Integration of Look-Ahead Multicast and Unicast Scheduling for Input-Queued Cell Switches

Transcription

1 22 3th nternational Conference on High Performance Switching and Routing ntegration of Look-Ahead Multicast and Unicast Scheduling for nput-queued Cell Switches Hao Yu, Member,, Sarah Ruepp, Member,, Michael S. Berger, Member,, and Lars Dittmann, Member, Abstract-This paper presents an integration of multicast and unicast traffic scheduling algorithms for input-queued cell switches. The multi-level round-robin multicast scheduling (ML RRMS) algorithm with the look-ahead (LA) mechanism provides a highly scalable architecture and is able to reduce the head-ofline (HOL) blocking problem that the weight-based algorithm (WBA) suffers from. Leveraging the FLter & Merge (FLM) scheme, multicast and unicast traffic are independently scheduled based on their requests. Decisions are integrated following a specific policy. Remainder is looped back to the filtering module that filters out the conflicting requests to ensure fairness. Simulation results show that comparing with the scheme using WBA for the multicast scheduling, the scheme proposed in this paper reduces the HOL blocking problem for multicast traffic and provides a significant improvement in terms of latency. ndex Terms-Multicast, unicast, integrated scheduling, inputqueued switch.. NTRODUCTON VN though the vast majority of network traffic today consists of unicast connections, multicast will become widely used as video applications turn popular in the nextgeneration communication networks, where the bandwidth of the access network is no longer the bottleneck. Thus, integrating multicast and unicast scheduling into a single switching system becomes necessary because it provides network operators with simplicity and flexibility in terms of network infrastructure and management. n terms of buffer allocation, switch architectures are usually categorized into several queuing mechanisms. The nput Queuing (Q) mechanism is accepted for large-scale switch designs due to its high scalability but can suffer from the head-ofline (HOL) blocking problem []. The Output Queuing (OQ) avoids the HOL blocking by increasing the output buffer speed by a factor of N, the number of input/output ports. This results in a poor scalability that limits the applications of the OQ mechanism. Similar to the OQ mechanism, the shared-buffer mechanism can also avoid the HOL blocking by speeding up the buffer speed, which makes it hard to scale up the switch size. By maintaining a separate queue for packets bound for the same the output port, Virtual Output Queuing (VOQ) eliminates the HOL blocking for unicast traffic at the cost of the increase of number of input buffers. Based on the VOQ architecture, several unicast scheduling algorithms have been This work was supported in part by the Danish Advanced Technology Foundation in the project The Road to Gigabit thernet. H. Yu. S. Ruepp, M. S. Berger, and L. Dittmann are with the Department of Photonics ngineering, Technical University of Denmark, Kgs. Lyngby, Denmark; {haoyu.srru.msbe.ladit}@fotonik.dtu.dk proposed to provide efficient scheduling performance, such as PM [2] and islp [3]. When the VOQ architecture to multicast traffic is applied to avoid multicast HOL blocking, a separate queue must be maintained for each possible destination combination, resulting in 2 N - queues for multicast per input port. The exponential increase of the number of multicast queues per input as the switch size grows strongly limits the integration of unicast and multicast. To tackle this problem, several schemes using reduced number of multicast queues are proposed [4], [5]. TATRA [4] provides a good scheduling performance but the complexity is too high for practical implementations. WBA [4] is proposed as a replacement to TATRA due to its low implementation complexity but it can suffer from the HOL blocking due to the algorithm only operates on the HOL cells in the queues. Using k ( < k < 2 N - ) multiple queues to schedule multicast traffic, the k-mc-voq [5] can alleviate the HOL blocking. However, a cell dispatching scheme, determining which queue the incoming cell should be forwarded to, needs to be carefully designed. f the dispatching scheme is based on the multicast fan-out information, each arriving cell has to be speculated before entering the queues, which can limit the throughput. The FFOMS [6] and the CMF [7] utilize the VOQ architecture to schedule multicast traffic. nstead of assigning a separate queue to each combination of N destinations, only N queues are allocated for each input port. Up to N address tokens are generated for each arriving cell, each of which is stored in the a queue corresponding to a destination. The arrived multicast cell is stored in a memory pool and is linked by its address tokens. Based on the scheduling decisions from the scheduling algorithms executed on the address tokens, the multicast cell is sent and is removed