Adaptive Real-Time Query Scheduling for Wireless Sensor Networks

Moutaz Saleh Mustafa Saleh Department of Computer Science & Engineering

College of Engineering, Qatar University 2713 Doha, Qatar

moutaz.saleh@qu.edu.qa

ABSTRACT With the recent evolution of high data rates in real-time sensor network applications, there is always an increasing demand for high performance query services in such networks. To meet this demand, we propose an adaptive approach for scheduling real-time query transmissions in wireless sensor networks (WSNs). Our purpose is to schedule multiple real-time queries efficiently by maximizing the overall network throughput while meeting the queries’ deadlines. The proposed scheduler allows multiple conflict-free queries to execute concurrently in order to achieve maximum throughput, besides the scheduler works preemptively to satisfy the real-time queries’ deadline constrains. The simulation results showed that our proposed scheduler can be adopted to effectively increase network throughput and eliminate query prioritization inversion.

Categories and Subject Descriptors C.2.1 [Computer-Communication Networks]: Network Architecture and Design – wireless communication, network communications.

General Terms Algorithm, Management, Performance.

Keywords Real-time Query; Scheduling; Preemptive; Priority; Throughput; Conflict-free; WSN.

1. INTRODUCTION Recent years have witnessed a necessary need towards wireless sensor networks (WSNs) that must support real-time data collection at high data rates. Examples of such systems include, patient monitoring [1], emergency response [2], and health monitoring [3]. These systems, however, pose significant challenges. First, the system must handle various types of traffic with different deadlines. Thus, a real-time communication protocol should provide effective prioritization between different

traffic classes while meeting their respective deadlines. Second, the system must support high throughput since it may generate high volumes of traffic. Therefore, it is important for the system to achieve predictable and bounded end-to-end latencies.

Many WSN’s real-time applications use query services to periodically collect data from sensors to a base station. Real-time applications refer to those performance critical applications that require bounded service latency. There exists a class of WSN applications that is real-time, and requires bounded latency on data delivery. Since the nodes communicate over a shared medium, it is possible that if multiple transmissions overlap in time, some of them will collide. This problem is alleviated by either sensing the medium for possible ongoing transmissions to avoid collisions, or by carefully timing and scheduling the transmissions so that no collision occurs. The collision avoidance algorithms that attempt to schedule transmissions are referred to as transmission scheduling algorithms. While contention based collision avoiding algorithms offer a simpler and more dynamic solution to the medium access problem, schedule based algorithms can provide deterministic service delay bounds that is crucial for real-time applications where the transmission delays should be known and bounded.

In this paper, we present a novel Real-Time Query Scheduling (RTQS) approach designed to meet the communication needs of high data rate real-time WSN applications. Our scheduling approach takes advantage of the common properties of WSN query services to construct a concurrent scheduler for executing conflict-free query transmissions. This can maximize the overall system throughput. Also, the scheduler works adaptively with priority-driven preemption to satisfy the real-time queries’ deadline constrains. Specifically; the proposed scheduling scheme has the following attractive features: first, it achieves high query throughput by allowing concurrent execution of conflict-free query transmissions. Second, it eliminates query prioritization inversion introduced by traditional non-preemptive query scheduling. Third, it enhances the query worst-case response time of the ordinary preemptive scheduling. Finally, the associated schedulability analysis bridges the gap between the WSN queries and real-time scheduling theories.

The rest of this paper is organized as follows: Section 2 introduces related works on WSNs query scheduling and then Section 3 describes query, network, and scheduler models. Section 4 details the design and analysis of the proposed scheduling scheme. After exhibiting simulation results in Section 5, Section 6 concludes this paper.

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, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MSWiM’11, October 31–November 4, 2011, Miami, Florida, USA. Copyright 2011 ACM 978-1-4503-0898-4/11/10...$10.00.

235

2. RELATED WORK In-network Query processing techniques have been widely accepted for on-line sensory data management in WSNs [4] [5]. Instead of pulling all the sensory data from nodes into a central server to process, these techniques make the sensor nodes to cooperatively process the queries from users within the network, hence can increase the flexibility and reduce the network traffic of WSNs. Scheduling schemes have also been proposed to guarantee the efficiency and reliability of in-network sensor query processing [6] [7] [8] [9] [10]. These scheduling schemes are able to coordinate the timings of nodes to avoid communication collisions and make receivers active when a transmitter transmits packets towards them. Figure 1 shows the operation of query processing in WSN.

