a fair resource share [11]. A fair queuing scheduler serves flows in proportion

to certain predetermined shares, as well as pro-tecting from the interference of ill-behaved sour-ces. Several fair queuing service disciplines have been discussed extensively in the literature. Ex-amples are weighted fair queuing [12], virtual clock [13], packet-by-packet generalized processor shar-ing [14,15], and self-clocked fair queushar-ing [16,17]. However, the drawbacks of these dynamic time-priority schemes are the high processing overhead for tracking the progress of tasks and scheduling the time-stamped packets. It has been suggested in Ref. [4] that a simple rate-guaranteeing discipline (i.e., weighted round robin) with per-VC queuing is indeed necessary to ensure GFR service. How-ever, the per-VC queuing would greatly compli-cate the switching system design. In Section 3, we propose a control approach that uses FIFO queu-ing in network switchqueu-ing nodes to support the GFR service requirements. We demonstrate that it is possible to use FIFO queuing instead of per-VC queuing to fulﬁll the requirements of GFR service.

3. Proposed GFR traﬃc control

GFR can be used between ATM edge devices. For example, IP routers connected through an ATM network can set up GFR VCs between them for data transfer. Fig. 1 illustrates such a case, in

(4)

which the ATM cloud connects LANs and routers. ATM end systems may also establish GFR VC connections between one another to obtain a mini-mum throughput guarantee. In ATM networks, the GFR traffic usually coexists with other higher pri-ority CBR/VBR traffic. Consequently, the available bandwidth allocated to GFR VCs varies dynami-cally. However, it is expected that dedicated band-width should be allocated for GFR services to avoid the service starvation. The GFR traffic should be served in the manner, such as class based weighted round robin, to achieve the configured service rate. To provide guaranteed GFR services, we pro-pose a packet discard and push-out buffering approach, which consists of a selective packet-discard mechanism with tracking of the ABS, and a packet push-out buffering scheme. The proposed buffering scheme ensures the fair sharing of net-work resources and avoids the misbehaved con-nection from occupying excessive buffer space. From the previous discussion, the selective EPD strategy could achieve 100% goodput that allevi-ates fragmentation problem. To further improve the TCP throughput, we introduce a selective packet-discard strategy with ABS tracking. Our scheme not only fulfills the GFR service require-ments, but also achieves 100% goodput that pro-vides a nearly optimal TCP throughput.

3.1. Selective packet-discard strategy with tracking of available buﬀer space

Given an ATM switch, we assume that an output port has a FIFO queue of size Q (in cell units) allocated for n GFR connections; also we assume that the maximum packet length is M cells. The EPD strategy is based on either a static threshold or dynamic threshold. In the latter, the switch sets the threshold according to the number of active virtual circuits. Our proposed method also uses a dynamic threshold based on the ABS. The switch maintains a state variable B to estimate the ABS, and a state variable Li to count the

in-coming packet length for connection i. The pro-posed method operates as follows:

1. Initially, B is set to Q, and Li is set to zero

(1 6 i 6 n).

2. If B is no less than one, the first cell of an ar-rived packet will be admitted into the buffer, and B will be set to B M. Otherwise, the switch will drop the first arriving cell and all subsequent cells belonging to the same packet. Note that if the incoming line is pumping cells into the buffer faster than the outgoing line of the switch, it still causes packet partial dropped. Therefore, if the arriving cell causes buffer over-flowthen the push-out process will be activated. The detail discussion of push-out process will be illustrated in the later.

3. Whenever a cell is transmitted, B is incremented by one. When the ﬁrst cell of a packet from con-nection i is admitted into the buﬀer, the Li is

incremented by one, and the switch starts to count the length of this incoming packet until its last cell is received, then B is updated to Bþ M Li.

There may be multiple cells arriving within one cell-slot time; under such situation, the cell service sequence will follow the predetermined order. The switch can identify the last cell of the incoming packet by checking the ATM-layer user-to-user indication bit speciﬁed in the cell header. Once a switch discards the ﬁrst cell of an incoming packet, it continues to discard the subsequent cells of that packet.