from the memory until all its destinations are reached. The FFOMS and CMF are able to achieve low latency and high throughput, but the bottlenecks of the architecture can hinder its scalability. However, the hardware complexity of the address token generator can be O(N), since up to N tokens are generated for each arriving cell, and the address token generating rate is required to be N times the cell arrival rate due to that multiple tokens are generated for each arriving cell within one cell transmission time. Besides, this architecture requires a complex buffer management mechanism to send a multicast cell using the link address in an address token because the actual cell to be sent is not always the HOL cell. n addition, the number of token queues in total is N2, which can be a obstacle for the /2/$

2 switch to scale up to hundreds or even thousands of ports. To design an integrated scheduling algorithm for both unicast and multicast in a high-speed data rate environment, a low-complexity architecture with high-performance scheduling integration is required. An input-queued integrated switch using N FFO queues for unicast and k FFO queues for multicast traffic ("N + k") is proposed in [8] and % throughput is achieved. However, the design is based on a load-balance policy [9], where each incoming multicast cell must be investigated by the switch to acquire the destination vector, which can bottleneck the scalability. Using a single queue to store the multicast cells and the VOQ architecture to handle the unicast is thus considered potentially practical for a unicast and multicast integrated scheduling system. n [] and [], the "N + " architecture is applied, where only one FFO queue is installed for multicast and N queues for unicast. islp is used for the unicast scheduler and WBA for multicast. However, the HOL blocking problem of WBA is not addressed, resulting in a reduced performance for the multicast scheduling. The multi-level round-robin multicast scheduling (MLRRMS) [2], [3] uses one FFO queue to store the incoming multicast cells, and to reduce the potential HOL blocking, the look-ahead (LA) algorithm is used to search for any potentially blocked multicast cells to send. Analysis and simulations show that by enabling the scheduling system to be capable of looking one cell stored further in the queues, the maximum performance gain is reached compared to the added complexity. Thus, this paper investigates into an "N + " unicast and multicast integration scheme, where MLRRMS is used in the architecture for multicast and islp for unicast. The remaining parts of this paper is constructed as follows. Section presents the reference architecture of the integrated scheduling system. Section describes the integrated scheduling scheme in details. Section V analyzes the simulation results under different scenarios. At last, Section V concludes the paper.. NTGRATD S YSTM ARCHTCTUR The system architecture is shown in Fig.. The integrated scheduler contains unicast and multicast schedulers, both running independently. The switch is assumed to have equal number of input and output ports due to the fact that an input and an output port usually reside in pair on the same line card. Fixed-size packets, cells, are used in the system due to the fact that fixed-size switching technology is able to achieve high switching efficiency. Therefore time is divided into cell times, which is the transmission duration of a cell. ach input has N + first-in-first-out (FFO) queues, where N queues are for unicast traffic in a VOQ architecture and for the multicast traffic. A queue is denoted as,j, :::; i :::; N -, :::; j :::; N.,j, :::; j :::; N -, Vi, are used in the unicast VOQs and store unicast cells bound for output j. Q i,n is the queue that stores multicast cells. A cell position pointer, p, is used to indicate the cell position in the multicast queue, where the HOL cell has p O. n one cell time, an input can only send at most one cell to the switch core and an output can only receive at most one cell from the switch core. f more than Fig.. r l r l Unicast scheduler Multicast scheduler -. ' i : ntegrated scheduler, LJ, i-o a i-] a > u." : j-n -,..J!l'l Multicast FFO r , j-o j- j-2 j-n-] r l : j-n :.j.qgl Multicast FFO r System architecture of the uni-multicast integrated switch. one cells are bound for an output within a cell time, an output contention is occurred and only one cell can be scheduled for transmission according to the scheduling algorithm with other cells left to be scheduled in the next cell time. t is also assumed that the switch core has a intrinsic multicastlbroadcast capability, e.g. a crossbar fabric. A unicast cell arriving at input i and bound for output j is inserted into,j. A multicast cell arriving input i is stored in,n and is characterized by its fan-out set, i.e. the set of the output ports for which the cell is bound. The fan-out set is represented by an-bit fan-out vector, denoted as bi,j, where i and j are