Implementation of Evolvable Fuzzy Hardware for Packet Scheduling Through Online Context Switching

Transcription

1 Imlementation of Evolvable Fuzzy Hardware for Packet Scheduling Through Online Context Switching Ju Hui Li, eng Hiot Lim and Qi Cao School of EEE, Block S Nanyang Technological University Singaore Tel: Abstract: - Real-time cell scheduling is a romising area for the alication of evolvable hardware (EHW) In this aer, we describe an intrinsic evolvable and online adative EHW for solving the acket switching roblem Based on the coding and evolution scheme, we refer to it as evolvable fuzzy hardware (), an extension of our roosed evolvable fuzzy system (EFS) [, ] and reconfigurable fuzzy inference chi (RFIC) [3] EFS is a good framework for online adatation RFIC on the other hand is a hardware architecture which suorts context switching or online reconfiguration By combining the advantages of the two, we can achieve an intrinsic evolvable and online adative The whole scheme is illustrated through alication in solving the acket switching roblem and the advantages are highlighted Key-Words: - Evolvable Fuzzy Hardware, Evolvable Fuzzy System, Cell Scheduling, Reconfigurable Fuzzy Inference Chi, On-line Adatation, Intrinsic Evolvable Hardware Introduction For modern networks, various services are currently suorted Voice, video, file transfer, etc, are some of the imortant services rovided by Internet, broad area network, and local area network The IPv6 (Internet Protocol version 6) is now starting to mature This new tye of communication rotocol can connect instruments all over the world into a single network It has significant advantages over the IPv4, which is currently widely used With the emergence of the advanced rotocol, new services not currently available may be needed in the future This tendency requires network nodes to accommodate for greater flexibility The multilexer is a very imortant network comonent for modern network infrastructure It is used mainly to rovide bandwidth sharing of high-seed link for network terminal equiments or network inter-nodes There are many scheduling algorithms available, such as first-in-firstout (), weighted round-robin (WRR), virtual clock [], dynamic weighted riority scheduling () [] and deficit round-robin (DRR) [] Each of these schemes has its own strengths and weaknesses In this aer, we investigate the alication of evolvable hardware (EHW) and some of its unique characteristics in cell scheduling Other than emulating general rocessor sharing (GPS) or designing the algorithm in a acket-by-acket mode, can rovide a latform of designing a scheduling system directly from the standoint of QoS requirements Evolvable hardware (EHW) was introduced in 99 by Higuchi et al [4], combining evolutionary scheme and reconfigurable hardware device to solve roblems EHW is suitable for working under a dynamically changing environment It is therefore romising in some situations whereby the oerating environment besides being dynamic, is also unknown or unredictable any research works till now deal with extrinsic EHW which carries out evolution by means of a software model and only downloads the elite into the reconfigurable hardware On the contrary, intrinsic EHW tries to do away with the software simulation model In an ideal case, both the evolution and reconfiguration can be carried out within the EHW such that it can trace and adat to the changing oerating environment to maintain good system erformance Due to ractical imlementation issues, intrinsic EHW achieved so far cannot entirely