For instance, TDMA scheduling is attractive for high data rate sensor networks because it is energy efficient and can provide higher throughput than CSMA/CA protocols under heavy load. Two types of TDMA scheduling problems have been investigated in the literature: node scheduling and link scheduling. In node scheduling, the scheduler assigns slots to nodes whereas, in link scheduling, the scheduler assigns slots to links through which pairs of nodes communicate. Early TDMA scheduling protocols were designed for static or uniform workloads [11][12][13][14]. Such approaches are not suitable for dynamic applications with variable and non-uniform workloads. Several recent TDMA protocols can adapt to changes in workload. A common method to handle variable workloads is to have nodes periodically exchange traffic statistics and then adjust the TDMA schedule based on the observed workload [11][15][16]. However, exchanging traffic statistics frequently may introduce communication overhead.

In [17], a novel query scheduling technique called DCQS is presented. Instead of assigning slots to each node or link, slots are assigned to transmissions based on specific communication patterns and temporal properties of queries in WSNs. This approach allows achieving high throughput and low latency. Also, DCQS adapts to changes in workloads by exploiting explicit query information provided by the query service. Furthermore, it features a local scheduling algorithm that can accommodate changes in query rates and completions without explicitly reconstructing the schedule. In spite of all these advantages for DCQS, it ignores the quality requirements from different applications and users queries with different quality preferences are equally treated and are processed one by one

according to their arrival order. As a result, those queries with high quality requirements, real-time queries, are likely to be unsatisfied, whereas the lower request queries that arrive earlier than the high request queries may be overly satisfied. Hence, for real-time query scheduling, the research work in [18] proposes a Slack-stealing Query Scheduling (SQS) algorithm which adapts preemption decisions to improve throughput while still meeting real-time queries’ deadlines.

3. SYSTEM MODELS

3.1 Query Model RTQS assumes a common query model in which source nodes produce data reports periodically. This model fits many applications that gather data from the environment at user specified rates. A query l is characterized by the following parameters: a set of sources, a function for network aggregation [19], the start time Sl, a query period Pl, a query deadline Dl, and a static priority. A new query instance is released in the beginning of each period to gather data from the WSN. We use Il,u to refer to the uth instance of query l whose release time is Rl,u = Sl + u · Pl. The priority of an instance is given by the priority of its query. If two instances have the same query priority, the instance with the earliest release time has higher priority. For each query instance a node i need Wl[i] slots to transmit its aggregated data report to its parent.

A query service works as follows: a user issues a query to a sensor network through a base station, which disseminates the query parameters to all nodes. To facilitate data aggregation, the query service constructs a routing tree rooted at the base station as the query is disseminated. The execution of a query instance entails network data aggregation. Accordingly, each non-leaf node waits to receive the data reports from its children, produces a new data report by aggregating its data with the children’s data reports, and then sends it to its parent. We assume that there is a single routing tree that spans all nodes and it is used to execute all queries. This assumption is consistent with the approach adopted by existing query services [19]. During the lifetime of the application the user may issue new queries, remove queries from execution, or change the parameters of existing queries. RTQS is designed to support dynamic queries efficiently.

3.2 Network Model RTQS models a WSN as an Interference-Communication (IC) graph as shown in Figure 2. The IC graph has all nodes as vertices and has two types of directed edges: communication and interference edges. A communication edge ab indicates that a packet transmitted by a may be received by b. A subset of the communication edges forms the routing tree used for data aggregation. An interference edge ab indicates that a’s transmission interferes with any transmission intended for b even though a’s transmission may not be correctly received by b. The IC graph is used to determine if two transmissions can be scheduled concurrently. We say that two transmissions, ab and cd are conflict-free and can be scheduled concurrently if a, b, c, and d are distinct and ad and cb are not communication/interference edges. Each node needs to know only its incoming/outgoing communication and interference edges. To do so, Zhou et al in [20], present a Radio Interference Detection (RID) mechanism that can be used by WSN nodes in IC graph to determine its adjacent communication and interference edges. Also, clocks are

Figure 1. Query Processing in WSN

236

assumed to be synchronized as clock synchronization is a fundamental service in WSN since many applications must time-stamp their sensor readings to infer meaningful information about the observed events [2].

4. QUERY SCHEDULING RTQS is designed to achieve high data throughput and differentiated query latencies through prioritized concurrent conflict-free transmission scheduling. The adopted approach relies on using two components: a planner and a scheduler. The planner constructs a plan for executing all the instances of a query. A plan is an ordered sequence of steps, each comprised of a set of conflict-free transmissions. The scheduler that runs on every node determines the time slot in which each step in a plan is executed. To improve the throughput, the scheduler may execute steps from multiple query instances in the same slot as long as they do not conflict with each other. To summarize, RTQS works as follows: (1) When a query is submitted, RTQS identifies a plan for its execution. Multiple queries be executed using the same plan, and therefore RTQS may reuse a previously constructed plan for the new query. When no plan is reused, the planner constructs a new one. (2) RTQS determines if a query meets its deadline using a schedulability analysis which is performed on the base station. If the query is schedulable, the parameters of the query are disseminated; otherwise, the query is rejected. (3) At run-time the scheduler executes all admitted queries.