the input and output index, respectively. bi,j indicates that the multicast cell arriving at input i should be sent to outputj, else bi,j o. Fan-out splitting [4] is considered and applied so that copies of multicast cells can be delivered to output ports over any number of cell times. Unless all the destinations in the fan-out set are reached, the cell will not be removed but remain in the queue. A multicast scheduler makes scheduling decisions prior to each cell time and grants cell transmissions accordingly.. NTGRATON OF LOOK-AHAD MULTCAST AND UNCA ST SCHDULNG A. ntegrated scheduling framework Before each cell time, multicast and unicast requests are transmitted to the scheduling system in parallel as shown in Fig. 2. A request filter module filters out some of the 6

3 Look-ahead notification decisions Me-FFO i Multicast requests Fig. 2. VOQ Unicast requests Multicast scheduer Multicast t Request Filter Unicast scheduler decisions Remainder Unicast i ntegration ntegrated decisions llustration of unicast and multicast scheduling integration. requests based on the feedback information from the integration module. The filtered requests are passed to the unicast and multicast schedulers that run independently in a parallel fashion. Scheduling decisions from both schedulers are read by the integration module and combined together to form the integrated decision. Due to the independency, scheduling decisions of unicast and multicast may overlap on some output ports, which cause output contention. The integration module solves this problem by leaving a remainder to the request filter each scheduling round so that the filter only passes the requests that are from the unreserved inputs in the remainder and will not compete the output ports with the remainder. The parallel scheduling characteristic of such a framework allows the system to provide different scheduling features based on different filtering and integration policies. Different from the framework presented in [], the look-ahead mechanism [2] is added as part of the scheduling process. Only when idle outputs are found in the scheduling decisions, will the multicast scheduler ask for the fan-out information of the cells stored further in the multicast FFO (Me-FFO) queue, if the maximum look-ahead depth is not reached. B. ntegration policy As the scheduling decisions made by the multicast and the unicast scheduler are independent, overlapping on the reservation of output ports are very likely to happen. A multicast cell is usually bound for several output ports, meaning that it generate more traffic on the outputs than what is received on the inputs. Since fan-out splitting is applied, a multicast cell will remain the FFO queue until all it is served to all its destinations. This makes multicast traffic more sensitive to scheduling decisions than the unicast traffic stored in the VOQs. t is therefore reasonable to choose an integration policy that gives strict priority to multicast over unicast []. f the mutlicastlunicast scheduling decisions and the remainder of time slot n are described in N x N matrices Mn, un and Rn respectively, where Mij,Uij {,}, < i,j < N, the integration policy can be formulated as follows: f l'vr;,y cj,. f U;,j cj, R,j, U;,j, 'Vj 2. And if Uiy cj, R'i,y, Ury, 'Vi where l'vlrj indicates the multicast cell in input queue i should be copied and sent to output j, and Urj means the unicast cell in VOQ j of input i should be sent to output j. An o Multicast decision / MultlCast dccslon ntegration / Q Unicast decision / Q Umcast dccslon Rcmamdcr Fig. 3. xample of integration of multicast and unicast scheduling decisions. example where N 6 is of the integration process is shown in Fig. 3. The multicast decision is fully served because it has the higher priority than the unicast. Since column 3 and 4 of the multicast decision matrix are zero, the unicast decision matrix has an opportunity to have column 3 and 4 integrated into the final decision matrix. The unserved columns of the unicast decision matrix, i.e. column,, and 5, form the remainder of the integration round. C. Filtering policy As shown in Fig. 2, the remainder generated by the integration module is sent back to the request filter module used as a mask to filter out the requests that are in conflict with the remainder. The remainder should be served by the scheduler with a priority over the scheduling decisions of the current time slot, meaning that either multicast or unicast scheduling decisions in conflict with the remainder should be filtered out. By conflicting, it means that the new decisions contain either the inputs or the outputs included in the remainder. Using the notations described previously in Section -B, the filtering policy can be formulated as follows: f R, l cj,. M;,j, 'Vj and Mry, 'Vi, 'Vi 2. U;,j, 'Vj and Ury D. ntegrated scheduling scheme Scheduling algorithms are run independently on the multicast and the unicast schedulers. Multi-level round-robin multicast scheduling (MLRRMS) [3] first schedules on the HOL cells in the multicast FFO queues. f any idle output is found in the decisions, the fan-out requests of the cells behind filtered by the request filter according to the filtering policy are taken into scheduling process in the next iteration. This process continues until either the maximum look-ahead depth is reach or all cells in the queues are examined, or no idle outputs are found in the decisions. To allow the scheduling system to look ahead into the multicast queue without limit can eliminate the multicast HOL blocking but this is impractical to implement. 6