2 break away from the need for an external comutational latform [5,6,7] The comlexity ertaining to the imlementation of intrinsic EHW deends very much on the alication areas One otential area of alication for intrinsic EHW is acket switching network There have already been some works done on EHW alication in AT cell scheduling [8,9] In their works, the authors resented schemes to solve this kind of roblem using functional EHW The functional EHW systems successfully evolved a circuit which can achieve similar service erformance as traditional scheduling scheme However, some significant limitations make the system not suitable for ractical alications The most significant limitation is that the system cannot evolve intrinsically and adat online We believe intrinsic evolvable and online adative EHW for such kind of alication is viable according to our revious work [,,3] The combination of these works can be termed as evolvable fuzzy hardware () The rest of this aer is structured as follow Section discusses the alication roblem and the evolution scheme in the roosed The roosed relies mainly on a reconfigurable fuzzy device termed as reconfigurable fuzzy inference chi (RFIC) In Section 3, our roosed RFIC will be outlined In Section 4, the simulation results of on AT cell scheduling will be given Section 5 will then resent the conclusion and give directions on some future work Evolution Scheme ultilexer is a very imortant comonent in acket switching network It adots time division sharing scheme to rovide bandwidth sharing for different flows In this aer, two classes of cell flows with fixed acket size are mainly considered Class refers to cell flow which is sensitive to cell delay, for examle, Constant Bit Rate (CBR) On the other hand, Class refers to cell flow such as non-real-time Variable Bit Rate (nrt-vbr) which are not sensitive to cell delay but to cell loss The whole system structure of for acket scheduling is described as in Fig BUF# and P unit are normal comonents as in traditional scheduling scheme BUF# in Fig is used to store cells waiting for time slots P is the multilexing unit for allocating time slots to Class and Class In this roblem, the size of BUF# is fixed It is also assumed that the caacity of OUT channel is the same as the caacity of the inut channels, 555hz The erformance of the multilexing system can be described by the quality of service (QoS) which includes Class cell delay, delay variation and Class cell loss Due to the system s comlexity, the evolution granularities such as transistor, gate, functional unit are not very suitable for evolving owerful circuits in a very short time In order to overcome such a roblem, we can adot fuzzy rules as the evolution granularity In Fig, TB# is the training buffer for acquiring training data online RFIC block is a secific imlementation of reconfigurable fuzzy inference chi described in [3] Evolution odule is a hardware comonent to erform GA oerations in order to evolve the desired fuzzy rule sets Scheduling odel is used to simulate the behavior of P unit on the cell flow stored in TB# In a simle way, Scheduling odel can be an embedded scheduling system It includes two comonents, P simulator unit and RFIC unit Every fuzzy rule set evolved by the Evolution odule can be evaluated by downloading onto the RFIC in the Scheduling odel By incororating the Scheduling odel, the evolved fuzzy rule sets can be evaluated without affecting the system s oeration Because of the unredictablility of cell flow, it is difficult to train a system that works equally well for all cell flow scenarios The rationale for choosing an aroriate data set for training can be justified based on the rincile of locality If we assume that the time eriod is very small, the cell flow of the next time eriod will be very similar to the current time eriod Based on this justification, it is suitable to emloy TB# of finite length to collect training data for the evolvable system In order to control the scheduling behavior using fuzzy rule set, we define two fuzzy variables c and c to describe the status of BUF and BUF resectively c is the ratio of Class cell rate and the caacity of the OUT channel c is the ratio of the number of emty units in BUF and the length of BUF The membershi functions for c and c are as shown in Fig A fuzzy rule set for cell scheduling is as shown in Table In Table, T means that the OUT channel is allocated to Class and F means that Class will be sent through the OUT channel For genetic coding, the string,,,, reresents the fuzzy rule set in Table The value corresonds to T and corresonds to F A in

3 the chromosome indicates there is no fuzzy rule defined for the corresonding inut situation The rule set in Table is designed based on human knowledge and intuition It can be emloyed as a core rule set in the system In this system, the evolution system tries to derive a good chromosome for a secific situation, not necessarily an otimal one After a fixed number of generations, if a better chromosome is derived, the current working chromosome is relaced immediately Otherwise, the current rule set continues to be alicable In order to guarantee a certain level of minimum accetable erformance, the system relies on the core rule set shown in Table as a default startu rule set Class Class Fig Adatation framework for Fig embershi function Table A fuzzy set for AT cell scheduling c Training buffer (TB) c BUF BUF Scheduling odel VS Training buffer (TB) S VS S L VL VS T F F F F S T T T F F T T T T F L T T T T F VL T T T T T L P RFIC Evolution odule VL SEL OUT In Table, each cell entry is interreted as a fuzzy rule that mas the inuts c and c to the outut SEL For examle, the first fuzzy rule reresents if <c is VS> and <c is VS> then <SEL is T> This means that the strength of firing is taken as the minimum of the degrees of matching between the two inuts and the antecedents of the rule The fuzzy aggregation is carried out by averaging the fuzzy conclusions derived for all the rules There are two aggregation results for the T and F oututs The larger value determines the final cris conclusion In this system, we fix the rule number to be a small number to kee the search sace manageable The fitness function adoted in the roosed system is described by Eq, and 3 F = κ AveDelay λ DelayFactor () τ AveDelay = m() i τ () i= DelayFactor = ρ υ (3) In Eq, κ is a very large constant λ is an adjustable arameter to indicate the desired Class cell delay AveDelay reresents the average cell delay suffered by the cell flow stored in TB DelayFactor is a constant used as a reference for scaling the value of λ based on the desired Class cell delay In Eq, τ is a variable denoting the number of Class cell units in TB sent during evaluation mi () is the waiting time of the ith cell in TB before being sent In Eq3, ρ is a constant corresonding to the time required to send a cell through the outut channel The value of ρ deends on the bandwidth caacity of the outut channel υ denotes the size of TB# 3 Hardware Imlementation In the system architecture of Fig, there are two RFIC blocks, which are the key comonents One RFIC block is for acket scheduling control and the other is included in the Scheduling odel block For the Scheduling odel, chromosomes need to be evaluated within a very short time eriod The duration of the evaluation can significantly affect the system erformance The concet of RFIC is illustrated by the block architecture in Fig3 [3] There are 4 main blocks that make u the RFIC, FI (fuzzy inference ma), CU (context memory unit),

