Dynamic Tag Estimation for Optimizing Tree Slotted Aloha in RFID Networks

Transcription

1 Dynamic Tag Estimation for Optimizing Tree Slotted Aloha in RFID Networks Gaia Maselli, Chiara Petrioli, Claudio Vicari Computer Science Department Rome University La Sapienza, Italy {maselli, petrioli, ATRACT The emergent commercial use of techniques for Radio Frequency-based IDentification of different items (RFID) requires the investigation and testing of collision resolution mechanisms for the efficient and correct communication between the system reader and the tags labeling the items that need to be identified. Several MAC protocols have been proposed to resolve collisions in RFID networks. A recent solution, named Tree Slotted Aloha (), has been shown to outperform previous ones with respect to the time it takes for identifying all tags, and the total number of bits transmitted to complete the identification process. However, almost half of the time needed by for identifying tags is spent in collisions. This depends on operation and in particular on the way estimates the number of colliding tags. We have observed that in the case of realistically large networks, highly underestimates this number, with non-negligible impact on the protocol performance. In this paper, we propose a Dynamic Tree Slotted Aloha (Dy ) protocol that exploits the knowledge acquired during ongoing readings to refine the estimation of the number of colliding tags. In so doing, Dy adapts the length of the following reading cycles to the actual number of tags still requiring identification. Through ns2-based simulations we show that the proposed method is effective for tag identification and results in significantly improved performance over. Specifically, the length of the identification process is up to 2% lower than that of. Furthermore, the amount of transmitted bits needed for identifying all tags decreases up to 3%. Categories and Subject Descriptors C.2.1 [Network Architecture and Design]; C.2.5 [Local and Wide-Area Networks] General Terms Algorithms, Performance. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MSWiM 8, October 27 31, 28, Vancouver, BC, Canada. Copyright 28 ACM /8/1...$5.. Keywords Radio Frequency IDentification, passive tags, anti-collision protocol. 1. INTRODUCTION Radio-Frequency IDentification (RFID) is considered a key technology for item identification, fast and efficient object tracking, and enables fundamental operations such as automatic inventory and management. An RFID system consists of radio frequency identification devices, named tags, that are able to communicate wirelessly to one or more readers. are attached to objects that need to be identified and answer with their ID when inquired by a reader. Typical applications require cheap and unobtrusive identification tags to be attached to various different items (T-shirts, cereal boxes, books, animals, etc.). To this aim tags should be small, light, low cost and able to operate independently of batteries or external sources of energy. Passive tags fulfill this purpose by receiving the energy needed for communication by the RF carrier of the inquiring reader. The request message from a reader gives the tags enough energy to remodulate the signal so that their ID information is backscattered to the reader. The major challenge for RFID systems is that of avoiding or solving collisions due to interference that might occur among readers and/or tags. The presence of multiple readers in the same area may cause reader-reader collisions, if the signals of two or more readers interfere with each other, and reader-tag collisions, when more than one reader attempts to communicate with the same tag. Collisions can also be caused by two or more tags simultaneously transmitting to the same reader (tag-tag collisions). In this paper, we specifically address this problem, called tag collision, in a single-reader scenario. Since the system is highly asymmetric (the reader is resource-rich while tags have very limited storage and computing capabilities, and are unable to hear the signal transmitted by the other tags and to detect collisions) channel access should be arbitrated by the reader. This is the case of the many anti-collision protocols that have been proposed in the literature so far. The different solutions can be classified into two major categories: Alohabased and tree-based. Mimicking slotted aloha [9], protocols in the first class consider the channel to be slotted into intervals of time, whose duration is equal to the tag s ID transmission time. are required to begin an ID transmission at the beginning of a time slot, that is signaled by the reader with a short message. As the reader is not aware of the num- 315