By tracking the ABS, the packet-discard method with dynamic threshold effectively improves the buffer utilization. As a result, our proposed method can achieve 100% goodput. Moreover, it improves the TCP throughput to a nearly optimal level. 3.2. Packet push-out buffering scheme

The design of simple and eﬃcient buﬀer man-agement approaches for accommodating GFR ser-vice requirements is an important issue toward the successful development of GFR. As described in Section 2.3, several sophisticated scheduling dis-ciplines and the per-VC queuing would greatly complicate switching system design, such as large separate queue structures, and scheduler state. In contrast, a FIFO service discipline is easy to implement and requires very little scheduler state. However, it has no protection between

(5)

well-behaved and misbehaved connections. It needs a specific buffering scheme to police the incoming packets to fulfill the GFR service re-quirements. A flowcan acquire a greater service share by sending more traffic to keep a higher occupancy in the FIFO queue. It is thus important to fairly allocate the network resources even in the presence of ill-behaved sources under heavy-load conditions. One possible solution is either to serve or to discard the arriving packets according to the fairness policy. The fairness of GFR services is defined that excess traffic from each GFR con-nection should share available resources fairly. The guaranteed minimum packet rate is first allo-cated for each VC, then the rest of the available bandwidth is shared equally (i.e., minimum packet rate plus equal sharing, MPRþ ES). There are other definitions of fairness, such as weighted al-location, which assigns available bandwidth to all active connections according to their relative per-formance demands.

A switch can use per-VC accounting to realize the dynamic sharing of the buffer space among all TCP flows. Let W be the current bandwidth available for GFR traffic in an output port, Q be the buffer size in cells and ribe the requested

ser-vice rate of VCi, where 1 6 i 6 n. In a FIFO service

discipline, the service rate rican be achieved if VCi

keeps an average buﬀer occupancy of bicells in the

FIFO queue, where ri W 6 bi Q6 bi Pn j¼1bj ð1 6 i 6 nÞ:

Then, VCi can obtain a service rate Wðbi=PbjÞ

which is no less than ri. That means the

through-put experienced by a connection VCi is

propor-tional to its average fraction of buﬀer occupancy, which can be preset to a threshold TLi. Hence, if

we keep the buﬀer occupancy of VCi at a desired

level, its service rate can be controlled. In this work, we adopt a simple MPRþES method, and assign an appropriate TLifor each VCias follows:

TLi¼ MPRi " þ W X n j¼1 MPRj ! wi # Q W; where wi¼ MPRi Pn j¼1MPRj ;

MPRi is the guaranteed minimum packet rate of

VCi and n is the total number of GFR

connec-tions.

Let NCibe the actual number of cells for VCiin

the buﬀer. Our packet push-out buﬀering scheme operates as follows:

1. When a ﬁrst packet cell of VCiarrives, if B > 0

the whole incoming packet will be admitted into the FIFO. If B 6 0 and NCiþ M is no greater

than its threshold TLi, the push-out process will

be activated. The push-out mechanism will also be activated for any incoming cell, excluding the first, that causes buffer overflow.

2. The queue manager selects a VCjwhich has the

maximum value of NCj TLj ð1 6 i; j 6 n; i 6¼

jÞ and pushes out its last packet. Then, the length of the pushed-out packet is added to B. Note that the queue manager should push-out packets continuously until B > 0. Then, accord-ing to our approach, it may ﬁnd ABS for the in-coming packet of VCi. It is possible that the

queue manager will push out a packet which has not yet completely arrived. Therefore, the queue manager will need to update state infor-mation so that cells from the push-out packet that have not yet arrived at the buﬀer will be discarded when they do arrive. The detailed im-plementation will be addressed in next section. 3. Otherwise, all cells of the incoming packet will

be discarded.

The functional diagrams of cell receiving and transmission for our proposed control approach is illustrated in Figs. 2 and 3, respectively.

It is true that smaller buffer size reduces good-put for TCP traffic over ATM. Since the proposed strategy is packet-based discard and push-out buffering scheme, it will cause certain problems that can be referred as a goodput beat down prob-lem, explained as follows. If the FIFO queue size is much smaller than nM (i.e., Q nM, where n is the number of VCs passing though the buffer) then most packets will experience partial discard. In some circumstances, it is likely that sources will be prevented from getting their effective throughput. To provide a certain level of goodput, a minimum buffer size is required. The proper buffer size

(6)

should be proportional to the number of VCs and the packet size.