4 AE (address encoding mechanism) and OA (outut aggregation mechanism) In Fig3, k indicates the number of inut bits refers to the width of the data bus v and w are the sizes of the linguistic term sets for Inut and Inut resectively m deends on the size of the linguistic term set for the outut variable In the alication for cell scheduling, k is 5, is 5, v and w are 5 and m is In the RFIC, the FI is a critical art for accommodating the various fuzzy contexts It stores all the fuzzy conclusions for every inut situation For a rule set of 5 rules as in Table, the FI block is divided into 5 grouings called artition blocks (PB) Each PB stores the fuzzy conclusions covering all ossible inut situations For examle, PB <,> maintains a maing for the fuzzy antecedents such as if <c is VS> and <c is VS> The digitized inuts c, c and genes of the working fuzzy rule set stored in the context register of CU can be combined together to address a secific memory location within this PB The content in every PB is determined by the adoted inference scheme Inut Inut k k AE Sel Sel Selvw m m m m k PB <v,w> PB <,3> PB <,> PB <,> CU v w The content of the context register in CU is the working fuzzy rule set CU generates Ena and Sel signals for the PBs based on the context register The Ena signal is used to decide which PB is active The Sel signal combines with the outut of AE to form the address signals to access secific memory location in every active PB AE is the address generator unit It is resonsible for generating major address signals for the PBs of the FI Its outut is derived by combing the two digitized inuts OA is a circuit to fulfill fuzzy aggregration It includes two arts of aggregation for T and F resectively which is comosed of Ave_ blocks and Ave_3 block to erform -averaging and 3-averaging [3] The inuts of OA are from every PB If one PB is active, it oututs the data stored in the unit secified by the FI Ena D D D4 D5 Fig3 Scheme of RFIC OA Fuzzy_OUT outut of AE and the outut signal Sel of CU to OA, otherwise it oututs 4 Simulation In order to demonstrate the viability of for solving acket scheduling roblem, we carried out simulations on the cell flow scenario in Fig4 The simulation results of will be comared with firstin-first-out () and dynamic riority scheduling () [] is a very general scheduling scheme which can achieve very good balance of cell losses but very oor Class cell delay is an imrovement of round-robin scheduling scheme It can adat to the changes of the cell flow Class Class ON OFF Fig4 Two classes of cell flows For the cell flow scenario, Class is the cell flow with cell bit rate of 555Hz while Class is the cell flow also with cell bit rate of 555Hz The difference between Class and Class is that Class has constant cell bit rate while Class has ms ON eriod and ms OFF eriod For the simulation, the length of BUF# is units, the length of TB# is 3 units, the oulation size is and the generation number is 4 The simulation results of ( λ =35), and are as given in Fig5, 6, 7, 8, and 9 From Fig5 we can see that can achieve lower Class cell delay than and From Fig6 it can be seen that the balance of cell losses by using these three systems are good, which means that none of them biases much on Class cell flow Fig 7, 8 and 9 resent the delay distribution suffered by Class In Fig7, it can be seen that most of the Class cells suffer delay between 7µS and 4µS for the The robability of cell delay above 4µS aroaches, indicating a small delay variation for Class ackets The delay for in Fig8 is mainly concentrated around two bars, 7µS and 53µS The robabilities for these two values are both very significant The delay for in Fig9 is in the range of 7µS and 47µS But besides the high robability of 7µS, the

