Improving RLP Performance by Differential Treatment of Frames

Transcription

1 Improving RLP Performance by Differential Treatment of Frames Mainak Chatterjee Department of Electrical and Computer Engineering University of Central Florida Orlando, FL Samrat Ganguly Broadband and Mobile etworking EC Labs Princeton, J 84 samrat@nec-labs.com Jaideep Sarkar Department of Electrical and Computer Engineering University of Central Florida Orlando, FL jaideep@cs.ucf.edu Abstract In this paper, we propose enhancements to radio link protocols by identifying decisive frames and categorizing them as crucial and non-crucial. The fact that initial frames from the same upper layer segment can afford a few trials of retransmissions and the later frames cannot, motivates our work. We treat the frames differentially with respect to FEC coding and ARQ schemes. We consider specific cases of FEC and ARQ strategies and qualitatively show how the differential treatment of frames can improve the performance of the RLP. I. ITRODUCTIO To support end-to-end services to wireless and mobile hosts in 3G cellular systems, it is necessary that transport layer protocols such as TCP be supported over the wireless links so that the increasing demand for data services are met. Unfortunately, the design of TCP has been done in such a way that the performance degrades in a wireless environment where the channel error rates are high. Due to this lossy nature of the wireless channels, there are frequent packet losses which are misinterpreted by TCP as congestion related losses and it unnecessarily reduces its transmission window size resulting in reduced throughput. To counter this problem, radio link protocols (RLP [] have been proposed which shields the channel related losses from TCP. RLP is generally employed in the logical link control layer, between the physical layer and the TCP layer. The main function of the RLP is to conceal the losses from TCP by quickly recovering the dropped packets by means of local retransmissions. RLP fragments the segments received from TCP into equal sized frames and adds header to each frame before transmitting them over the physical channel. In case of a RLP frame loss during transmission, the RLP uses an Automatic Repeat request (ARQ mechanism to recover lost or damaged frames. The retransmissions by the RLP are considerably faster because the RLP expiration timer is usually much smaller than the TCP timeout and hence RLP is able to recover the frames before the TCP timer expires. RLP aborts the frame recovery process after the number of allowed retransmissions fail to recover a lost frame. In that case the RLP hands over the packet (with missing frames to the upper layer. Then TCP starts its own retransmission scheme to recover the lost frames. RLP uses a number of retransmission schemes [4], depending on the channel conditions and the performance required for the session it is supporting. The combination of forward error correction FEC and ARQ is known as hybrid ARQ. Oftentimes, hybrid ARQ s [6] are also used to enhance the performance of RLPs. Hybrid ARQs incorporate certain FEC schemes through which it ensures that there is a higher probability of the packets reaching the receiver end. If the two schemes are matched to the channel conditions then the hybrid ARQ can significantly change the system performance. FEC reduces the number of retransmissions by correcting the detectable and correctable errors. However, all errors cannot be corrected and the receiver sends for a retransmission request of the RLP frame rather than passing on the uncorrected frame to the upper layers to be corrected by TCP. Thus by properly combining FEC and ARQ, the overall system throughput and reliability can be increased. In this paper, we demonstrate how the relative position or the sequence number of the RLP frames plays an important role in the delay performance of the RLP. We show that the timely delivery of a fraction of the frames are more vital than others, and hence categorize them into crucial and non-crucial. The crucial frames are those that have greater impact on the delay performance of the RLP. We find this fraction and the sequence number at which the frames become crucial from non-crucial. We propose differential treatment of the crucial frames with respect to FEC coding and ARQ scheme. With specific case examples we qualitatively analyze the failure probability, delay and goodput as achieved by the RLP. Most of our analysis assumes parallel transmission of original and retransmit frames, as in HSDPA systems. With the proposed framework, depending on the application requirements, the desired levels of FEC and ARQ can be chosen so as to obtain a certain level of performance from the RLP. The rest of the paper is organized as follows. We present the basis of our framework in section II and show the existence of crucial frames. In section III we quantitatively demonstrate with the help of specific examples of FEC and ARQ how differential treatment of RLP frames can enhance the RLP. Conclusions are drawn in the last section. IEEE Communications Society /4/$. (c 4 IEEE

