* Correspondence to: Chyan Yang, National Chiao Tung University, Institute of Information Management, Mb307, 1001 Ta Hsueh Road, Hsin Chu 300, Taiwan.
R E-mail: cyang@cc.nctu.edu.tw
Received December 1999
A comparison of congestion control and time slot algorithms
in Internet transmission performance
Chyan Yang * R and Chen-Hua Fu
National Chiao Tung University,Institute of Information Management,Mb307,1001 Ta Hsueh Road,Hsin Chu 300,Taiwan National Defense Management College,Graduate School of National Defense Information,
P.O. Box 90046-15,Chung-Ho,Taipei 235,Taiwan
SUMMARY
The characteristics of TCP and UDP lead to di!erent network transmission behaviours. TCP is responsive to network congestion whereas UDP is not. This paper proposes two mechanisms that operate at the source node to regulate TCP and UDP #ows and provide a di!erential service for them. One is the congestion-control mechanism, which uses congestion signal detected by TCP #ows to regulate the #ows at the source node. Another is the time-slot mechanism, which assigns di!erent number of time slots to #ows to control their #ow transmission. Based on the priority of each #ow, di!erent bandwidth proportions are allocated for each #ow and di!erential services are provided. Simulation results show some insights of these two mechanisms. Moreover, we summarize the factors that may impact the performance of these two mecha-nisms. Copyright 2001 John Wiley & Sons, Ltd.
KEY WORDS: di!erential service; network congestion; congestion-control mechanism; time-slot mechanism; #ow priority
1. INTRODUCTION
TCP and UDP are the two major protocols over the Internet. These two protocols have di!erent tra$c transmission operations. TCP is connection orientated whereas UDP is connectionless. These characteristics of TCP and UDP lead to di!erent network-transmission behaviours. Since most of the Internet applications are based on TCP, the performance of TCP will impact on the Internet e$ciency. The focus of this study is how to improve the TCP transmission performance and restrict the excessive bandwidth taken by UDP transmissions.
Note that there are about a dozen internet-drafts and RFCs related to our subject using the term of &di!erentiated services', &quality of service' or various types of &forwarding' behaviours [1,2]. In the near future, however, Di!Serv-aware devices will still be rare [3]. This is why this
the di!erential services can be provided at the source node to enhance the transmission perfor-mance of higher priority tra$c #ows.
1.1. Congestion control mechanism
The congestion control mechanism is a source-based tra$c #ow-control mechanism. The conges-tion signal from TCP #ows can be used as a congesconges-tion indicator for the source node; this could help the source node control the TCP and UDP tra$c transmissions. When the transmission path is congested, the source node can stop the transmissions of lower priority #ows and let higher priority #ows keep their transmissions. With regulated transmission, higher priority #ows can have better transmission performance.
Depending on the importance and time constraint of a transmission, network administrators may assign a proper transmission priority to TCP/UDP #ow at the source nodes. A time-critical #ow can receive a higher transmission priority. The congestion-control mechanism collects the priority information of #ows. TCP and UDP #ows can concurrently transmit their packets. This mechanism will routinely check the congestion signal issued by TCP #ows to detect congestion on the transmission path. If the network is congested, it stops the transmissions of lower priority #ows to release some bandwidth share for higher-priority #ows to get a better transmission performance. Otherwise, if there is no congestion, perhaps due to more bandwidth available on the network, it starts the transmission of higher-priority #ows to enhance the bandwidth utilization. To prevent the transmission starvation of lower-priority #ows, this mechanism uses a priority ageing method to upgrade the priority of lower-priority #ows. After a transmission period elapses, lower-priority #ows can have the higher priority and allocate more bandwidth share to transmit packets.
1.2. Time-slot mechanism
The time-slot mechanism is an application of the time-sharing concept. The bandwidth is divided into many transmission units. Each transmission unit is a time slot. In each time slot, the source node only allows one #ow to transmit its packets and this #ow can use all available bandwidth or as much as it can. All other #ows must yield the right of way during the time slot. The number of time slots that a #ow can get depends on the priority, assigned by network administrators at a source node, to a #ow.