2 ber of tags to be identified, it does not know how many slots are required to identify all tags. For this reason, the reader issues consecutive groups of slots, or frames, until it completes the identification process. Specifically, the reader starts by issuing a first frame whose length is fixed to a predefined value. that need to be identified, randomly and uniformly pick one slot in the frame for replying to the reader. At the end of the frame, tags that are successfully identified become silent for the rest of the process, while tags that generated collisions keep trying in the following frame. The length of the new frame is established according to the outcome of the previous one: information such as the number of idle slots, that of slots where collision occurred, and the number of slots where tags were identified, allow to estimate the tag population and properly dimension the following frames. Protocols of this kind differ in the way tags are grouped into frames, the way tag population is estimated and the way the frame size is decided. In [1], for instance, all tags which have collided in the previous frame participate in the next one. The frame size is based on the estimate of the population of unidentified tags, which is obtained at the end of the previous frame by means of the Chebyshev s inequality (Section 2). The Enhanced Dynamic Framed Slotted Aloha (EDFSA) [6] addresses the problem of selecting large frame sizes. As tags are very simple devices, they are not supposed to generate large random numbers (no more than 256 [6]). For this reason, EDFSA defines a predefined set of frame sizes (varying from 8 to 256) for different ranges of estimated unread tags. Typically, the frame size is around the mid point of the range. As an example, if the estimated number of unread tags is between 82 and 176, than the selected frame size is 128, while for a range between 177 and 354 tags, the frame size grows to 256. For larger tag populations, EDFSA randomly splits tags into groups of the maximum frame size (i.e., 256 tags). To give an example, if the number of unread tags is in the range (355, 77), than 2 groups are created. Only the tags associated with one of the groups are queried in the following frame to reply back to the reader. Chebyshev s inequality is used at the end of the frame to estimate the number of tags which have participated in it. Such an estimate is then used to refine the estimate of the global tag population, possibly adjusting the number of tag groups and the size of the following frames. Tree Slotted Aloha () [3] improves over EDFSA and [1] by more efficiently dealing with collisions. After the first frame, a new set of frames is allocated, each devoted to solving the collisions which have occurred in a given slot of the first frame. Only the (few) tags which transmitted into that slot participate in the corresponding frame. The approach is repeated: If collisions occur in one of the frames allocated to solve collisions (say, frame i), new frames are allocated to solve such collisions (one for each collision slot in frame i). Tree-based protocols draw on tree algorithms for packet broadcast channel [4][7], and take a substantially different approach. The tag identification process is deterministic, and it is based on iteratively querying a subset of tags which match a given property until all tags are identified. These protocols are called tree-based because the identification process can be represented as a tree where the root is the set of tags to be identified, intermediate nodes represent groups of colliding tags answering the same request from the reader, and the leaves correspond to single-tag responses. Tree-based protocols are grouped into two classes: binary splitting protocols and query tree protocols. Binary Splitting () [8] protocols recursively split answering tags into two sub-groups and keep splitting until single-tag groups are obtained. Each tag maintains a counter (initially set to zero). with the counter value equal to zero answer the reader query, while others remain silent until their counter decreases to zero. The value of the tag counter is modified if the query induces a collision, an identification, or no-answer (idle slot). In case of a collision, colliding tags add a random binary value to their counter. As a consequence, they are split into two subsets: Those whose counter is zero and those whose counter is one. not involved in collisions increase their counter by one. In case of identification or idle, all the tags decrease their counter by one. Those with the counter at zero will answer the next query. Query Tree (QT) protocols [5] decide which tags will answer each query quite differently. are interrogated by the reader based on their ID. Specifically, those tags whose ID has a prefix matching the value communicated in the query reply. Initially, the reader starts querying all tags, which is obtained by including a NULL prefix in the query. If a collision occurs, the prefix length is increased of one bit until the collision is solved and a tag is identified. The reader then starts a new query with a different prefix. In particular if tag identification occurred with a prefix q the reader will query for prefix q1. This corresponds to exploring the binary tree of the possible prefixes according to a layer-by-layer tree traversal. Such a binary tree has nodes at the i th level labeled with all