4 To control the complexity, the maximum look-ahead depth that is the maximum number of multicast cells examined by the scheduler further in the queues is defined. Results in [2] show that for uniform fan-out distribution by allowing the scheduler to look cell stored further in the queues, the switch can achieve a maximum improvement-to-complexity ratio. ither islp [3] or PM [2], which uses the VOQ architecture, needs multiple iterations to gain performance improvement. Simulation results show that the number of iterations needed for islp to converge is approximately log2 N, similar to PM [3]. Since the multicast scheduling decisions have a higher priority than the unicast scheduling decisions in the integration phase, the whole unicast scheduling decisions can be left as the remainder in an extreme case. Then in the next time slot ' the remainder is guaranteed to be served, achieving fairness. V. SMULATON PRFORMANC The proposed integration scheme with islp used for unicast scheduling and MLRRMS [3] for multicast scheduling algorithm is evaluated and compared with the integration scheme using islp and WBA [] by simulations carried out in OPNT modeler [5]. The simulated system is assumed to be an 32 x 32 input-queued switch with infinite buffers. The unicast scheduler uses islp with 5 iterations so that unicast scheduling The duration of each simulation is million cell times. The Bernoulli process and the bursty process, or Correlated Arrival Process, are used to generate cells for each input, respectively. For the Bernoulli process, each input receives a cell with a probability of p, which is equal to the arrival rate. For the bursty process, there are two states, busy and idle. Cells are generated only in the busy state. The process stays in each state for a random number of cell times following the geometric distribution with mean values of [B] and [], respectively. The arrival rate is calculated as p [B]/([B] + []). ach cell has a probability a of being a multicast cell. The fan-out of the multicast set, i.e. the number of destinations that the cell is bound for, is uniformly distributed between 2 to N, and the average fan-out thus becomes P 2 N. A unicast cell is uniformly bound for each output. The output load, A, that both traffic generate on the switch is thus: A p [ap + ( - a) l, (::; A ::; ) () and we define the fraction of the multicast load in the total output load to be: (3 pap pap + p( - a) Combining C!:... and q. 2, we have the output load A as a function of p, P, and (3: AP [ - ( - p-l)(3 (2) (::; A ::; ) (3) When (3, only unicast is contributing to the output load and A p. When (3, only multicast traffic exists and the output load becomes A pp. a; T ;:::r:: ;-r- --'-r- -rzjm o a; c..> Fig. 4. Average delay vs. output load (Al for the proposed integration scheme, f3.5 Fig. 4 compares the average delay between the unicast and the multicast traffic under two traffic models when the proposed integration is applied. The fraction of the multicast output load, (3, is set to.5. The maximum look-ahead depth, L, is set to, meaning that the switch can examine 2 cells in the queues within cell time. Under the Bernoulli process, the multicast traffic has slightly lower average delay than the unicast traffic until the output load reaches.9. The multicast average delay is calculated as the time it takes to send the multicast cell to reach all its destinations. Therefore multicast cells can stay in the queue for a longer period than unicast traffic. When the switch output load is higher than.9, a decreased average delay of the multicast traffic is observed since the integration policy gives the highest priority to the multicast traffic. Under the bursty process, the multicast average delay is significantly lower than the unicast delay because of the the integration policy. Fig. 5 shows the comparisons between the proposed integration scheme and the Fair ntegration (F) [] that uses WBA for multicast scheduling under two traffic processes. The proposed integration scheme uses MLRRMS for the multicast scheduler and by looking up to one cell stored further in the multicast FFO queues, the multicast delay is reduced significantly under high output load. Using MLRRMS to schedule multicast causes an increased delay to the unicast traffic under high output loads. This is due to that the lookahead mechanism of the MLRRMS can reduce the multicast HOL blocking, leaving fewer outputs for the unicast traffic comparing to the WBA. Fig. 6 compares the average delay of unicast traffic under the F and the proposed integration scheme with maximum look-ahead depth L and 2 under both Bernoulli and bursty processes. As the L increases, we can observe an increase of the unicast delay. This is due to that multicast scheduler grants more transmission opportunities to the multicast traffic and following the multicast-first policy, more unicast decisions are left for the next time slot. However, if the L is limited to, the unicast average delay is not significantly increased. Fig. 7 presents the multicast average delay comparisons 62