A plan has two properties: First, it respects the precedence constraints introduced by data aggregation: a node is assigned to transmit in a later step than any of its children. Second, each node is assigned in sufficient steps to transmit its entire data report. We use Tl[i] to denote the set of transmissions assigned to step i (0 ≤ i < Ll) in the plan of query l, where Ll is the length of the plan. To facilitate in-network aggregation, a node waits to receive the data reports from all its children before transmitting the aggregated data report to its parent. Therefore, to reduce the query latency, the planner assigns the transmissions of a node with a larger depth in the routing tree to an earlier step in the plan. This strategy reduces the query latency because it reduces the time a node waits for the data reports from all its children. We also assume that two queries belong to the same query class if they can be executed according to the same plan. This can be defined as follows: Let Wl[i] be the number of slots node i needs to transmit its data report to its parent for an instance of query l. If the planner constructs a plan for a query l, the same plan can be reused to execute a query h if Wl[i] = Wh[i] for all nodes i. Since queries with different temporal properties and aggregation functions may share a same plan, a WSN may only need to support a small number of query classes. This allows RTQS to amortize the cost

of constructing a query plan over many queries and effectively reduces the overhead.

The scheduler performs a query instance according to its query plan and improves the query throughput by overlapping the transmissions of multiple instances such that: (1) all steps executed in a slot are conflict-free. (2) the steps of each plan are executed in order: if step Il,u.i is executed in slot si, step Il,u.j is executed in slot sj < si then Il,u.j < Il,u.i. This ensures that the precedence constraints required by aggregation are preserved. Additionally, the scheduler maintains a record of the start time, period, and priority of all admitted queries. This can help in determining the step numbers in which the host node is assigned to transmit or receive in each plan together with the plan’s length. The scheduler considers the released instances in the order of their priority and executes all steps that don’t conflict in given slot s. For high performance, this operation should be done with low time complexity. Accordingly, the concept of Minimum Step Distance (MSD) is used [6]. The MSD ∆(l, h) is the smallest step distance between Il,u and Ih,v such that the two steps Il,u.i and Ih,v.j may be executed concurrently without any conflict. The MSD ∆(l, h) depends on the IC graph and the plans of l and h.

Here, we consider two common RTQS algorithms: Non-preemptive Query Scheduling (NQS) and Preemptive Query Scheduling (PQS). In NQS, once a query instance is started then it can’t be preempted. For that, the earliest time at which an instance Il,u may start is after the previous instance Ih,v completes step ∆-1. Since the execution of Il,u and Ih,v cannot be preempted, if we enforce the MSD between the start of the two instances then their concurrent execution is conflict-free for their remaining steps since steps Il,u.i=x and Ih,v.j=x +∆ are executed in the same slot. Therefore, to guarantee that a non-preemptive scheduler executes conflict-free transmissions in each slot, it suffices to enforce a MSD of ∆ between the start times of any two instances. Indeed, NQS enforces a MSD of at least ∆ between the start times of any two instances by starting an instance in two cases: (1) when there is no instance being executed and (2) when the step distance between the two instances is larger ∆.

Obviously, NQS has the drawback of introducing priority inversion. To eliminate such drawback, PQS is introduced with a key feature which preempts the instances that conflict with the execution of a higher priority instance. PQS maintains a run queue and a release queue which are keyed by the query instance priority. When a new instance is released, it is added to the release queue. On the other hand, when an instance is started for execution, it is added to the run queue. PQS starts an instance Il,u (Il,u € release queue) in two cases: (1) If the next step Il,u.i can be executed concurrently with all instances in the run queue without conflict i.e. MSD of ∆ between the start times of any two instances. (2) If Il,u has higher priority than all the instances in the run queue. In the last case, all running instances that conflict the execution Il,u will be preempted. We mention here that when an instance is started it is moved from the release queue to the run queue. Oppositely, when an instance is preempted, it is moved from the run queue to the release queue. Now, by eliminating priority inversion, PQS achieves lower latencies for high priority instances when compared with NQS. However, the query throughput is lower because it allows less overlap in the execution of instances. This exemplifies the tradeoff between prioritization and throughput in query scheduling. In the next section, we will introduce our proposed scheduling algorithm to effectively combine this tradeoff.