2 II. PROPOSED RADIO LIK PROTOCOL The main drawback with current RLPs is that they do not differentiate the frames obtained from the same TCP segment. This motivates us to provide differential treatment to the RLP frames. We believe that the importance of each RLP frame from the same TCP segment is different and hence deserves differential treatment. Our claim is based on the fact that the reassembly of the RLP frames can only be done when all the frames belonging to the same TCP segment are correctly received by the receiver. The basic philosophy is that the last frame of a particular segment will decide the time of delivery of the reassembled TCP segment to the upper layer. The last frame here refers to the last frame received correctly (among the frames from the same TCP segment. ote that the last frame need not be the last one in terms of the sequence number. But, under ideal channel conditions (i.e., no frame loss the last frame in terms of the sequence number will arrive last simply because of the sequential nature of the transmission as shown in Figure. Suppose a TCP segment under consideration is fragmented into L frames. It can be seen from the figure how the L RLP frames obtained from a TCP segment are transmitted and received, the last (Lth frame arriving at time t ideal. At this time, all the L frames are ready to be reassembled to form the TCP segment. But in a realistic situation, frames will get dropped and due to retransmissions of the erroneous frames, the i th RLP frame might successfully arrive last, where i L. If i happens to be one of the initial frames of the TCP segment then few trials of retransmission are possible because the retransmissions would be complete before the last RLP frame successfully arrives at the receiver. But retransmissions of the later frames might delay the delivery of all the reassembled TCP segment to the upper layer. It is these later frames which are more important in deciding the total delay for the reassembly of all the RLP frames to form the TCP segment. Hence we call these frames as crucial, and the others as non-crucial. ext, we find the fraction of total frames that are crucial. 3 4 i L L r RLP frames obtained from a TCP Segment RTT = c r Fig.. A. Finding Crucial Frames 3 4 i L L Error-free transmission of RLP frames Let us assume that the size of a TCP segment be T bytes and it is fragmented into equal sized RLP frames. The number of RLP frames obtained from a TCP segment would be L = T R p, where R p is the payload of each RLP frame. The actual size of the RLP frame would be R p plus some header information and some redundancy checks. As shown in Figure, the time at which all the L frames are received at the receiver under ideal channel conditions (i.e., t i t ideal no frame loss is denoted by t ideal. Considering successive and pipelined transmission as in HSDPA we obtain t ideal = Lr T rtt /, where T rtt is the round trip time of each RLP frame and r is the transmission time of each RLP frame. But when frames get dropped or corrupted with a probability p, the correct reception of frames will get delayed due to retransmissions. If we do not restrict the number of allowed retransmissions then the expected delay, D, of a lost frame would be given by D =( p Trtt p( p 3Trtt p ( p Trtt. However in real implementations, the number of allowed retransmissions is usually three and so the correct expression for D will result in a truncated (finite series. As p is small, we can ignore the higher powers of p and approximate the finite series with the above infinite series and upon simplification, we obtain D = T rtt( p ( p. ( If this approximation leads to a high error margin, then we can as well take the first four terms (one transmission plus three retransmissions in calculating D. We now calculate the expected time of arrival of each frame at the receiver end. Let us denote the expected time of arrival of the ith frame as t e (i. It can easily be shown that t e (i = D (i r, where r is the transmission time of each RLP frame. We can thus find the frames whose expected time of arrival at the receiver will be more than t ideal. Equating t ideal and t e (i, and substituting for D and T rtt = cr, we get i = L c/ c( p. ( ( p Frame number i gives the starting index for the crucial frames. Of course, the fraction of crucial frames will