Load Sharing with OCGRR for Network Processors Which Supports Different services

Transcription

1 1 Load Sharing with OCGRR for Network Processors Which Supports Different services R.Sunitha, Assistant Professor, CMS College of science and commerce. Abstract A new scheme for packet transaction in a router is used, which provides load sharing among multiple network processors for various traffic patterns by introducing an OCGRR scheduler at the output port of the core router and multiprocessors. A feedback control mechanism prevents processor overload. The incoming traffic is scheduled by using a mapping formula derived from the robust hash routing ( HRW scheme). For the individual flow mapping no state information need to be stored, but for each packet, a predefined set of fields like a mapping function is computed over the identifier vector. To the HRW scheme an adaptive extension was provided to cope with biased traffic which possesses minimal disruption property. We use a scheduling algorithm called OCGRR to overcome the difficulty of handling different traffic patterns which influence the performance of load sharing scheme at the output port of the core router and the multiprocessors sharing its work.ocgrrt defines a stream to be the same-class packets from a given immediate upstream router destined to an output port of the core router. Each stream within a class can transmit a number of packets in the frame based on its available grant.provide, but only one packet per small round, thus reducing the inter transmission time from the same stream and achieving a smaller jitter and startup latency. Thus we introduce the OCGRR algorithm concept with the adaptation algorithm for sharing load with different traffic patterns like text, audio, video. We also verify and demonstrate the good performance of our scheduler by comparison with other algorithm in terms of packets dropped, queuing delay, jitter, and start-up latency. Keywords - QOS, HRW, load sharing, router, network processor, Packet scheduling, fair queuing, Differentiated Services, Round Robin. 1. Introduction : Network QoS refers to the ability of the network to handle this traffic such that it meets the service needs of certain applications. Dr. Antony Selvadoss Thanamani, Reader in Computer Science, NGM College, Pollachi. selvdoss@yahoo.com Networks interconnect hosts using a variety of network devices, including host network adapters, routers, switches, and hubs. Each of these contains network interfaces. The interfaces interconnect the various devices via cables and fibers. Network devices generally use a combination of hardware and software to forward traffic from one interface to another. Each interface can send and receive traffic at a finite rate. If the rate at which traffic is directed to an interface exceeds the rate at which the interface can forward the traffic onward, then congestion occurs. Network devices may handle this condition by queuing traffic in the device's memory until the congestion subsides. This paper first analyzes the results of experiments measuring the performance for sharing packet-processing tasks among multiple network processors within a router and then adds the OCGRR concept into it to improve its performance. We have extended the mapping with an adaptation discipline aimed at keeping the processor load below a dynamically derived threshold. The threshold reflects the total system workload intensity. 2. Existing System: The previously centralized router devices with a single general-purpose processor could not cope with the ever-increasing workloads and are being replaced by routers of more effective architectures, distributed or parallel [2]. In this thesis we propose QoS guaranteed adaptive load sharing by granting priority to the load incoming to the network processors. The main goal of this proposal is to maximize total network revenue so that quality of service (QoS) requirements for different traffic services is fulfilled. 2.1 Load Sharing for n/w processors : A feedback control mechanism prevents processor overload in the adaptive load sharing for n/w processors [7]. Incoming traffic is scheduled to multiple processors based on a deterministic mapping. The mapping formula is derived from the robust hash routing (also known as the highest random weight HRW [2]) scheme. 16