Figure 2. Example of IC Graph and its Plan

237

5. QUERY SCHEDULING Our proposed scheduler combines the benefits of NQS and PQS in that it improves query throughput while meeting all deadlines. The design of this scheduler is based on the observation that preemption lowers throughput, and hence it should be used only when necessary for meeting deadlines. We define Cl,u as the number of slots a lower priority instance u of query l has completed during its execution. Besides, a higher priority instance v of query h may start execution in any slot in the interval [rh,v , rl,u+∆], where rh,v is the release time of Ih,v (the vth instance of query h), rl,u is the release time of Il,u (the uth instance of query l), and ∆ is the MSD between Il,u and Ih,v. The scheduler postpone the start of the higher priority instance Ih,v if the lower priority instance Il,u satisfies the condition of having (∆ - Cl,u + Ph,v) ≤ D h,v , where Ph,v is the execution period of the higher priority instance Ih,v, and Dh,v is the maximum allowable deadline for the same higher priority instance Ih,v. The advantage for this approach is that it opportunistically avoids preemption and the related throughput reduction when allowed by query deadlines. This highlights that our proposed adaptive query scheduler (AQS) can adapt preemption decisions to improve throughput while meeting all queries’ deadlines. To clearly show the efficiency of the proposed scheduler, we present here the worst-case response time analysis for NQS, PQS, and proposed AQS. The worst-case response time of a query is the maximum query latency of any of its instances.

5.1 Analysis of NQS Since NQS is a non-preemptive scheduling algorithm, to compute the response time of a query l we must compute the worst-case interference of higher priority instances and the maximum blocking time of l due to the non-preemptive execution of lower priority instances. From the NQS operation, an instance is blocked for at most ∆−1 slots, and a higher priority instance interferes with a lower priority instance for at most ∆ slots. Accordingly, the response time Rl of query l is the sum of its plan’s length L and the maximum delay Wl that any query instance experiences before it is started.

5.2 Analysis of PQS In PQS, higher priority instance cannot be blocked by a lower priority instance. Also, after an instance completes ∆ steps, no newly released instance will interfere with its execution because their step distance would be at least ∆ which allowing them to execute concurrently. To compute the response time of a query l we split its instance Il into two parts: a preemptable part of length ∆ and non-preemptable part of length L − ∆. Higher priority instances may interfere with Il only during its preemptable part. Thus, the response time of a query l is the sum of response time of the preemptable part and the length of the non-preemptable part.

5.3 Analysis of AQS To compute the worst-case response time of a query l we split again a query instance Il into two parts: a preemptable part and a non-preemptable part. Under PQS, the preemptable part is ∆ slots. In contrast, with AQS, an instance Il may utilize from a higher priority instance Ih at least n slots. Thus, the length of the preemptable part is at most ∆ − n slots; and the length of the non-preemptable part is therefore L – (∆ − n) slots which is lower compared to both NQS and PQS. This is for sure a noticeable improvement for the AQS operation.

6. SIMULATION AND RESULTS Our simulation settings are similar to 802.11b radios. This is because we are interested in high data rate applications, and sensor nodes used for such applications often adopt high-bandwidth radios. Indeed, this is also reasonable since several real-world sensor systems use 802.11b interfaces to meet their bandwidth requirements. Accordingly, the network bandwidth is 2Mbps. The communication range is 125m. The power consumed for transmitting and receiving a packet is 1.6W and 1.4W, respectively. The size of a packet is 2KB, of which 20 bytes are used for packet headers. Based on packet size and bandwidth we computed the slot size to be 8.16ms. The queue size is 10 packets. Each experimental run takes 200s. In the beginning of the simulation we construct the IC graph and the routing tree. The IC graph is constructed similarly to the method described in. The routing tree is constructed as follows. The node closest to the center of the topology is selected as the base-station and is the root of the routing tree. The base station initiates the construction of the routing tree by flooding setup requests. A node may receive setup requests from multiple nodes and selects the node with the latest depth as its parent. Each node in the routing tree performs in-network aggregation when executing a query. We assume that each aggregated data report fits in a single packet. The queries issued involve all nodes in the network. In all experiments the queries belong to the same query class. The results presented in this section are the average of multiple runs for accuracy. All experiments are performed in a 750m × 750m area divided into 75m × 75m grids in which a node is placed at random. We simulate three queries with high, medium and low priorities. The query priorities are determined based on their deadlines such that the tighter the deadline, the higher the priority. Eventually, we compare the performance of our proposed AQS algorithm against both NQS and PQS with respect to the response time. The response time of a query instance is the time between its release time and completion time, i.e., when the base station receives the last data report for that instance. The average response times of NQS, PQS, and AQS when scheduling queries with high QH, medium QM and low QL priorities are shown in Figures 3, 4, and 5 respectively. By observing the plotted figures, NQS meets the deadline requirements only when QH deadline is large enough to be scheduled. NQS misses QH deadline when it is tight due to the priority inversion under non-preemptive scheduling. This indicates that NQS is unsuitable for high priority queries with tight deadlines. Interestingly, under AQS, the response time of QH changes depending on its deadline. As the deadline becomes tighter, the response time of QH also decreases and remains below the deadline. We also see an increase in the response times of the lower priority queries as QH deadline is decreased. This is because as QH deadline decreases the lower priority queries may interferes high priority queries. This shows that AQS adapts effectively based on query deadlines. Moreover, note that AQS provides smaller latencies for the lower priority instances than PQS. This is because AQS has a higher throughput than PQS since it uses preemption only when it is necessary for meeting packet deadlines.