the possible values of a prefix of length i (e.g., nodes at level 1 contain prefixes and 1, nodes at level 2 prefixes, 1, 1, 11 and so on). The exploration of a subtree is skipped in case there is only one tag matching the prefix stored in the subtree root (i.e., if a tag identification occurs when the reader queries with the subtree root prefix). The Query Tree Improved () [5] protocol optimizes the number of queries, avoiding those that will certainly result in collisions. As an example, consider the case in which prefix p generates a collision, while prefix p results in no tag answers. According to the reader skips prefix p1 that will certainly generate a collision and queries directly with p1 and p11. Query tree protocols are often referred to as memoryless, since the set of tags answering a query only depends on the current prefix included in the query, not on the past history. therefore do not need to keep any state information. Note that this is not the case for all the other classes of protocols (e.g., tags in maintain a counter, in the frame and slot number in which a given tag has last transmitted). Recent work on MAC protocols for RFID (confirmed by extensive simulations we have performed) has shown that the Tree Slotted Aloha protocol () [3] is one of the most effective anti-collision protocols for single reader RFID systems. When looking at the time needed to identify a given number of tags is remarkably faster (between 17 and 35%) than other ALOHA-based protocols and than treebased solutions. This is due to the effectiveness of the collision resolution mechanism adopted by and to the more limited number of collision slots in than in the other protocols. Indeed, different types of slots (idle, collision, identification) have different lengths. Since idle slots are shorter than collision ones it is convenient to develop solutions that try to minimize as much as possible the percentage 316

3 of collision slots. is clearly a first attempt in this direction. Simulations have also outlined that the frame size dimensioning and the tag population estimation are critical aspects which can significantly impact the overall protocol performance. Based on our experimental observations, this paper addresses the problem of achieving accuracy in estimating the number of tags, and proposes an optimization of the based on the new estimation method. Specifically, we propose a Dynamic Tree Slotted Aloha (Dy ) protocol, which adapts frame size and tag estimation based on the amount of tags identified in previous frames. This is possible because divides colliding tags into groups, and generates a new frame for each group. Only tags that transmitted in the same slot belong to the same group, so as to avoid re-collisions (i.e., two tags not colliding in a frame cannot collide in next frames). In, frame sizing for the new set of frames is established once and for all at the end of their parent frame, according to the observed outcomes on it. However, as frames are executed and tags are identified, the knowledge about the amount of discovered tags for each frame can be used to better estimate the number of tags to be identified and properly establish the size of the following frames. This is what is done by Dy. Simulation results show that Dy outperforms all other protocols in all relevant metrics, especially for what concerns the time needed for identifying all tags. Our tag estimation method improves protocol performance, achieving an increase of up to 2% of the system efficiency (in seconds) and consistently reducing the collision activity. The rest of the paper is organized as follows. Sections 2 and 3 detail respectively the and Dy protocols. Section 4 discusses the results of an ns2-based performance evaluation to compare Dy performance with that of the major representatives of the aloha-based and tree-based protocols. 2. TREE SLOTTED ALOHA () The operations of the protocol are based on dividing time in slots. Slots, in turn, are grouped into frames. The size of a frame (i.e., its number of slots) may vary in successive reading cycles, and it is communicated by the reader to the tags at the beginning of each frame in the query message. then randomly select one slot in the current frame: They will answer the reader inquiry with their ID only in that slot. Each slot includes enough time for a tag to answer and for the reader to send an acknowledgment (ACK) or a message of end of slot. At the end of each frame, tags that have been identified get silent. Colliding tags, instead, keep participating in the identification process. According the operations, the interrogation process progresses in new set of child frames. More precisely, a new frame is initiated by the reader for each collision slot in the previous frame. If collisions occur in a child frame, corresponding new child frames are generated, as if visiting the progress tree in a depth-first order. understand whether they collided or not in a frame by keeping the level and the slot in which they transmitted, and checking if the reader initiates a new frame for