2 2 then sent for resolution to NPU j, f( v r ) = j. At NPU j, the packet information vector w ur is processed and the resolution information about the treatment to be applied to the packet (next hop, outgoing switch port, QoS applied) is returned to the requesting unit. Then, the packet is switched to the correct outgoing port and the corresponding packet alterations or manipulations, based on the resolution results, are applied. fig 2. Load sharing with feedback No state information on individual flow mapping has to be stored, but for each packet, a mapping function is computed over an identifier vector. An adaptive extension to the HRW scheme is provided to cope with biased traffic patterns. It was proved that adaptive load sharing for network processors possess minimal disruption property with respect to the mapping and minimizes the probability of flow reordering. But it is not efficient for various traffic patterns. In order to support different services a credit based algorithm called OCGRR is used capable of handling different packet sizes and traffic types at the core router of a network. An OCGRR [1] scheduler resides at each output port of a core router which schedules traffic into a router at its immediate downstream. Every core router can request its edge routers to adjust their arrival rates to satisfy the QoS performance level. Fig 1. Load sharing scheme abstraction The HRW mapping possesses properties like Load balancing, Minimal disruption which is useful for providing fault tolerance. The control loop prevents over-utilization of a single processor. They also minimize the amount of packet-to-npr remapping The Adaptation scheme : 2.2. SCHEME OF LOAD SHARING Packet -to-npr mapping : In load-sharing scheme using deterministic mapping f (see fig.1) the load of each input is distributed for processing among the NPRs. Here the mapping f is computed over the identifier vectorv r. The computation f( v r ) = j determines the particular NPU j to which the packet is mapped for processing. The function f( v r ), f : V {1, 2,...,m}, splits the vector space V into m exclusive subspaces Vj. Packets from a particular subspace are all mapped to the same processor. Upon arrival of a packet at an input, the packet is parsed to extract the fields relevant for packet processing, i.e., the identifier vector v r and the packet information vector w ur. The packet is buffered, the mapping f( v r ) is computed and the packet information vector w ur is Fig : An adaptation algorithm scheme The adaptation scheme works in the following general way periodically, the CP gathers information about the workload intensity of the NPUs see fig 2. If an adaptation threshold is exceeded, the CP adjusts the weights of the mapping f. The new multiplicative weights vector is then downloaded to the NPUs. Let x(t) be the instance of weights vector x used to compute f(t) at time t. In order to evaluate the status of individual processors, we need a processor workload intensity indicator. For that purpose, we introduce a smoothed, low-pass filtered processor workload intensity measure. The adaptation 17

3 algorithm consists of two parts (see Fig. ): the triggering policy, which specifies the conditions to act, and the adaptation policy, which specifies how to act. A trigger is periodically evaluated and, based on the result, specific action is taken. 2. Output Controlled Grant_Based Round Robin Algorithm Concepts. In order to support different services in the core router of a network a credit based algorithm was used. This has the capability of handling different packet sizes and traffic types. Algorithms like SRR, DRR, DRR++, OCRR has its own drawbacks like higher service time, higher startup latency and jitter. To overcome it a common approach to support DiffServ traffic saves all same class packets from different sources in a shared FCFS (First Come First Served) buffer. It is difficult to control the service order of packets from different sources. An OCGRR(Output Controlled Grant-based Round Robin) scheduler resides at each output port of a core router which schedules traffic into a router at its immediate downstream. Every core router can request its edge routers to adjust their arrival rates to satisfy the QoS performance level. Each output port can receive R streams. The traffic monitor unit gives information on the arrival rate of streams.a dedicated buffer is used for each stream in a class. A shared buffer is used for all traffic streams within each class into which the overflow packets from buffer i dedicated to stream i within class J will be redirected. The OCGRR algorithm used the Grant parameters to specify If the stream i has underused or overused its allocation. Here Gj,i(t) is the available grants for the stream at time t. Uj,i(t) is the total used grants for the stream at time t. The rate parameter specifies the average arrival rate of stream i packets in class j. OCGRR handles the packets overflow by transferring it to the shared buffer. Backlogged stream is a stream with at least one packet in its buffer with the available grant. The active list contains the references of all backlogged streams within class J where a dedicated linked list is used for class J. Using the active list the traffic is scheduled within class J Frame based operation: The scheduling sequence starts from the EF traffic followed by the class-2 traffic and finally moves to the BE traffic. Each frame is divided into EF rounds, class-2 rounds, BE rounds. In a Backlogged stream a round of each class can send only one packet.within the frame we can limit the Fig. 4. The scheduling sequence model. number of bits to be transmitted within a frame using the variable T based on stream R, the average size of packet l and the class indices Ci. R+1 indicate the possibility of allocating upto R+1 buffers in each class. The frame can end when the traffic transmitted within the frame exceeds T or when the actual length is l >T. Note that the lower priorities streams access right to bandwidth is saved for the future frames. L<T when there is no backlogged stream in any class Algorithm Explanation (OCGRR) : Fig. 6 presents the OCGRR flowchart, where Q represents a stream, and Xbits is the total transmitted bits in a frame. The Enqueue process adds a new packet of any stream in its relevant buffer. It then appends the stream reference to the relevant ActiveList provided that the stream not in the ActiveList becomes backlogged (i.e., had a positive grant but was empty). The scheduling for each class is divided into two parts: 1) In the DequeueInit process, the grant of each stream inside class J is incremented by some quantum computed based on the frame beginning time. Then, if a non backlogged stream becomes backlogged, its reference is appended to ActiveList J; and 2) in the DequeueProcess, packet scheduling is performed. There are two scheduling processes: one for the classes that use a dedicated buffer per stream in a class, and the other for the classes that use one shared buffer for all streams in a class. The scheduler visits the backlogged streams one by one. This continues until the last backlogged stream in the schedule domain J is visited. Then, the scheduler starts visiting the backlogged streams from the head of ActiveList J. We call this scheme Multiple Round Robin (MRR) transmission. Scheduling for class J is stopped whenever there is no backlogged stream left in ActiveList J or the total transmitted traffic exceeds T. The latter condition is evaluated at the end of each packet transmission. When scheduling a packet from a stream, OCGRR first schedules the packet with any size and then 18