vary under different conditions. The fraction of crucial frames is simply defined as the ratio of the number of crucial frames to the total number of frames, i.e., L (i L. III. DIFFERETIAL TREATMET OF FRAMES So far we have demonstrated that there is a fraction of frames emanating from the same TCP segment which are more crucial than others, in determining the performance of the RLP. ow, we would like to treat these crucial frames differently compared to the non-crucial ones. Currently, all the existing RLPs have the same channel coding and ARQ mechanism for all the frames. In this paper, we do not propose any new channel coding scheme or ARQ mechanism but show comparatively how the performance of the RLP could be improved if we were to use different channel coding schemes and ARQ mechanisms for different frames. We propose differential FEC and ARQ treatment for the RLP frames. Both the FEC and the ARQ schemes to be applied to a frame will depend on the index of that frame. In obtaining equation ( we assumed that the number of retransmissions allowed was infinite, and therefore all packets were eventually recovered. However, in reality this is not true. In most cases, the number of retransmissions IEEE Communications Society /4/$. (c 4 IEEE

3 allowed is three, but the manner (number of duplicate frames transmitted in each trial in which these three trials are done is different. Thus, due to the finite number of retransmissions trials there would be a non-zero probability of a frame being not recovered by the RLP. Let us formally define RLP failure probability along with the two other metrics of interest- delay and goodput. Definition of RLP failure probability: We define RLP failure probability as the probability of the RLP failing to deliver all the L frames within its allowed number of retransmissions as a result of which the recovery mechanism will be handed over to the upper layer (e.g., TCP. Definition of Delay: We define the delay for a frame as the time taken for that frame to be received correctly at the receiver with respect to the first frame s transmission time i.e, the time the RLP started transmitting the first frame from a particular TCP segment. Definition of Goodput: We define goodput as the ratio of the actual number of information bits decoded correctly at the receiver to the total number of bits transmitted. It can be noted that for all the analysis that follows, a certain FEC and ARQ scheme was assumed. We just show how the differential treatment of the frames affects the performance of the RLP. Also, we would repeat notations as their meaning remain the same. F and F would represent the RLP failure probability, D and D the delay, and G and G the goodput. The indices and are for the two cases which we compare. A. Differential FEC When a receiver gets a corrupted RLP frame, it is in no position to correct the errors. However, if some redundant bits in the form of FEC codes are added to the RLP payload before transmission, then there is a probability that the receiver would be able to detect and possibly correct the errors. The correction capability of these codes will depend on the kind of codes and the length of the code used. Since this paper does not deal with FEC codes, we will just discuss in term of the simplest of codes- Block codes. In block codes, M redundancy bits are added to the information bearing bits. (ote that these extra M bits are generated using a generator matrix operating on the bits. If we consider a RLP frame of M bits, then the resulting bit loss probability is given by [] b = M i=m ( M ( b pl M i b i i pl i M (3 where b pl is the bit loss probability before decoding. Of course, different FEC schemes will yield different loss probabilities. From this equality (or any such relation between M and b we can calculate the number of redundant bits to be added to achieve a desired loss probability. The FEC coding need not be very robust for these non-crucial frames. This will also reduce the overhead since the redundancy bits will be less. On the other hand the later packets are more crucial and retransmission should be avoided or minimized. One way to avoid or minimize loss of the crucial frames is to use stronger FEC codes. Of course, there are some tradeoffs which we will discuss later. Case : Traditional 3 4 k L L FER = p Case : Proposed (an example with M redundancy bits 3 4 k L L FER = p with M redundancy bits FER = p ( M Fig.. Different FEC schemes Let us assume that a traditional RLP uses M bits to code each frame as shown in Case of Figure. It can be noted that each of the L frames is coded with the same number of bits, i.e., M, because of which the FER observed is p. ow let us consider that the RLP is made aware of the crucial