This strongly ensures that a high-priority #ow receives the required bandwidth. With the time-slot control mechanism, the transmission behaviours of TCP and UDP #ows will be regulated. UDP #ows can no longer occupy the bandwidth share irresponsibly. Moreover, the transmission performance of each #ow can be ensured with its priority. A round-robin scheduler is used by the time-slot mechanism to arrange each #ow's transmission. The time-slot mechanism
Figure 1. A topology of simulation scenario.
adopts the "rst-come-"rst-served principle to append a #ow to a round-robin scheduling queue and transmit its packets by turns. When a #ow takes turns at transmitting its tra$c, the time-slot mechanism assigns a round-robin transmission time to the #ow according to its priority. Then a transmission token is assigned to the #ow to start its transmission. With a time-slot mechanism, although each #ow can get an assigned period to send its packets, a transmission-starvation situation may happen to the lowest-priority #ow when higher-priority #ows continue to arrive. A priority-ageing method is also incorporated to such a delayed #ow.
2. A SIMULATION OF CONGESTION-CONTROL AND TIME-SLOT MECHANISMS Several scenarios are simulated to illustrate the operations of these two mechanisms. With the simulation results, one can obtain some transmission-performance statistics about these two mechanisms. Factors that may a!ect the algorithms were also investigated.
The topology of the simulation is shown in Figure 1. TCP/UDP tra$c #ows are simulated to transmit packets from the S1 and S2 nodes, all tra$c #ows have the same routing path and share the same bandwidth from N1 node to N4 node, then reach the D1 and D2 nodes. The ratio of Internet TCP/UDP tra$c #ow is basic to our simulation scenarios. From the MCI/NSF's very high performance Backbone Network Service (vBNS) project [4], one can "nd that the ratio of TCP and UDP tra$c #ows is 90 : 10. Based on this, with 100 tra$c #ows the TCP may vary from 81 to 99 whereas UDP varies from 19 to 1 during simulation. The transmission size is another factor that may impact the behaviour of tra$c #ows. We use a 10 kbyte "le as the smaller tra$c-#ow source and a 1 Mbyte "le as the larger tra$c-#ow source on the network.
For assigning priority, six bits in the TCP header are reserved to indicate the priority of the #ow [5,6]. A bit in a di!erent position represents a di!erent-level priority. The leftmost bit is the highest priority and the rightmost bit is the lowest priority. There are six levels of priorities available in both proposed mechanisms. For the congestion-control mechanism, priority is used to determine whether a #ow continues its transmission when the network is congested. If the network congests, the #ows of lower priority yield way to the #ows of higher priority. For the time-slot mechanism, priority is used to determine the number of time slots allocated for a #ow. Let P denote the priority of a #ow, where P"1, 2,2, 6. When P"1, the #ow is the highest priority, whereas when P"6 the #ow is of the lowest priority. Let tsn(P) denote the number of
Table I. A speci"cation of the "ve transmission environment settings.
Environment settings Purpose
1. cc(S, D) and cc(S, D): cc/cc Congestion-control mechanism 2. ts(S, D) and ts(S, D): ts/ts Time-slot mechanism
3. cc(S, D) and be(S, D): cc/be A congestion-control mechanism and a best-e!ort tra$c mechanism 4. ts(S, D) and be(S, D): ts/be A time-slot mechanism and a best-e!ort tra$c mechanism
5. be(S, D) and be(S, D): be/be Best-e!ort tra$c mechanism
research, four priority levels are assigned to the TCP #ows. Let TCPG denote a TCP #ow of priority i. Two priorities, 3 and 4, are assigned to the UDP #ows. Let UDPG denote a UDP #ow of priority i.
Queueing disciplines are also important since they may impact on the transmission perfor-mance of the two proposed control mechanisms. In our simulations, four di!erent queueing disciplines: "rst come "rst serve (FCFS), stochastic fair queue (SFQ) [7], random early detection (RED) [8] and de"cit round robin (DRR) [9] are implemented to schedule network applications' transmissions.
Various control mechanisms can be used in transmission over the Internet. The congestion-control mechanism and time-slot mechanism may coexist with other transmission-congestion-control mech-anisms and with the best-e!ort tra$c. Co-working with a di!erent control mechanism, the proposed mechanisms may have a di!erent transmission performance and behaviours. To simulate di!erent combinations, two groups of end-to-end tra$c transmissions with the indi-vidual #ow-control mechanism are investigated. Let cc(SG, DH) denote the congestion-control mechanism applied to the #ows from source SG to destination DH. Let ts(SG, DH) denote the time-slot mechanism applied to the #ows from source SG to destination DH. Let be(SG, DH) denote best-e!ort tra$c applied to the #ows from source SG to destination DH, &be' represents typical tra$c of sources not employing fairness techniques. The "ve di!erent transmission environments are illustrated in Table I.
3. RESULTS AND ANALYSIS
Parameters used for the simulation include transmission size, queueing disciplines, transmission performance, TCP/UDP ratios, environment setting and parameter settings.