4 4 updates both grant values of the stream with the transmitted packet size. This update may lead to a negative grant if the size of the transmitted packet is greater than the stream s grant. When the stream s grant is negative, the stream becomes non backlogged. OCGRR flags. If the frame ends right before processing stream i in class J, this stream in the Last ListPtr J parameter related to class J. At a future frame when it is the turn of class J to be processed, the scheduler would start processing class J from the stream referenced by LastListPtr J, but not from the beginning of the streams referenced by ActiveList J. This is handled by ListPtr = GetStartingStreamRefInRound in Fig. 6. This process ensures fairness for stream i. It is also in the direction of reducing the intertransmission time from the same stream. to manage the bandwidth of the streams. At time t, the total given grant to stream i in class J, Total g rant j, i( t ), must be proportional to its AAR (Average Arrival Rate), where the numerator is the total given grant to stream I until time t. Although the parameter K could be set as a constant value without considering the traffic arrival rate and available bandwidth, say K= 1, we determine K by the following heuristic parameter The parameter K is used to adjust the allocated quantum according to the output bandwidth C, the class indices, and the AAR values. If the average arrival rates remain unchanged, K is constant. This parameter can therefore control the transmission policy from different classes. For example, in a router with ψ =, C = 100 Mb / s. C1 = 1, C2 = 0.8, C = 0.5, λ1 = 10 Mb / s, λ2 = 0 Mb / s, and λ = 60 Mb / s, Fig. 6. The flowchart of the OCGRR scheduling algorithm Grant Calculation : Before processing the streams within class J, each stream in class J obtains some quantum proportional to its AAR. Let t denote the starting time of a new frame. The time variable t is a virtual time that can be measured from any convenient reference point like when the router powers up and everything is reset. This virtual time is the same for all streams of all classes. Hence, only one virtual time is enough we have K=1.56. Consider the second case withc 1 = 1, C 2 = 0.9, C = 0.7, one obtain K=1.26. Since R and l are the same for both cases, the parameter Γ in the first case is smaller than the second case. Hence, the service of the highest priority streams is much better in the first case. Lower traffic loads also lead to a higher K, and streams can use the unused bandwidth flexibly. Now, let us increase class arrivals to, λ 1 = 50 Mb / s, λ 2 = 70 Mb / s, and λ = 80 Mb / s and keep the other parameters as before. Then, we obtain K =0:68 and K = 0:59 in the first and second cases, respectively. This gives a 19