A switch uses per-VC accounting to accomplish the dynamic sharing of the buffer space among all TCP flows. Cells belonging to the same VC are logically organized in a double-linked list. To en-sure a fair share of the available resource, we use push-out scheme to prevent the misbehaved con-nections from occupying excessive space in the FIFO queue. Then, a simple FIFO service disci-pline can fulfills the requirements of GFR services. 3.3. Realization of the proposed traffic control method

We have mentioned that our proposed control approach consists of a selective packet-discard

Fig. 2. The procedure of cell receiving in proposed control approach.

Fig. 3. The procedure of cell transmission in proposed control approach.

(7)

strategy and a packet push-out buﬀering scheme, which can be realized with an architecture con-sisting of four major functional components: a packet-discard controller (PDC), a cell dispatcher (CD), a stack controller (SC) and a push-out controller (POC), as illustrated in Fig. 4. Packets from the ith virtual connection with an assigned threshold TLiare deﬁned as legal if the number of

cells from the same connection in the FIFO queue is no greater than TLi.

The PDC discards packets when there is no ABS, and manipulates the control variables NCið0 6 i 6

nÞ and B. When a ﬁrst packet cell arrives, it will be accepted by PDC if there is ABS for the whole packet (i.e., B P 1). Otherwise, the packet will be discarded. The control variable NCiis used to turn

on/off the POC and SC for the ith virtual connec-tion. Once the first cell of an incoming packet arrives and is admitted into the buffer, if NCiPTLi M then all cells’ physical addresses

(PAs) of the arriving packet must be stored into the associated stack for possible push-out operation in the future, because it may be an illegal packet. However, an illegal packet may become legal later on. Therefore, if TLi>NCiPTLi M the PDC

immediately activates the SC to clear the associated stack before the ﬁrst cell’s PA of the incoming packet is inserted into it. This is because that cells’ PAs in the stack become legal and thus their asso-ciated cells cannot be discarded. If NCiPTLi, the

PDC will activate the SC to push the cell’s PA into its associated stack. The procedure for maintaining the stacks of PDC is illustrated in Fig. 5.

An arriving cell will be classiﬁed by the PDC, then written into the cell pool. Prior to this pro-cess, the associated stack number is extracted by the CD and its PA is stored into the POC. The POC consists of several modules, through which PAs are stored. Once the POC receives a push-out signal, it activates the SC to pop out illegal packets

(8)

from the stack with associated connection occu-pying excessive buﬀer space. The selected candi-date will be the VC that has the largest number of illegal packets in the associated stack. Each time the push-out routine is activated, illegal packets will be popped out continuously until the number of popped cells is greater than M, then the cell PAs of the popped packet will be returned to POC.

The SC consists of a set of stacks, each of which is associated with a connection, but not all con-nections need to use a stack. The purpose of the SC is to keep the illegal packets of a VC in the asso-ciated stack for possible push-out process. Since the selective discard strategy is packet based, an indication bit is used to identify the last cell’s PA of a packet. As a result, once the SC pops the last cell of a packet it continues to pop out remaining cells of the same packet. The information regarding the popped-out packet (i.e., the number of popped-out cells) must be sent to the PDC for the sake of maintaining an accurate NC value. In addition, it is possible that the popped-out packet is not com-plete, because part of the packet cells might have not arrived yet, thus the SC must inform the PDC

to discard the subsequent cells of the popped-out packet and keep the value of B correct.

The service discipline of POC is simply FIFO, in which the PA of a cell is stored in a 16-bit register. Once the newPA of an incoming cell is to be inserted into POC, the controller just appends the newPA to the FIFO modules. When a head-of-line cell is scheduled for transmission, its PA is retrieved from POC to identify the cell location in the cell pool. Once a cell has been transmitted, the content of all 16-bit registers will be shifted one position to the right; hence the PA in the POC will be overwritten automatically. In the mean time the transmitted cell’s PA will be written into the idle-address pool for future incoming cells. The de-tailed realization can be found in Ref. [18].