that slot. If that happens, tags send again their ID in a randomly chosen slot of the new frame. Determining the size of a frame is a crucial part of, with noticeable consequences on its performance. Since the reader does not know the amount of tags that are in its interrogation zone, the initial frame size must be fixed to a predefined value. Previous studies [3] have shown that the choice of the initial frame size does not significantly affect the protocol performance when the number of tags is significantly smaller or higher than the frame size. Based on such studies and on experiments reported in [3] we have set the initial frame size to 128. When the first frame is completed, the knowledge of the slots outcome (i.e., idle, identification, and collision) allows the reader to estimate the number of tags that participated in the reading cycle. This information is used for deciding the size of the following frame. This estimation, based on Chebyshev s inequality, is computed by searching for the number of tags n such that the distance between the triple of estimated values a, a 1, a k and the triple of observed values c, c 1, c k is minimum, as defined by Equation 1 [1]. ɛ(n, c, c 1, c k ) = min a N,n a N,n 1 a N,n k 2 1 c c 1 c k 1 A. (1) Here N denotes the size of the completed frame, a is the expected number of empty slots (idle slots), a 1 is the expected number of slots with one responding tag (identification slots), and a k is the expected number of slots where multiple tags reply (collision slots). The values a,a 1, and a k depend on N and on the number of tags. The estimated number of tags is that value n which minimizes the right side of Equation 1. The minimum is computed over n ranging in [c 1 + 2c k, 2(c 1 + 2c k )]. Given a frame size N and a possible value of n the expected number of slots with r tags can be estimated as! a N,n r = N n «r «n r. (2) r N N The reader then queries only the tags colliding in a given slot in the following frames: One frame for each collision slot. The length of such frames is set based on the estimated number of colliding tags associated with it. More precisely, let n i be the estimated number of transmitting tags in reading cycle i, at level l. The value c i 1 indicates the number of tags actually identified during cycle i, and c i k the number of actual collision slots. Then the expected number of tags transmitting in each of the child frames of i at level l + 1 is given by: n i c i 1 c i k. (3) The size of each child frame is set to such a value. The accuracy of the estimator has significant impact on the performance of the identification process, as an inaccurate estimate of the number of tags may result in non-optimal frame size, with a consequent increase of the identification delay. estimation based on Chebyshev s inequality is observed to be inaccurate when the number of collisions is high (i.e., when c, c 1 ). This is because the estimator does not capture the possibly high variance of the number of tags (from few hundreds to thousands) that can collide in all slots. When this happens, the frame size is not set properly and the performance of the protocol degrades. 317

4 Table 1: Estimation accuracy in the case N = 128, and c, c 1 =, c k = N n vect. distance a a 1 a k DYNAMIC TREE SLOTTED ALOHA is efficient in solving collisions because it generates a new frame for each colliding slot, avoiding any re-mix of colliding tags. However, collision and idle slots still account for more than half of the protocol execution time [3]. This is mostly due to the inaccuracy of the estimation of the number of tags that is used for setting the length of the following frames. In the estimated number of tags n (Equation 1) is searched in the range [c 1 + 2c k, 2(c 1 + 2c k )]. While the lower limit is correct, as at least two tags transmit in a colliding slot, the upper limit is too low to reflect realistic situations comprising a large number of tags. This number is upper bounded by 4 times the initial frame size 128, i.e., 512. As a result, the size of the following frames is set to a value that is too small for networks with a large number of tags. For example, considering a scenario with 5 tags and the initial frame size of of 128 slots it is highly likely that c 1 =, i.e., that no tag is identified in the first frame. In this case, 2(c 1 + 2c k ) = 4c k, n i is bounded by it, and j therefore, k according to (3), the size of the frame at level 1 is 4ck c k = 4, which is largely underestimated as the expected number of tags colliding in a slot is around 4. Therefore, many collisions happen at this level and in the following one, in which the frame size will be around 4 and the number of tags per slot reduces to around 1. Our observation of the unsuitability of the upper limit 2(c 1 +2c k ) for large networks motivates our search for a better upper bound. Our approach starts by improving in case during the first frame no tag is identified (c k = N), which, as mentioned, is highly likely in networks with many tags. Table 1 shows the triple of estimated values a, a 1, a