frames and it codes each of the crucial frames using M bits, where M >M, as shown in Case. This usage of more redundancy bits will result in FER = p, where p <p. The exact reduction in the FER will depend on the values of M, M,, and the kind of coding used. We also assume that the ARQ scheme used is (,,. RLP Failure Probability: We need to calculate the probability that all the L frames will not be correctly received at the receiver. For Case, where the FER is p, the RLP failure probability (F is simply given by F = ( p 3 L. For the example in Case, where the first k frames experience aferofp and the last L k frames experience a FER of p (, the RLP failure probability (F is given by F = ( p 3 k ( p 3 L k. Due to stronger FEC in Case, the RLP is able to recover more frames than in Case. Delay Calculation: Recall that the expected delay of the ith frame is D (i r, where D = Trtt ( p. For Case, the delay at the RLP is D = Trtt ( p (L r. For Case, we will have to treat the crucial and non-crucial frames separately, because the frames would experience different loss rates. Irrespective of the losses, successive frames are always transmitted, i.e., crucial frames would not be prevented from transmission even if all the non-crucial frames are not received correctly. The last non-crucial (kth frame is expected to arrive correctly after a delay of D nc = Trtt ( p (k r. Similarly, the last crucial (L frame is expected to arrive correctly after a delay of D c = kr Trtt ( p (L k r. The term kr is the time it takes to transmit the noncrucial frames (not necessarily correctly before the crucial frames are transmitted. It is not known which frames would arrive later because the exact relation between p and p is not known. Therefore for Case, the delay at the RLP is D = max(d nc,d c. Although, the error performance of the transmission is improved by adding the redundancy bits, the goodput is compromised as we see in the next section. 3 Goodput Calculation: It is to be noted that the M goodput is ( p when the frame reaches in its first transmission. However, the goodput obtained after jth retransmission trail, j 3, will depend on IEEE Communications Society /4/$. (c 4 IEEE

4 RLP failure Probability (F Delay (D Goodput (G umber of redundancy bits (M 3 umber of redundancy bits (M D D F F G G course, the improvement saturates beyond M = which confirms that the error correcting capabilities of FECs are bounded. Also, from the goodput perspective it is advisable that the redundancy is not made arbitrarily large which we see from Figure 3. There is slight reduction in the goodput with the increase in redundancy. This is the trade-off for better RLP failure probability and delay performance. B. Differential ARQ In the differential FEC case (Section III-A we considered that the underlying ARQ scheme was (,, but with varying number of redundancy bits. ow, we consider two different ARQ schemes (,, for Case and (,, 3 for Case as shown in Figure 4. We also assume that the FEC codes used is uniform across all the frames (say M bits per frame, therefore all frames would experience a FER of p (say. Of course, these can be generalized with the only condition that the crucial frames in Case must have a stronger ARQ than the non-crucial ones. Case : Traditional 3 4 k L L ARQ (,, Case : Proposed (an example 3 4 k L L umber of redundancy bits (M Fig. 3. Performance due to Differential FEC the probability of previous failures and the total number of frames transmitted to eventually recover that frame. Therefore, for the jth retransmission, the goodput will be j pj ( p M. Thus, the goodput due to the original transmission and three retransmissions in Case would be G = j=3 j= j pj ( p M. Similarly, the goodput for CaseisG = ( k L G j=3 j= j pj ( p M. As expected, it can easily be verified that there is a reduction in goodput in Case,i.e., G <G. umerical Results for Differential FEC: In this section let us briefly discuss the performance of the RLP with respect to the three metrics when the proposed differential FEC is applied to the crucial frames. We assume L =3, T rtt = and p=% in calculating the results. It is also assumed that the number of information bits per frame ( is and the number of redundancy bits M is varied from to. M was maintained at, implying that no FEC was applied to the non-crucial frames. We observe from Figure 3 how the delay performance of the RLP improves because of adding more robustness to the crucial frames. We also observe that there is a variation in delay with the increase in redundancy bits of the FEC scheme used. Last, there is an improvement in the RLP failure probability with the increase in the redundancy bits. This ensures that the RLP is more effective in recovering the