4. Simulation and performance evaluation

4.1. Simulation environment for small network conﬁguration

Three network conﬁgurations are chosen to illustrate the eﬀectiveness and scalability of our

(9)

proposed approach. We evaluate the small net-work configuration first, as shown in Fig. 6. Sim-ulations are performed based on this configuration for newReno TCP connections with 10 sources, given that the link transmission rate is 155 Mbps. The cell transmission time is about 2.78 ls, which is defined as a cell-slot time. The propagation delay is two slot times between the host and the switch, and eight slot times between switches. Ten TCP connections complete for network resources (i.e., buffer space, bandwidth) simultaneously and 15.5 Mbps is allocated for each connection. Assume that the 10th TCP source is greedy, and it sends packets at almost three times (i.e., 45 Mbps) of the allowed rate. The simulation time was 54 s which is much longer than any round-trip delay in real applica-tions; this is sufficient for the simulation process to reach the steady state as well as to collect the nec-essary information. The amount of transferred data is more than 1.2 Gb during the simulation period. The maximum TCP segment size is 512 bytes, common in IP networks. A buffer size of 200–1000 cells per port is appropriate for a small switch.

Under the FIFO service discipline, we investi-gate and compare the performance characteristics of our proposed strategy with EPD and EPDþ push-out strategies. The EPDþ push-out strategy is a combination of partial buﬀer sharing (PBS) with the well-known push-out scheme, in which an arriving packet is admitted only if the queue length is less than the EPD threshold. In addition to this, if the queue occupancy exceeds the EPD threshold and NCiþ M is no greater than its threshold TLi,

it will activate the push-out process.

The major performance measure considered here are the goodput, the TCP throughput and the fair sharing of the network resource. The following section explains the results of the proposed traﬃc control scheme.

4.2. Numerical results for small network conﬁgura-tion

4.2.1. Total goodput versus EPD threshold

The relationship between the total goodput and the buffer size beyond the EPD threshold is pre-sented here. The buffer size in excess of the EPD threshold varies from 1 to 5 segments. From the results shown in Fig. 7, the proposed traffic control method always experiences a 100% goodput. The major reason is that the selective packet-discard strategy tracks the ABS for each incoming packet, which will be admitted into the buffer only when there is enough buffer space. In addition, the push-out scheme is also a packet-based mechanism. Therefore, none of the partially received packet will occur in our control approach. Obviously, the proposed control scheme can optimize the goodput. On the other hand, in EPD and EPDþ push-out schemes the total goodput increases gradually as the buffer size in excess of the EPD threshold increases, as illustrated in Fig. 7. When the buffer size is equal to n times the maximum packet size, where n is the number of connections, EPD can achieve 100% goodput. Due to the effect of push-out process, the EPDþ push-out scheme achieves higher goodput than EPD scheme. The push-out process has little effect under the light traffic load, but it does improve the goodput under the heavy traffic condition. This is because a partially trans-mitted packet in the EPD control scheme may be transmitted completely with the EPDþ push-out control, which reduces the number of partially received packets. The major difference between EPDþ push-out and the proposed approach is the setting of the EPD threshold. Our proposed method uses a dynamic threshold based on the ABS. By tracking the ABS, the packet-discard method with dynamic threshold effectively improves the buffer utilization. The degree of throughput improve-ment may depend on the buffer size in excess of the EPD threshold.

(10)

4.2.2. Total goodput versus buﬀer size

The relationship between the total goodput and the buffer size is presented in Fig. 8. The total buffer size varies from 200 to 1000 cells. The buffer size in excess of the EPD threshold is two seg-ments. Although the total goodput increases as the buffer size increases in EPD and EPDþ push-out schemes, the total goodput of the proposed

ap-proach is even better, it is always maintained at 100% regardless of the change of buﬀer size. 4.2.3. TCP throughput versus EPD threshold

Fig. 9 shows the relationship between the total TCP throughput and the buffer size in excess of the EPD threshold. The total buffer size is 200 cells, and the buffer size in excess of the EPD threshold varies from 1 to 10 segments. Since increasing buffer uti-lization improves the total TCP throughput, the selective packet-discard strategy will drop a packet only if there is not enough buffer space. In EPD or EPDþ push-out strategy, it is possible that the buffer space is available but the incoming packet is still discarded because the buffer occupancy exceeds the EPD threshold. Through tracking the ABS, it can achieve much higher buffer utilization. Obvi-ously, our proposed strategy provides higher TCP throughput than EPD strategy, as shown in Fig. 9. Furthermore, we have also observed that in EPD and EPDþ push-out schemes, the total TCP throughput decreases gradually as the buffer size in excess of the EPD threshold increases for more than two segments. This is because the increase of the

Fig. 7. The goodput versus the buﬀer size in excess of the EPD threshold.