5 delay for also concentrates around 4µS From these comarisons, it is concluded that can rovide smaller delay variation than the other two algorithms Another very good roerty of is that the system s erformance can be adjusted very intuitively by decreasing or increasing the value of λ in Eq The smaller the value, the smaller the Class cell delay This roerty cannot be achieved conveniently using traditional scheduling methods The tunability of can be seen from the simulation results Results of simulation for different λ values are as shown in Fig and In Fig and, when λ is 4, the Class cell delay and Class cell loss are very small Accordingly, the Class cell delay and Class cell loss are very big If good balance of Class cell loss and Class cell loss is desired, a bigger value can be assigned to λ In Fig and, both the Class cell loss and Class cell loss are moderate when λ is 6 In the case when the QoS of Class needs to be significantly emhasized, the value of λ can be further increased The larger the λ, the better the QoS of Class The QoS for Class when λ is 8 is the best in Fig and In rincile, Class cell delay can be adjusted in the range from to ρ υ if λ is within and This means that Class cell delay has a very wide range of tunability In another word, according to the correlation between cell loss and cell delay, Class cell loss and Class cell loss can also be tuned to a very wide range when λ is adjusted According to the fitness function, a rough Class cell delay can be estimated after deciding the value of λ On the other hand, if one knows the satisfactory Class cell delay requirement, the value of λ can also be aroximated This delay tunability is very useful to give some riorities to other flows under the case when a flow does not require much on small delay but has requirement of small delay variation This hilosohy can take better advantage of the network bandwidth and resource without much effects on network QoS 5 Conclusion and Future Works In this aer, we described the acket switching roblem and the evolution scheme of for solving this roblem We also addressed the imlementation issues of by describing the RFIC, a reconfigurable fuzzy inference chi that is able to average cell delay (us) handle real-time context switching For the demonstration of the viability of, the simulation results for the acket switching were given which shows that the system can erform as well as other scheduling schemes, namely and The can also rovide better roerties than other schemes in terms of flexibility The more advantageous asect of is the tunability This was also demonstrated through simulation results From the research work shown in this aer, it is evident that is very suitable for controlling data flow It can adat to the changes of the cell flows and effectively control the multilexing of the cell flow Our future work will focus further on studying the characteristics of system and its hardware imlementation Class Cell Average Delay x 6 Fig5 Cell delay of Class and Class 5 x 5 Class Cell Loss x 6 EFS Class Delay Distribution x average cell delay (us) Fig7 Class delay distribution for EFS Class Cell Average Delay 35 x 5 Class Cell Loss Fig6 Cell loss of Class and Class x 6

6 average cell delay (us) Class Delay Distribution Fig8 Class delay distribution for Class Delay Distribution Fig9 Class delay distribution for Class Cell Average Delay x 6 35 x 5 Class Cell Loss average cell delay (us) Fig Cell delay tunability x 6 Fig Cell loss tunability Class Cell Average Delay sim ulate tim e (us) x 6 5 x 5 Class Cell Loss sim ulate tim e (us) x 6 References: [] JH Li and H Lim, Evolvable fuzzy system for AT cell scheduling, LNCS66, (Proc of ICES 3), 8-7, 3 [] JH Li and H Lim, "A Framework for Evolvable Fuzzy Hardware," Secial session on Evolutionary Comuting for Control and System Alications, ICONIP/SEAL/FSKD, 8- Nov, Singaore [3] Q Cao, H Lim, JH Li and WL Ng, A Context Switchable Fuzzy Inference Chi, submitted to IEEE Transactions on Fuzzy Systems [4] T Higuchi, T Niwa, T Tanaka, H Iba, HD Garis, and T Furuya, Evolving Hardware with Genetic Learning, Proc of Simulation of Adative Behaviour, 47-44, IT Press, 99 [5] Iwata, I Kajitani, Y Liu, N Kajihara, and T Higuchi, Imlementation of a Gate-Level Evolvable Hardware Chi, LNCS (Proc of ICES), 38-49, [6] J Langeheine, K eier and J Schemmel, Intrinsic Evolution of Quasi DC Solutions for Transistor Level Analog Electronic Circuits Using a COS FPTA Chi, Proc NASA/DoD Conference on Evolvable Hardware, 75 84, [7] J Langeheine, S Folling, K eier and J Schemmel, A COS FPTA Chi for Intrinsic Hardware Evolution of Analog Circuits, Proc of 3rd NASA/DOD Worksho on Evolvable Hardware, 7-75, Long Beach, CA, USE, IEEE Press [8] W Liu, urakawa and T Higuchi, AT Cell Scheduling by Function Level Evolvable Hardware, Proc of ICES996, LNCS 59, 8-9, Sringer-Verlag, 997 [9] W Liu, urakawa and T Higuchi, Evolvable Hardware for On-line Adative Traffic Control in AT Networks, Genetic Programming 997, Proc of the Second Annual Conference, 54-59, organ Kaufmann Publishers, 997 [] T Lizambri, F Duran and S Wakid, Priority Scheduling and Buffer anagement for AT Traffic Shaing, The Seventh IEEE Worksho on Future Trends of Distributed Comuting Systems, Dec 999 [] L Zhang, VirtualClock: A new traffic control algorithm for acket switching networks, in Proc AC SIGCO 9, Aug99, 9-9 [] Shreedhar and G Varghese, Efficient fair queuing using deficit round robin, in Proc AC SIGCO 95, 995, 3-4