lost frames and thereby preventing the information of losses propagating to TCP. Of ARQ (,, Fig. 4. Different ARQ schemes ARQ (,,3 RLP Failure Probability: The RLP failure probability (F for the example in Case is obtained as F = ( p 3 L. For the example in Case, where the probability of correctly receiving a non-crucial frame is ( p 3 and a crucial frame ( is ( p 6, the RLP failure probability (F is F = ( p 3 k ( p 6 L k. As expected, we observe that F <F. Delay Calculation: We would only calculate the delay for the frames which are ultimately decoded correctly at the receiver. The frames which do not, are accounted for in the RLP failure probability. The expected delay for each noncrucial frame will be due to the delay contributions from the original transmissions and the three trials of retransmission. For Case, where the ARQ scheme is (,,, the expected delay for any frame is the expected delay due to the initial transmission and possibly the three retransmissions. The total expected delay (D in getting all the L frames in Case, will be dictated by the last frame to arrive at the receiver. If we calculate the expected time of arrival for all the L frames then we find that the expected delay for the Lth frame equals the delay for the first frame plus the delay in transmitting the Lth frame. Thus, we get D =( p T rtt p( p 3T rtt p ( p T rtt p 3 ( p 7T rtt (L r. For Case, we obtain the expected delay for the non-crucial frames ( through k in the same manner as Case. The expected delay (D nc for the non-crucial frames is IEEE Communications Society /4/$. (c 4 IEEE

5 D nc =( p Trtt p( p 3Trtt p ( p Trtt p 3 ( p 7Trtt (k r. The calculation for the crucial frames (k through L will be a little different because of the ARQ scheme. Therefore, expected delay (D c from crucial frames would be D c =( p Trtt p( p 3Trtt p ( p Trtt p 4 ( p 3 7Trtt (L r. The term (L r appears because we are calculating the delays with respect to the time of transmission of the first frame. Since, we do not know the values of the variables used, we cannot determine for sure whether the non-crucial or the crucial frames arrive later. Physically it means, the arrival of the frames would depend on the observed FER and also on the ARQ used. Therefore, the overall delay for Case is simply determined by finding the greater of D nc and D c as the time of reassembly will depend on the last frame that arrive at the receiver. Therefore, D = max(d nc,d c. 3 Goodput Calculation: Following the logic from the earlier goodput calculation, the goodput for Case would be RLP failure Probability (F Delay (D Frame Error Rate Frame Error Rate.9 D D F F G = j=3 j= j pj ( p M. For Case, we consider (,, and (,,3 schemes for the non-crucial and crucial frames respectively. We observe that the goodput for the first k frames would be the same as in Case. The goodput for the crucial frames will again depend on the probability of previous failures and the total number of frames transmitted to eventually recover that frame. For the original transmission the goodput will be ( p M.For the three retransmissions the goodput will be p( p M, 4 p ( p M and 7 p3 ( p 3 M respectively. So, the total goodput for all the L framesincasewouldbe G = k L G L k ( ( p L j=3 M j= p j ( p j. j ote that this expression for G is for the ARQ scheme considered, i.e., (,,3. The expression can be made general for any ARQ. For ease of demonstration, the specific scheme (,,3 has been worked out. umerical Results for differential ARQ: In this section we discuss the performance of the RLP with respect to the three metrics. Just to focus on the ARQ performance, we assume that there are no redundancy bits added to any RLP frame, so M =. We varied the frame error rate p from to.3. From Figure, we observe how the RLP failure probability is lowered when differential ARQ is applied. We do not see appreciable gain in the delay with the better ARQ. This is due to the fact that the ARQ scheme (,,3 successfully recovers more frames in a given time than the (,, scheme which results in an additional delay. It is clear that from the second retransmission onwards the (,,3 scheme starts recovering more frames than the (,, scheme. Last, we see that there is hardly any degradation in the goodput for scheme (,,3. This is because, the loss in goodput due ormalized Goodput (G Frame Error Rate Fig.. Performance due to Differential ARQ to the transmission of duplicate frames in (,,3 scheme is compensated by the recovery of more frames. C. Differential FECARQ In sections III-A and III-B, we have shown how differential RLP would perform if only FEC or ARQ was applied. In this section we apply both differential FEC and ARQ. Similar to the previous sections, we consider cases as shown in figure 6. In Case, the frames are subjected to an ARQ scheme of (,, and each frame is coded with M bits. In Case, the non-crucial frames are treated as the frames in Case but the crucial frames are subjected to an ARQ scheme of (,,3 with M redundancy bits per frame. Case : Traditional 3 4 k L L FER = p (M, ARQ (,, Case : Proposed (an example 3 4 k L L FER = p (M, ARQ (,, Fig. 6. FER = p (M, ARQ (,,3 Different (FECARQ schemes RLP Failure Probability: The RLP failure probability for Case is obtained as F = ( p 3 L. Similarly we ( obtain the RLP failure probability in Case as F = ( p 3 k ( p 6 L k. G G IEEE Communications Society /4/$. (c 4 IEEE