5 5 chance to all classes to have access to the bandwidth. If we keep K = 1, class- will unlikely be served at all. Thus, by controlling K, the starvation of lower priority traffic can be avoided in a congested network. At time t, stream i should obtain some quantum of QJ, i ( t ) for the new frame until the end of the frame, t + T f,so that the total given grant to stream i in class J at timet + Tf still satisfies the above condition, i.e., 4. Result and Discussion : We have used the java 1.6 environment on an Pentium core2duo with Microsoft Windows XP machines to simulate a model of a router with multiple NPUs and line cards. It compares the number of packets dropped under the two scenarios Adaptive load sharing and Adaptive load sharing with OCGRR concept integrated within it. The dropped packet is very low in AdaptiveLs with OCGRR. Adaptive LS[7]shows maximum packet drops. This shows that QOS guaranteed should be implemented for adaptivels. Table 2 : Packets dropped at Time t Time Adaptive LS with OCGRR Adaptive LS where T f is the expected frame transmission time obtained by Since the exact frame period is unknown at the start of the frame, T f is taken as an estimate for the duration of the logical frame period. Rearranging, Then, the available grant for stream i is calculated from. Now, if becomes positive and the stream is nonempty, the stream reference is appended to ActiveList J.. Integration of OCGRR scheduling with network Processors : The OCGRR scheduling is integrated with all network processors to improve the quality of service. When the incoming traffic reaches the router, an OCGRR scheduler residing at each output port of a core router schedules the traffic into a router at its immediate downstream and grants a priority to it, as it is specified in the OCGRR algorithm. A traffic monitor unit gives the scheduler information on the arrival rate of streams. Thus it can handle various traffic patterns efficiently. The incoming load is shared among the NPU s. Incoming traffic is scheduled to multiple processors based on a deterministic mapping. Using the Adaptation algorithm the NPU s workload intensity is maintained below a dynamically derived threshold value. The threshold reflects the total system workload intensity. Figure 7 Comparisons of dropped packets Fig. 8 compares the per-processor workload intensity for each method. Clearly, AdaptiveLS workload intensity remains within close vicinity of the ideal workload intensity when AdaptiveLS with OCGRR sharing is deployed. Table 1 : Workload Intensity at Time t Time Adaptive LS With CGRR Adaptive LS

6 6 Figure 8 Comparison of work load intensity In Fig. 9, the number of fraction of traffic per flow remapping for the adaptive load-sharing method is examined. Note that the number of flows remapped per iteration is several orders of magnitude smaller than the number of flows appearing per iteration. Table : Peformance achieved for the maximum fraction of traffic flow Max Adaptive fraction LS with OCGRR Adaptive LS The maximum fraction of the total rate of an interface fig.9 (in packets per second) a single flow is allowed to occupy. If,a single flow it may occupy up to 50% of the interface transport capacity. Note that the study of flow results suggests that such limits on the maximum per-flow rate are in agreement with reality. Figure 9 Max fraction of traffic flow vs result 5. Conclusion: This scheme is proposed for sharing packetprocessing tasks among multiple network processors within a router. The scheme is based on an adaptive [7] deterministic mapping of flows to processors. The proposed load-sharing scheme requires no flow state information to be stored within a router. The OCGRR scheduling is integrated with all network processors to improve the quality of service. Adaptive load sharing with OCGRR keeps the processor load below a dynamically derived threshold. The threshold reflects the total system workload intensity. The adaptation is performed by adjusting the weights of the packet-to-processor mapping, thus reducing or increasing the amount of flows a processor must handle. Thus, the probability of packet reordering within a flow is kept low. OCGRR has the features of using smaller frame lengths and rounds; sending traffic packet by packet in smaller rounds; reducing the intertransmission time from the same stream; reducing queuing delay, jitter, and startup latency; controlling the starvation of lower priority classes; and beginning the transmission in each class from a delayed stream in the previous logical frame to ensure low latency and fairness. It can also keep the fairness for streams at an acceptable level. A desired QoS performance can be obtained by adjusting class indices. References : [1]Akbar Ghaffar Pour Rahbar, Oliver Yang,OCGRR:A new scheduling algorithm for differentiated sevices networks,ieee Transactions on parallel and distributed systems,may [2] K. W. Ross. Hash routing for collections of shared web caches. IEEE Network, Vol. 11, No. 6, November-December [] Cisco Express Forwarding (CEF). Cisco Systems white paper, [4] G. C. Fedorkow. Cisco Edge Services Router (ESR) technology overview, [5] A. Asthana, C. Delph, H. V. Jagadish, P. Krzyzanowski. Towards a Gigabit IP router. Journal of High Speed Networks, Vol. 1, No. 4, pp , [6] D. G. Thaler, C. V. Ravishankar. Using namebased mappings to increase hit rates. IEEE/ACM Transactions on Networking, Vol. 6, No. 1, pp. 1-14, February [7] ukas Kencl, Jean-Yves Le Boudec, Adaptive load sharing for Network processors,ieee/acm Transaction on networking. 141