k and their distance from the observed values c, c 1, c k when varying the number of tags n, N = 128 and c, c 1, c k =,, 128. We observe that the bound (512 tags) results in high vector distance. When n 9 the vector distance becomes extremely small, and the estimated number of collision slots truly estimates the c k. Since for higher values of n the vector distance does not improve noticeably (networks with higher number of tags behaves similarly: Every slot is a collision slot), we choose n = 9 in this case. We applied the new upper limit to the protocol and we obtained a slight improvement of protocol performance: System efficiency increases from 4 to 5%. (This variant of will be called unbounded in the following.) However, the improvement is still limited because although the number of estimated tags is higher, there is no difference for networks with more than 9 tags: The estimation of the frame size at the second level is the same for 1 as well as 5 tags. Figure 1(a) shows the mean error for tag estimation obtained by applying the two different limits, where we call bounded the limit used in the basic (i.e., 2(c 1 + 2c k )). The two estimation methods are comparable for less than 5 tags, as the probability to collide on all slots is low and the estimated number of tags does not reach the upper bound. For larger networks the bounded estimation experiences increased error, as it always returns the upper bound 2(c 1 + 2c k ). Instead the unbounded version keeps the error below.1 until n = 9. The error then increases with n. To cope with this issue we propose to estimate the number of tags participating in a frame, exploiting the knowledge gained during previously completed frames. Our improvement is based on the assumption that the allocation of a tag in a slot is completely independent of the behavior of other tags. Therefore, tags are uniformly distributed in available slots, and the number of tags X that fall in a slot is given by the binomial distribution: P{X = r} = n r! «r «n r. (4) N N The expected value of the number of tags in a slot is E[X] = n N. (5) Let us consider a frame of size N, with a population of n tags. We expect to have n/n tags in each slot. At the end of the frame, in case there were colliding slots (i.e., c k > ), then Chebyshev s estimation is used to estimate the number of tags that participated in the frame, and c k new child frames are issued, one for each colliding slot in the parent frame. The participants in each new frame are only the tags that collided in the corresponding slot in the parent frame. The new frames to be executed are estimated to have size as in Equation 3. However, when the first of these sibling frames is completed, the knowledge on the number of tags that were found in such frame can be exploited to refine tags estimation and accordingly adapt the size of sibling frames to be executed. When also the second frame is completed, the knowledge on the first two frames can be useful to refine the size of the remaining frames, and so on. The more frames are completed in a group of sibling frames, the higher the accuracy achieved in estimating the number of tags that are going to participate in the following frames. More in general, let us consider the execution of the ith frame at level l, 1 < i c l 1 k, where c l 1 k represents the number of colliding slots in the parent frame at the previous level. The size of the ith frame S i at level l is estimated as the mean value of the tags that have been identified in the sibling frames previous to i. Specifically, if t j is the number of tags that participated in the frame j at level l, then S i = 1 Xi 1 t j. (6) i 1 j=1 As proceeds in a depth-first order, in case a frame experiences collisions, then they are resolved going down in 318

5 mean estimation error unbounded bounded dynamic tags (a) Error in Chebyshev s estimation for different upper limits. estimated tag number unbounded bounded dynamic observed slots (b) Number of tags estimated for n = 2. Figure 1: Results on tag estimation functions. the tree, and only when all collisions in a frame have been resolved, the next sibling frames at the same level can be executed. This allows not only to exploit knowledge on previous frames, but also to recursively apply the estimation method on deeper levels of the tree, whenever multiple collision slots are present in a frame. Figure 1(a) shows the effectiveness of our estimation method (called dynamic), based on the mean of observed values on previous sibling frames. Tag estimation is very accurate independently of the number of tags, keeping the estimation error always close to zero. The higher accuracy in case of larger networks is due to the fact that a higher number of tags makes higher the probability that they distribute uniformly on all slots. Figure 1(b) confirms the good performance of our novel estimation technique by showing the number of tags estimated by the three methods for a network of 2 nodes. The x-axis represents the number of tags estimated after having resolved each slot of the first frame. We observe