(11)

buffer size in excess of the EPD threshold will lessen the available buffer size to reduce, and it affects the loss probability of the incoming packets.

4.2.4. TCP throughput versus buﬀer size

With FIFO queuing, the data ﬂowof a misbe-haved TCP source may always occupy a large portion of the buﬀer space. In EPD, when

con-gestion occurs, even the well-behaved sources may suffer a large number of packet retransmission. By using our control approach, the number of packet retransmission is reduced by 35–40% for well-behaved sources with respect to EPD and 10– 12% with respect to EPDþ push-out. Moreover, the number of packet retransmissions for misbe-haved sources in our scheme is almost same as that of EPD. This result is the effect of combining the selective packet-discard strategy with packet push-out buffering scheme. It is obvious that the pro-posed strategy improves the total TCP throughput because it reduces the total number of retrans-mitted packets, as shown in Fig. 10.

The relationship between the TCP throughput and the buffer size is presented in Fig. 11. The total buffer size varies from 200 to 2000 cells. The buffer size in excess of the EPD threshold is two seg-ments. Fig. 11 shows that the larger the buffer size, the higher the TCP throughput. When the buffer size is set to 2000 (cells), the total TCP throughput of the proposed strategy is 96.8%, while it is 92.8% and 96.4% for EPD and EPDþ push-out strategy, respectively. Clearly, the total TCP throughput is improved in our proposed approach.

Fig. 9. The TCP throughput versus the buﬀer size in excess of the EPD threshold.

(12)

Note that the total TCP throughput is not af-fected significantly by increasing the buffer size from 400 to 800 cells in the EPD strategy, since the sender will recover from N segment losses in N round trips with new Reno TCP. The increase of buffer size leads to a longer round-trip delay, which causes the counter effect that reduces TCP throughput. As the increase of buffer size is large enough, it can improve the TCP throughput again. Obviously, a larger buffer size reduces the number of dropped and corrupted packets, and tends to improve the TCP throughput.

4.2.5. Fair sharing

As discussed earlier, it is an important issue to provide fair sharing of available resources for GFR service. The fairness measure is based on the fairness index, which is deﬁned asðPxiÞ

2

=ðnPx2 iÞ

[7], where xiis the TCP throughput of the ith TCP

source, and n is the number of TCP sources. If all sources share the available bandwidth more fairly, the fairness index will be much closer to 1. To further clarify the eﬀect of the proposed scheme, we also enlarge the propagation delay between switches to 600 slot times (i.e., equivalent to about 500 km) and examine the performance.

Under FIFO discipline, the proposed strategy prevents the misbehaving source from obtaining an unlimited share of the bandwidth. The misbe-having source will be punished through the packet push-out buﬀering scheme. Our control approach almost achieves a rate-guaranteed service and fair sharing for TCP sources, as illustrated in Table 1. Since the growth of round-trip delay increases the cost of packet recovery, it in turn reduces the TCP throughput. However, with the proposed scheme, the ill-behaved source gets more signiﬁcant pun-ishment for longer round-trip delay.

4.3. Simulation for large network topology

We have examined TCP traﬃc with a single TCP per VC in previous subsection. However, in a real network the edge switch of the IP/ATM net-work may multiplex multiple TCP connections over a single VC. In this subsection, we assume that each VC carries multiple TCP connections and a more realistic scenario of GFR [19] is used for the simulation, as shown in Fig. 12. There are local IP/ATM edge switch pairs interconnected by two ATM backbone switches through GFR VC. Each GFR VC carries traﬃc of 10 TCP connec-tions. The amount of allocated bandwidth for each VC is 10, 20, 30, 40 and 50 Mbps, respectively,

Fig. 11. The TCP throughput versus the buﬀer size.