5 en (J) gj (a) Bernoulli tratlic (a) Bernoulli tratlic en en (J) (J) (b) Bursty traffic (b) Bursty tratlic Fig. 5. Average delay vs. output load (oa) for the proposed integration scheme and the Fair ntegration,f3.5 Fig. 6. Unicast average delay vs. output load (oa) for the proposed integration scheme and the Fair ntegration,f3.5 under the F and the proposed integration scheme with max imum look-ahead depth L and 2. t can be observed that the multicast average delay decreases as the L increases from to 2, especially obvious when the output load is high. We can see that the improvement gained on the increase of L in comparison with the complexity added to the switch (improved performance/increase complexity) is the largest under both traffic models when L. This allows us to design an integrated scheduling system with limited hardware implementation complexity. stored in the queues, MLRRMS reduces the HOL blocking problem for the multicast traffic. At most, two iterations are required by the multicast scheduler before it generates the decisions. This can align with the iterative islp to provide significant improvements on both traffic. Simulation results demonstrate that the proposed scheme outperforms the scheme that uses WBA in terms of switching latency, under both Bernoulli traffic process and bursty traffic process. V. CONCLUSON This paper proposes a integration scheme of VOQ unicast and the multi-level round-robin multicast scheduling (MLR RMS) that reduces the potential multicast HOL blocking prob lem. Unicast and multicast traffic are scheduled independently. Scheduling decisions are integrated following a multicast-first policy. During the integration, some decisions are left for the next cell time, which is called remainder. The remainder is looped back to the filtering function to ensure it is served in the next cell time. By allowing the switch to look one cell further 63 RFRNCS [] M. Karol,M. Hluchyj, and S. Morgan, "nput versus output queueing on a space-division packet switch," Trans. on Commun., vol. 35, no. 2,pp ,987. [2] T. Anderson, S. Owick,. Saxe, and C. Thacher, "High-speed switch scheduling for local-area networks," ACM Trans. on Net., vol., no. 4,pp ,993. [3] N. McKeown, "The islp scheduling algorithm for input-queued switches," ACM Trans. on Net., vol. 7,no. 2,Apr [4] B. Prabhakar, N. McKeown, and R. Ahuja, "Multicast scheduling for input-queued switches,". Sel. Areas Commun., vol. 5, no. 5, Jun [5] A. Bianco,P. Giaccone,C. Piglione,and S. Sessa,"Practical algorithms for multicast support in input queued switches," in nti. Can! on High Performance Switching and Routing, 26,pp

6 ., "" (a) Bernoulli tratlic " " "" (b) Bursty tratlic Fig. 7. Multicast average delay vs. output load (A) for the proposed integration scheme and the Fair ntegration, f3.5 [6] D. Pan and Y. Yang, "FFO-based multicast scheduling algorithm for virtual output queued packet switches," Trans. on Comput., vol. 54, no., pp ,25. [7] --, "Bandwidth guaranteed multicast scheulding for virtual output queued packet switches,". of Parall. and Distributed Comput., vol. 69, no. 2, pp , 29. [8] W. Zhu and M. Song, "ntegration of unicast and multicast scheduling in input-queued packet switches," Computer Networks, vol. 5, no. 5, pp , 26. [9] A. Bianco, P. Giaccone,. Leonardi, F. N eri, and C. Pigione, "On the number of input queues to etliciently support multicast tratlic in input queued switches," in nt. Can! on High Performance Switching and Routing, 23, pp. -6. []. Schiattarella and C. Minkenberg, "Fair integrated scheduling of unicast and multicast tratlic in an input-queued switch," in nt. Can! on Commun., vol , Jun. 26. [] Q. Zhang, A. Marshall, and R. Woods, "A traffic manager for integrated queuing and scheduling of unicast and multicast P traffic," in nt. Can! on Telecom., Jul. 29, pp [2] H. Yu, S. Ruepp, and M. S. Berger, "nhanced fifo based roundrobin multicast scheduling algorithm for input -queued switches," lt Commun., vol. 5, no. 8, May 2. [3] --, "Multi-level round-robin multicast scheduling with look-ahead mechanism," in nti. Can! on Commun., Jun. 2. [4] J. Hayes, R. Breault, and M. Mehmet-Ali, "Performance analysis of a multicast switch," Trans. on Commun., vol. 39, pp , 99. [5] OPNT Modeler, 22. [Online]. Available: 64