that dynamic estimation converges quickly to the actual number of tags, while both the bounded and unbounded versions of are far from it. The latter estimators do not exploit the information made available from the identification process in sibling frames. Therefore, the frame sizes do not change because of the number of identified tags in previous sibling frames. Dy instead exploits this information for refining the size of frames to come. Figure 1(b) shows that, according to Equation 6, Dy quickly converges to the correct value of n. 4. PERFORMANCE EVALUATION In this section we show the results of a thorough ns2- based comparative performance evaluation among our protocol and the major mechanisms proposed so far for single reader identification. We have implemented an RFID extension of the Network Simulator ns2 (v. 2.3) [1] accounting for all the unique features of reader tag communications. Using our extension we have simulated the,,, Dy, QT and EDFSA protocols. In the following discussion we focus on the performance of the first four protocols, as QT and EDFSA always behave worse than and, respectively. 4.1 Metrics Protocols for RFID networks are usually compared by considering the number of bits transmitted during the identification process, and by the number of rounds needed for identifying all tags. A round is defined as the time needed for a reader s request (e.g., query or time slot signaling) and relative tag response. Rounds can be classified depending on whether there is no tag response (idle round), one response (identification round) or there are multiple responses (collision round). In our evaluation, beyond the aforementioned classical metrics we have also measured the time needed to complete the identification process. This is the primary metric of interest for realistic RFID applications. We also revisited traditional metrics such as system efficiency considering the actual time spent in different kinds of slots (i.e., idle, identification and collision slots). The resulting time system efficiency metric is a much more effective way for evaluating the performance of a given protocol as it captures the different impact on performance of different slots. For instance, the time of an idle round is much shorter than that required by a collision round (EPCglobal standard [2]). Specifically, after sending a message to the tags (e.g., a query or a time slot signaling), the reader waits for a tag to respond. In case the reader does not get any response before a reception threshold elapses, the reader realizes that no transmission is coming back from tags (idle round), and issues a new command for the next round. Instead, when one or multiple tags respond, the reader has to wait until the end of tags transmission (lasting for the time needed to transmit tags ID), before issuing a new command. Therefore, idle slots are much shorter than collision and identification ones. In our experiments we focus on the following metrics. Latency. Also called protocol execution time, latency is defined as the time (in seconds) for identifying all tags. System efficiency. This metric indicates the fraction of rounds/time spent by the various protocols identifying tags. In terms of rounds the system efficiency is measured as SE r = R id /R tot, where R id is the amount of identification rounds (which is equal to the number of tags), and R tot is the total number of rounds. In terms of time (time system efficiency), the system efficiency is SE t = T id /T tot, where T id is the time spent in identifying tags, and T tot is the total protocol execution time. Transmitted bits. This metric measures the total amount of bits transmitted during the identification process (both by the reader and by the tags). 319

6 Figure 2: Link timing for reader-to-tag and tag-toreader transmission during a round. 4.2 Transmission time model To perform a realistic evaluation of protocols temporal aspects, it is necessary to define a reference model, that specifies time requirements for reader-to-tag (R => T ) and tag-to-reader (T => R) communications. The transmission model we consider has been derived by the EPCglobal Specification Class-1 Gen-2 [2], that defines the physical and logical requirements for a passive-backscatter, interrogatortalks-first, radio-frequency identification system. From this standard, we drew out two important aspects that must be considered when evaluating transmission time. First of all, physical signaling has to be considered, as both R => T and T => R transmission should begin with a preamble. As R => T preamble is not transmitted at each round, but only in the first issued query, its effect is negligible. Instead, T => R preamble is sent in each tag transmission and hence it has an impact on protocol performance. This preamble depends on data encoding, and in the case of FM, that is usually employed for single-reader scenarios, is 6 bits long. The second aspect to take into account is link timing: when estimating transmission duration, propagation delay, transmission delay and devices reaction time must be