Table 1

The fair sharing of EPD, EPDþpush-out and proposed strategies

TCP 1 2 3 4 5 6 7 8 9 10 Fairness index EPD strategy Short 12.61 12.68 12.72 12.65 12.92 12.62 12.63 12.79 12.6 24.7 0.937 Long 11.73 11.81 11.58 12.19 11.85 11.96 12.10 11.88 11.87 16.17 0.989 EPDþpush-out strategy Short 13.35 13.32 13.35 13.36 13.36 13.37 13.33 13.33 13.35 21.56 0.971 Long 12.71 12.61 12.63 12.76 12.74 12.82 12.42 12.78 12.59 13.74 0.999 Proposed strategy Short 13.49 13.43 13.45 13.48 13.44 13.5 13.41 13.43 13.42 21.42 0.973 Long 12.90 12.78 12.83 12.74 12.88 12.90 12.80 12.79 12.97 13.94 0.999

(13)

shared by the connected TCP sources. In addition, all sources are greedy TCP sources, i.e. they transmit data with twice of allowed transmission rate. The propagation delay is set to 1200 slot times (equivalent to approximately 1000 km). The remaining control parameters are the same to that in the small network conﬁguration.

Since TCP segments from diﬀerent sources are multiplexed on a single VC, the selective packet-discard strategy is unable to control the TCP rate accurately since the TCP source of the discarded packet in a VC may not be the malicious one. Thus it is diﬃcult to achieve intra-VC fairness. In this scenario, the selective packet-discard strategy should be able to dominate the per-VC rate instead of per-TCP rate for inter-VC fairness.

4.3.1. TCP goodput versus buﬀer size

The goodput performance is shown in Fig. 13. Again, the proposed scheme provides 100% good-put even with the multiplexed connections. Con-versely, the multiple greedy TCP connections will makes the buﬀer occupancy high, and thus reduce the goodput of EPD-based strategies.

4.3.2. TCP throughput versus buﬀer size

As shown in Fig. 14, the throughput is im-proved significantly as buffer size increases. This is mainly because all sources are greedy. Addition-ally, push-out based schemes benefit more than the EPD strategy when the buffer space is sufficiently

Fig. 12. TCP conﬁguration of large topology (50 sources).

Fig. 13. The goodput versus the buﬀer size.

(14)

large. In the EPD strategy, random drops may cause congestion to persist for a longer time; this results in further discarding and hence causes poor performance. Our proposed strategy pro-vides greater buﬀer utilization than the EPD-based strategy by dynamically tracking the ABS, and thus oﬀers better throughput.

4.3.3. Fair sharing

When a TCP source detects packet loss, the TCP CWND decreases, as does the transmission rate. Push-out based schemes drop packets from the worst behaved VC. However, the TCP source of the dropped packet may not be the most mis-behaved one (i.e. the TCP source with the largest CWND/allowed_rate ratio). Since the packet dis-carding controls the rate less precisely, the fairness index of our scheme is around 0.98, which is little lower than in the small conﬁguration simulation. However, since the packet dropping is random in the EPD strategy, the discarded packet may be-long to another irrelevant VC, thus it causes the congestion to last longer. As a consequence, more packets not belonging to the worst behaved VC may be dropped; further unfair dropping may result, as shown in Fig. 15.

4.4. Simulation for chain network topology

In this subsection, we examine a chain topol-ogy, as shown in Fig. 16. The greedy UDP source

is added into the topology and transmits data with twice of the allowed rate. The UDP and traffic of four TCP sources will pass through four switches and compete for resources with other TCP traffic. The propagation delay between two switches is 200 slot times (about 160 km), and the buffer size is 1000 cells. The throughput of both UDP source and TCP sources 1–9 are then analyzed.

4.4.1. Goodput, throughput and fair sharing The total goodput is shown in Table 2. The total goodput of the EPD-based strategies are re-duced. This is because the greedy UDP source keeps a high buﬀer utilization. It does not reduce the transmission rate since there is no ﬂowcontrol

Fig. 15. The fairness index versus buﬀer size.

(15)

in UDP. Thus, it lowers the goodput of EPD-based strategies, which is similar to the phenome-non in the large network topology, as described in Section 4.3.