3.1. Sensitivity to transmission size
The simulation results show that the congestion-control mechanism provides a signi"cant di!erential service among the TCP/UDP #ows at the transmission size of 1 Mbyte. Most of the 1 Mbyte TCP/UDP #ows receive di!erent transmission performance based on their transmission priorities. In cc/cc and cc/be transmission environments, when the transmission size is large, the priority of the #ow dictates the transmission performance. For 10 kbyte TCP/UDP #ows, the congestion-control mechanism also provides a di!erential service. But, there are cases where the transmission performance is inconsistent with their transmission priorities. Some tra$c #ows without higher priority showed better transmission performance.
Likewise, the di!erential service transmission behaviours only occur in the 1 Mbytes TCP/ UDP #ows for ts/ts and ts/be settings. When the transmission size is 10 kbytes, however, the performance of each TCP/UDP #ow behaves as a "rst-come-"rst-served transmission for both settings. Except for the "rst #ow, to start transmission immediately, all subsequent data #ows must wait for their turns even though there are available fragments within under-utilized time slots. 3.2. Diwerential service operations in the control mechanisms
Table II shows the summary of average-transmission performance of tra$c #ows from the S1 source node to the D1 destination node. These #ows are cc, ts or be. In each environment settings, four queueing disciplines (FCFS, DRR, RED and SFQ) are used. The transmission size of each #ow is 1 Mbyte.
Examining Table II, di!erent queueing disciplines do not show signi"cant di!erence in performance. For both ts and cc mechanisms, performance of a data #ow is only dictated by its transmission priority. Figure 2 shows the average transmission performance of TCP/UDP #ows with the FCFS queueing discipline in the di!erent transmission environments.
3.3. A relationship between the control mechanisms and queueing disciplines
Table II shows that the transmission performance is not too sensitive to the queueing disciplines. The queueing disciplines, however, do impact the transmission performance when di!erent priorities are imposed on the tra$c. Several observations occur from Table II:
(1) The RED queueing discipline has a better transmission performance for the TCP #ows in the cc/cc and cc/be environment. Moreover, the RED queueing discipline does not favour the TCP #ows in the ts/ts environment. On the contrary, RED favours the UDP #ows. (2) With DRR-queueing discipline, the lower-priority TCP/UDP #ows get worst transmission
performance in the cc/cc and cc/be environments.
(3) In the ts/ts and ts/be environments, the four di!erent queueing disciplines do not show much impact on the TCP/UDP #ows' transmission performance.
(4) In the di!erent transmission environments with the FCFS- and SFQ-queueing disciplines, the #uctuation of average transmission performance for TCP/UDP #ows is smaller than that of the RED and DRR queueing disciplines.
3.4. Transmission performance of the control mechanisms
Table II clearly shows that the congestion-control mechanism in each case outperforms the time-slot mechanism. The time-slot mechanism might su!er from underuse since the required
Figure 2. Di!erential service operations of the cc/cc, cc/be, ts/ts and ts/be environment.
burst bandwidth is smaller than the time slot assigned in the ts/ts environment. Moreover, most of the bandwidth in the ts/be environment may be taken by best-e!ort tra$c, leaving only a little bandwidth available for the time-slot mechanism to regulate the #ows' transmission. On the other hand, the congestion-control mechanism always keeps higher priority tra$c #ows to take the bandwidth whenever possible. This may be another reason why the congestion-control mecha-nism outperforms the time-slot mechamecha-nism in 1 Mbyte simulations.
Table II also shows that UDP has better performance than that of TCP, and UDP has better performance than that of TCP. Therefore, by assigning a lower priority to UDP #ows when they are not time critical, ensures that TCP #ows transmit before UDP #ows.
The transmission performance of these two control mechanisms in the cc/be and ts/be environments is interesting. From the 3rd and 4th row blocks in Table II, one can "nd that the congestion-control mechanism has a better transmission performance than that of the time-slot mechanism when they operate with best-e!ort tra$c #ows. In contrast to the cc/be TCP #ows being better than all the be/be TCP #ows, the transmission performance of TCP, TCP and TCP in the cc/be is worse than the corresponding TCP #ows in the be/be environment. Only the cc/be TCP #ows get a guaranteed service. In other words, there is no need to have too many priority levels in a di!erential-service mechanism. Two levels are enough. A high-priority #ow will receive a guaranteed service and has a better performance than #ows of lower priority. On the other hand, the low priority or the best-e!ort tra$c #ows will compete for bandwidth left by #ows of the highest priority.
Figure 3. Flows' transmission performance with di!erent ratios of TCP/UDP #ows.