considered. Figure 2 shows the link timing for messages exchanged between a reader and a tag. Whenever the reader sends a message to tags, such as a query or a timeslot signaling (in the figure we generically call this message as inventory), the reader transmission arrives to tags after the propagation delay and lasts T X I time. T X I depends on the transmission datarate and on the amount of bits to be transmitted. When the tags stop receiving the reader transmission, they need a R1 time to react and send back an answer. R1 has been set to 1 T 1 (where T 1 = 1/datarate) according to EPC specification. Again a propagation delay is needed before the reader starts receiving tag responses, that will last T X R time. After receiving tag responses, the reader needs a reaction time R2 before being able to issue a new command (new query or timeslot signaling). In case the reader does not get any response (idle round), the reader realizes that no transmission is coming back from tags after a RX threshold that is the time at which the reader should receive the first bit of tag transmissions. This means that in case of no response by the tags, the round ends when the RX threshold elapses, and the reader issues a new command after a reaction time R2. We applied this model to all simulated protocols, so as to have a common time reference that depends only on datarate, and devices physical characteristics. The model highlights that the duration of a round strictly depends on the amount of bits transmitted by the reader (in the inventory message) and by the tag (in the response message). As a consequence, the duration of an idle round (that does not involve any tag response) is shorter than the time taken by an identification or collision round, that involves the transmission of tag s ID. These aspects play an important role when comparing protocols from a temporal point of view. It is also clear that different protocols have different lengths of the inventory and response messages, resulting in different times to complete a R => T and T => R communication. 4.3 Scenarios We consider an RFID system where a single reader with transmission range equal to 1m communicate with n = 1,..., 5 tags. The channel data rate is 4 Kbps. Reader tags communication occurs at a frequency of 866Mhz as specified by the EPCglobal standard [2]. Tag IDs have a length of 96 bits (which is the most commonly used ID length). Longer IDs (e.g., 256bits) do not affect protocol behavior, except for an increase in protocol latency and in the amount of transmitted bits. The initial frame size for aloha-based protocols is set to 128 slots. In terms of distribution of tag IDs we have considered tags with uniformly distributed IDs. Results have been obtained by averaging over 1 runs. 4.4 Simulation Results Results on time and rounds related metrics are shown in figures from 3(a) to 3(d) for both dense (n = 1,..., 1) and very dense (n = 1,..., 5) networks. Results highlight the superiority of slotted aloha protocols (i.e., and Dy ) compared to tree-based protocols ( and ). In terms of system efficiency (figures 3(a) and 3(b)), tree-based protocols have a similar behavior (around 34% for and 37% for ) that remains constant independently of n. This is due to the deterministic operation of tree based-protocols. The protocol splits a group of colliding tags into two groups of similar cardinality at each step. The protocol behaves similarly, as the IDs are uniformly distributed and hence the identification tree is similar to the tree. The number of deployed tags has instead a greater impact on the performance of aloha-based anti-collision schemes. Dy is comparable to when n 7, while for larger networks Dy significantly outperforms, increasing its system efficiency with n. This depends on the fact that when the number of tags is very high the Dy assumption that tags are uniformly partitioned into different slots holds more tightly. The Dy advantage over grows considerably if we switch our attention from the system efficiency, which is an indirect measure of the effectiveness of the identification process, to the time system efficiency, which instead directly reflects performance experienced by the user (see figures 3(c) and 3(d)). When focusing on the protocol execution time, Dy is consistently more efficient than any other protocol, for all considered values of n. For networks composed of thousands of nodes, the improvement achieved by our protocol is 2% with respect to the performance. The difference in protocol performance when considering round and time-based system efficiency depends on the fact that round-based system efficiency does not account for the different time a protocol needs to complete idle or collision and identification rounds. Insights on the reasons behind protocols different execution times is provided by results shown in 32