6 Delay Calculation: The delay in Case will just be the same as that of D in the differential ARQ case but p would be replaced by p in the expression for D. Similarly for Case, the delay for the non-crucial frames, D nc, would be the same as the equation for delay for the non-crucial frames in the ARQ case with p replaced by p. Hence the delay for the non-crucial frames would be D nc =( p Trtt p ( p 3Trtt p ( p Trtt p 3 ( p 7Trtt (k r. The delay for the crucial frames (k through L will now have the combined effect of both FEC and ARQ. The expected delay (D c from the crucial frames would be D c =( p Trtt p ( p 3Trtt p ( p Trtt p 4 ( p 3 7Trtt (L r. The term (L r appears as we are calculating delay with respect to the first frame. So, the overall delay for Case is again obtained by the greater of the two D nc and D c. Thus, D = max(d nc,d c. 3 Goodput Calculation: The goodput due to the original transmission and three retransmissions in Case would be G = j=3 j= j pj ( p. M Similarly for Case, where the non-crucial frames are subject to an ARQ scheme of (,, and redundancy bits of M and the crucial frames with an ARQ scheme of (,,3 with redundancy bits M the goodput is given as G = k L G L k ( ( p L RLP failure probabilty (F Fig M RLP failure Probabilty (F j=3 j= RLP failure probability vs. umber of redundancy bits and FER p j j ( pj. frame error rate is high. ormalized Goodput (G Fig. 9.. ormalized Goodput (G Goodput vs. umber of redundancy bits and FER umerical Results for ARQFEC: To see the effect of differential FECARQ, we maintain the same range for p and M as in our previous results. We assume M =. We plot the RLP failure probabilities in Figure 7. We observe how the RLP failure probability is lowered in Case (right side plot. We observe from Figure 8 how the increase in the redundancy bits improves the delay performance in Case. The plots for goodput are shown in Figure 9. Interestingly we find that initially the goodput decreases due to the increasing redundancy bits and also more number of retransmissions due to (,,3 scheme. However, when the frame error rate is high the goodput increases because with the increasing redundancy bits and the (,,3 scheme will actually recover more packets which then contributes towards increasing the goodput. Or in other words, the differential RLP is more effective when the IV. COCLUSIOS This paper demonstrates how the performance of radio link protocols can be improved if the RLP frames are treated differentially. We categorize the frames into crucial and noncrucial and find their ratio as a function of segment size, roundtrip time and frame error rate. We applied differential FEC and ARQ based on their relative position of the frames. We considered parallel transmission (as in HSDPA of original and retransmit frames, and specific FEC and ARQ schemes to show the qualitative gain. Similar quantitative analysis can be done for generalized FEC and ARQ. Our results signify that if the performance of the differential RLP is known for various FEC and ARQ schemes under different channel conditions, then the RLP can choose the appropriate hybrid mechanism which will sustain the promised level of reliability expected from the applications to be supported REFERECES Delay (D Fig. 8. Delay (D Delay vs. umber of redundancy bits and FER [] 3G Partnership Project, Release 99. [] G. Bao, Performance Evaluation of TCP/RLP Protocol Stack over CDMA Wireless Link, Wireless etworks, o., 996, pp [3] M.C. Chan, R. Ramjee, TCP/IP Performance over 3G Wireless Links with Rate and Delay Variation, ACM MobiCom,, pp [4] A. Chockalingam and G. Bao, Performance of TCP/RLP protocol stack on correlated fading DS-CDMA wireless links, IEEE Transactions on Vehicular Technology, Volume 49, Issue, Jan, pp [] Henrik Lundqvist and Gunnar Karlsson, TCP with End-to-End Forward Error Correction, Radio Vetenskap och Kommunication, RVK,Stockholm, June. [6] E. Malkamaki, D. Mathew,and S. Hamalainen, Performance of hybrid ARQ techniques for WCDMA high data rates IEEE VTC Spring, Volume: 4, pp IEEE Communications Society /4/$. (c 4 IEEE