Various ratios of TCP/UDP data #ows are simulated and the FCFS results are shown in Figure 3. The TCP/UDP ratios do not show signi"cant e!ects on the tra$c performance. According to the di!erent transmission environments, we draw six line charts to demonstrate the variations of transmission performance of TCP/UDP #ows as the TCP #ow number increases and the UDP number decreases. From the transmission performance line charts' #uctuations, there is no obvious evidence showing a relationship between a ratio of TCP/UDP #ows and their transmission performance.
3.5. Parameter settings of the control mechanisms
Three important key parameters are investigated: priority ageing, round-robin transmission time, and the number of time slots assigned to each priority. Proper parameter settings allow the control mechanisms to have better control. With numerous simulation settings, one can "nd that the priority-ageing time is important. The priority-ageing time impacts both mechanisms. Too short a priority-ageing time allows the lower-priority #ows to be upgraded sooner than other-wise. In that case, soon all the #ows become the highest priority. This tra$c pattern in turn degenerates into a best-e!ort tra$c and the di!erential service is not supported any more. Too long a priority-ageing time, however, may cause a #ow with lowest priority to starve because other higher-priority #ows may keep coming and jumping ahead of the queue.
The round-robin transmission time is also important in the time-slot mechanism. A #ow's round-robin transmission time depends on the number of time slots. A proper number of time slots bene"ts both TCP and UDP #ows. If the number of time slots is too large, a long round-robin transmission time will lead to a "rst-come-"rst-served operation. If the number of time slots is too small, the round-robin transmission time can be shorter than the round-trip time. TCP #ows cannot get their ACKs from the destination, the retransmission of TCP #ows will happen repeatedly and their performance will be poor. Additionally, if di!erences among round-robin transmission times of di!erent priority #ows are too large, this may starve a low priority #ow. Otherwise, di!erential service behaviours do not signi"cantly vary between priority #ows. For the time-slot mechanism, the number of time slots assigned to each priority is important. The binary-bandwidth allocation guarantees that high-priority #ows receive better performance than low-priority #ows. Meanwhile, this allocation scheme does not starve the low-priority #ows. The optimal di!erential bandwidth allocation of the round-robin transmission time deserves further study.
4. CONCLUSION
The congestion-control mechanism and time-slot mechanism are the two source-based #ow-control mechanisms studied in this research. These two mechanisms are applied at the source node to regulate the transmissions of TCP and UDP #ows. In these two mechanisms UDP #ows are regulated and are not irresponsible to the network congestion.
Because these control mechanisms regulate the TCP and UDP #ows at the source node, they are compatible with the current-transmission operation environment over the Internet. No additional device, protocol, or control mechanism is needed to implement these two mechanisms. The only operational cost of these two mechanisms is the execution time at the source node.
ACKNOWLEDGEMENTS This research is sponsored in part by NSC 89-2416-H-009-011.
REFERENCES
1. Bernet Y, Blake S, Grossman D, Smith A. An informal management model for Di!serv Routers. draft-ietf-di+serv-model-04.txt, July 2000.
2. URL http://www.ietf.org/html.charters/di!serv-charter.html
3. Blake S, Black D, Carlson M, Davies E, Weiss W. An architecture of di!erentiated services. RFC 2475, 1998. 4. URL http://www.vbns.net/stats/#ows/html/index.html
5. Tanenbaum AS. Computer Networks (3rd edn). Prentice-Hall Inc.: New Jersey, 1996; 413}416.
6. Wright GR, Stevens WR. ¹CP/IP Illustrated, vol. 22¹he Implementation. Addison-Wesley Publishing Company: Reading, MA, 1995; 211}212.
7. McKenney PE. Stochastic fairness queueing. Internetworking: Research and Experience 1991; 2:113}131.
8. Floyd S, Van Jacobson. Random early detection gateway for congestion avoidance. IEEE/ACM ¹ransactions on Networking 1993; 1(4):397}413. URL http://www-nrg.ee.lbl.gov/nrg-papers.html
9. Shreedhar M, Lambda GV. E$cient fair queueing using de"cit round robin. Proceedings of the Conference on Applications, ¹echnologies, Architectures, and Protocols for Computer Communication, September 1995; 231}242.
management.
Chen-Hua Fu received his BS degree in the Department of Information Management, National Defence Management College, Taiwan, R.O.C., and his MS degree from the Institute of Computer Science, U. S. Naval Postgraduate School in 1993. He is a PhD student in the Institute of Information Management, National Chiao Tung Univer-sity, Taiwan, R.O.C., where he is working towards his PhD on Network Communica-tion.