7 Dy_ Dy_ System Efficiency System Efficiency (a) System efficiency: 1-1 nodes (b) System efficiency: 1-5 nodes. Time System Efficiency Dy_ Time System Efficiency Dy_ (c) Time system efficiency: 1-1 nodes (d) Time system efficiency: 1-5 nodes. Figure 3: Time/rounds related metrics Dy_ coll idle 6 5 Dy_ coll idle Execution Time (s) Execution Time (s) (a) Execution time: 1-1 nodes (b) Execution time: 1-5 nodes. Figure 4: Protocol latencies. Figures 4(a) and 4(b). The bars in the figure show the execution time of each protocol, displaying with different colors the time spent performing idle rounds (lower white part of the bar), the time spent performing collision rounds (gray mid section of the bar) and finally the time spent for identification rounds (upper section of the bar, colored with different colors depending on the protocol). Idle rounds have a negligible impact on overall protocol execution time, ranging from 2% to 4% of protocol execution time for all protocols. This aspect has a twofold reason: collision rounds tend to occur more frequently than idle rounds, and idle rounds have a shorter duration. Aloha-based protocols experience much less collisions than tree-based protocols. Dy collision percentage is always below 4%, and decreases by incrementing the number of tags. At n = 5 Dy collision percentage is 24.4% lower than s and 35.8% (4%) lower than (). The reduced time spent by Dy to solve collisions results in a remarkable decrease in terms of overall execution time. When 5 tags are queried the time needed by Dy to identify them is 2% shorter than the time needed by. The difference increases compared to and. For the same number of tags the execution time is 49.8% (61.17%) higher in () than in Dy. A higher or lower number of collision translates into higher 321

8 Transmitted bytes Dy_ Transmitted bytes Dy_ (a) Transmitted bits: 1-1 nodes (b) Transmitted bits: 1-5 nodes. Figure 5: Transmitted bits. or lower bit complexity. This is clearly shown in figures 5(a) and 5(b)), which display the number of transmitted bytes during the protocol execution for the four different protocols. Dy is the protocol that involves the minimum amount of bytes to perform tag identification. On the contrary, tree-based protocols require a significantly higher number of transmitted bytes due to their high number of collisions (collisions in such protocols account for almost 6% of protocol execution time). 5. CONCLUSIONS We have proposed and evaluated the performance of the Dynamic Tree Slotted Aloha protocol (Dy ) for a variable frame identification of RFID tags. We started by observing that one of the best performing protocol for tag identification, the protocol presented in [3], suffers performance-degrading limitations in networks with a high number of tags. These limitations are the consequence of the method used by for the estimation of the number of tags. Dy copes with this issue by exploiting the knowledge dynamically acquired during protocol execution, and refining the estimation as more and more tags are identified. Time system efficiency induced by Dy improves up to 2% with respect to, given an up to 3% reduction in collisions. As a consequence, other relevant metrics, such as the amount of transmitted bits, are also improved. A detailed comparative performance evaluation of Dy with several previously proposed anti-collision protocols shows the clear advantage of our method for tag identification. Overall, with Dy we have shown that being able of dynamically refining the length of successive frames is beneficial for metrics that are important for realistic RFID applications. [3] M. Bonuccelli, F. Lonetti, and F. Martelli. Instant collision resolution for tag identification in RFID networks. Elsevier, Ad Hoc Networks, (5): , 27. [4] J. Capetanakis. Tree algorithms for packet broadcast channels. IEEE Transactions on Information Theory, 25(5):55 515, [5] C. Law, K. Lee, and K.-Y. Siu. Efficient memoryless protocol for tag identification (extended abstract). In DIALM : Proceedings of the 4th international workshop on Discrete algorithms and methods for mobile computing and communications, pages 75 84, New York, NY, USA, 2. ACM. [6] S.-R. Lee, S.-D. Joo, and C.-W. Lee. An enhanced dynamic framed slotted aloha algorithm for rfid tag identification. In MOBIQUITOUS 5: Proceedings of the The Second Annual International Conference on Mobile and Ubiquitous Systems: Networking and Services, pages , Washington, DC, USA, 25. IEEE Computer Society. [7] J. Massey. Collision resolution algorithms and random-access communication. Multi-Users Communication Networks, CISM Courses and Lectures, (256):73 137, [8] J. Myung and W. Lee. Adaptive binary splitting: a rfid tag collision arbitration protocol for tag identification. Mob. Netw. Appl., 11(5): , 26. [9] L. G. Roberts. Aloha packet system with and without slots and capture. SIGCOMM Comput. Commun. Rev., 5(2):28 42, [1] H. Vogt. Efficient object identification with passive rfid tags. In Pervasive 2: Proceedings of the First International Conference on Pervasive Computing, pages , London, UK, 22. Springer-Verlag. 6. REFERENCES [1] The Network Simulator - ns-2. [2] EPCTM Radio-Frequency Identification Protocols Class-1 Generation-2 UHF RFID Protocol for Communications at 86MHz-96MHz, EPCglobal. Dec