The per-connection throughput, the total throughput, and the fairness index are listed in

Table 3. We can see that the UDP source gains much more bandwidth than other TCP sources, thus the throughput of TCP sources is greatly re-duced. However, the proposed strategy can still provide relatively high throughput and fairness. The misbehaved UDP ﬂowwill be dropped by the push-out control scheme, and the throughput is improved by adopting dynamic threshold.

It is known that there is a bias for TCP flows with different round-trip time. In Table 3, the propagation delay of the TCP source traffic 1–4 is much longer than that of TCP sources 5–9. Again, the push-out control scheme minimizes the

Table 2

The goodput of EPD, EPDþ push-out and proposed strategies EPD strategy EPDþ push-out strategy Proposed strategy Total goodput (%) 99.51 99.72 100 Table 3

TCP throughput and fairness index of diﬀerent control strategies

UDP TCP Total throughput Fairness index 1 2 3 4 5 6 7 8 9 EPD strategy 28.96 11.58 11.50 11.29 11.11 14.49 14.64 14.56 14.55 14.59 147.27 0.897 EPDþ push-out strategy 26.91 12.00 12.13 12.18 12.20 14.47 14.34 14.54 14.47 14.45 147.69 0.925 Proposed strategy 26.44 12.39 12.41 12.48 12.34 14.37 14.30 14.30 14.42 14.33 147.78 0.932

(16)

diﬀerence of throughput for TCP ﬂows with dif-ferent round-trip times.

4.4.2. TCP synchronization problem

A good buffer management scheme should avoid the occurrence of TCP synchronization be-cause it will degrade the throughput. The TCP synchronization might happen when the buffer overflowoccurs. The packets of multiple TCP con-nections are dropped almost simultaneously and result in most TCP sources decreasing their CWND. As a result, the buffer utilization is sig-nificantly reduced. In the EPD strategy, there is no explicit scheme to control the misbehaved source; it just drops the incoming packets as buffer over-flowoccurs, and might result in TCP synchroni-zation, as shown in Fig. 17.

In contrast, the proposed strategy always drops the packets of the most misbehaved source; hence the TCP synchronization can be avoided eﬀec-tively, as shown in Fig. 18.

The simulation results showthat our proposed approach achieves a minimum packet-level guar-antee and fair resource allocation for GFR services in various network conﬁgurations. Moreover, it

achieves 100% goodput and avoids the TCP syn-chronization problem. As a result, it brings the TCP throughput to a nearly optimal level.

5. Conclusions

The GFR services may be a key solution for supporting the TCP traffic over ATM networks. The implementation cost would affect the success of GFR service, however. The per-VC queuing is necessary to support the quality of GFR service, but may complicate the switch design. In this work, we have shown that it is possible to achieve the quality of GFR service through FIFO queuing. It is essential to have an efficient traffic control scheme to support the service requirements of GFR in network switches. Our proposed scheme consists of a selective packet-discard strategy and a packet push-out buffering scheme. This combina-tion ensures the fair sharing of network resources and prevents a misbehaved connection from oc-cupying excessive space in the FIFO queue. Al-though the push-out scheme complicates the buffer management, it achieves fair sharing of the

(17)

network resources and improves the total TCP throughput because it signiﬁcantly reduces the number of packet retransmissions. To further improve the TCP throughput, a selective packet-discard strategy using ABS tracking is introduced. This strategy increases the buﬀer utilization that leads to a nearly optimal TCP throughput.

The simulation results showthat our proposed approach achieves a minimum packet-level guar-antee for packets and the fair allocation of resource for GFR services. Moreover, it achieves 100% goodput and brings the TCP throughput to a near optimal level. The proposed GFR traﬃc control approach can be implemented in most switch ar-chitectures with state-of-the art ASIC technology.

References

[1] N. Yin, S. Jagannath, End-to-end traﬃc management in IP/ATM internetworks, ATM Forum Cont. 96-1406, October 1996.

[2] R. Guerin, J. Heinanen, UBRþ service category deﬁni-tion, ATM Forum Cont. 96-1598, December 1996. [3] S.K. Pappu, D. Basak, TCP over GFR implementation

with diﬀerent service disciplines: a simulation study, ATM Forum Cont. 97-0310, May 1997.

[4] D. Basak, S.K. Pappu, GFR implementation alternatives with fair buﬀer allocation scheme, ATM Forum Cont. 97-0528, July 1997.

[5] R. Goyal, R. Jain, S. Fahmy, B. Vandalore, S. Kalyanar-aman, GFR––proving rate guarantees with FIFO buﬀer to TCP traﬃc, ATM Forum Cont. 97-0831, September 1997. [6] W. Stevens, TCP slowstart, congestion avoidance, fast retransmit, and fast recovery algorithms, Internet RFC 2001, January 1997.

[7] R. Goyal, R. Jain, S. Fahmy, B. Vandalore, X. Cai, Selective acknowledges and UBRþ Drop policies to improve TCP/UBR performance over terrestrial and sat-ellite networks, ATM Forum Cont. 97-0423, April 1997. [8] A. Romanow, S. Floyd, Dynamics of TCP traﬃc over

ATM networks, IEEE Journal on Selected Area in Communications 13 (4) (1995) 633–641.

[9] J.S. Turner, Maintaining high throughput during overload in ATM switches, Proceedings of the IEEE INFOCOM’96. [10] M. Casoni, J.S. Turner, On the performance of early packet discard, IEEE Journal on Selected Area in Com-munications 15 (5) (1997) 892–902.

[11] P.-C. Wang, C.-T. Chan, Y.-C. Chen, An intelligent buﬀer management approach for GFR services in IP/ATM

internetworks, IEEE ICON’99, Brisbane, Australia, Sep-tember 1999, pp. 156–162.

[12] A. Demers, S. Keshav, S. Shenker, Analysis and simulation of a fair queueing algorithm, Journal of Internetworking Research and Experience (October 1990) 3–26.

[13] L. Zhang, Virtual clock: a newtraﬃc control algorithm for packet switching networks, Proceedings of the ACM SIGCOMM’90, Philadelphia, PA, September 1990, pp. 19– 29.

[14] A. Parekh, R. Gallager, A generalized processor sharing approach to ﬂowcontrol in integrated services networks: the single-node case, IEEE Transactions on Networking 1 (3) (1993) 344–357.

[15] A. Parekh, R. Gallager, A generalized processor sharing approach to ﬂowcontrol in integrated services networks: the multiple node case, IEEE Transactions on Networking 2 (2) (1994) 137–150.

[16] S. Golestani, A self-clocked fair queuing scheme for broadband applications, IEEE INFOCOM’94, Toronto, CA, June 1994, pp. 636–646.

[17] H. Zhang, Service disciplines for guaranteed performance service in packet-switching networks, Proceedings of the IEEE 83 (10) (1995) 1374–1396.

[18] C.-T. Chan, Y.-C. Chen, S.-C. Hu, P.-C. Wang, A guaranteed fair queuing approach through FIFO buﬀering, IEEE Communication Letters 4 (6) (2000) 205–207. [19] R. Goyal, R. Jain, S. Fahmy, B. Vandalore, Buﬀer

management for the GFR service, ATM Forum Cont. 98-0405, July 1998.

Chia-Tai Chan received his Ph.D. de-gree in Computer Science and Infor-mation Engineering from National Chiao Tung University, Hsinchu, Taiwan in 1998. He is now with the Telecommunication Laboratories Chunghwa Telecom Co. Ltd. His re-search interests include the design, analysis and traﬃc engineering of broadband multiservice networks.

Pi-Chung Wang received his Ph.D. degree in Computer Science and In-formation Engineering from National Chiao Tung University, Hsinchu, Taiwan in 2001. He is now with the Telecommunication Laboratories Chunghwa Telecom Co. Ltd. His re-search interests include the Internet multimedia communications, traﬃc control on high-speed network and L3/ L4 switching technology. He is a member of IEEE.

(18)

Yaw-Chung Chen received his Ph.D. degree in Computer Science from Northwestern University, Evanston, Illinois in 1987. During 1987–1990, he worked at AT&T Bell Laboratories as a member of Technical Staﬀ. In Au-gust 1990, he joined the faculty of the Department of Computer Science and Information Engineering, College of Electrical Engineering and Computer Science, National Chiao Tung Uni-versity, as an Associate Professor. He has been a Professor since 2000. His research interests include Internet traﬃc engineering, multimedia communications and high speed networking. He is a member of IEEE and ACM.