238

7. CONCLUSIONS This paper presents an adaptive query scheduler (AQS) approach designed to meet the deadline needs of real-time queries in WSN. The proposed AQS approach allows multiple conflict-free queries to execute concurrently in order to achieve maximum throughput, besides it works preemptively to satisfy the real-time queries’ deadline constrains. This scheduler combines the benefits of NQS and PQS in that it improves query throughput while meeting all deadlines. The design of this scheduler is purely based on the observation that preemption lowers throughput, and hence it should be used only when necessary for meeting deadlines. When comparing AQS with both NQS and PQS, the simulation results showed that AQS can be adopted to effectively increase network throughput and eliminate query prioritization inversion.

8. ACKNOWLEDGMENTS My great thanks to Qatar University in supporting an ideal research environment.

9. REFERENCES [1] Lorincz, K., Malan, D. J., Fulford-Jones, T. R. F., Nawoj, A.,

Clavel, A., Shnayder, V., Mainland, G., Welsh, M., and Moulton, S. 2004.. Sensor networks for emergency response: Challenges and opportunities. IEEE Pervasive Computing, vol. 3(4): pp.16–23.

[2] Mangharam, R., Rowe, A., Suzuki, R., and Rajkumar, R. 2006. Voice over sensor networks. In RTSS 06.

[3] Lynch, J. P. and Loh, K. J. 2006. A Summary Review of Wireless Sensors and Sensor Networks for Structural Health Monitoring. The Shock and Vibration Digest, 38(2):91–128.

[4] Madden, S., Franklin, M. J., Hellerstein, J. M., and Hong, W. 2005. Tinydb: an acquisitional query processing system for sensor networks. ACM Trans. Database Syst., vol. 30(1): pp. 122–173.

[5] Yao, Y. and Gehrke,, J. 2003. Query processing in sensor networks. In proceedings of CIDR.

[6] Cao, Q., Abdelzaher, T. F., He, T., and Stankovic, J. A. 2005. Towards optimal sleep scheduling in sensor networks for rare-event detection. In proceedings IPSN.

[7] Hohlt, B., Doherty, L., and Brewer, E. A., 2004. Flexible power scheduling for sensor networks. In proceedings IPSN.

[8] Sadagopan, G. Lu, N., Krishnamachari, B., and Goel, A. 2005. Delay efficient sleep scheduling in wireless sensor networks. In proceedings of IEEE INFOCOM.

[9] Sichitiu, M. L. 2004. Cross-layer scheduling for power efficiency in wireless sensor networks. In proceedings of IEEE INFOCOM.

[10] Wu, H., Luo, Q., and Xue, W. 2006. Distributed cross-layer scheduling for in-network sensor query processing. In PerCom. IEEE Computer Society, pp. 180–189.

[11] Cidon, I. and Sidi, M. 1989. Distributed assignment algorithms for multihop packet radio networks. IEEE Transactions on Computer., vol. 38, no. 10.

[12] Ephremides, A. and Truong, T. 1990. Scheduling broadcasts in multihop radio networks. IEEE Transactions on Communications, vol. 38, no. 4.

[13] Ramaswami, R. and Parhi, K. K. 1989. Distributed scheduling of broadcasts in a radio network,. In proceedings of IEEE INFOCOM.

[14] Arikan, E. 1983. Some complexity results about packet radio networks. NASA STI/Recon Technical Report, vol. 83.

[15] Rajendran, V., Obraczka, K., and Garcia-Luna-Aceves, J. J. 2003. Energyefficient collision-free medium access control for wireless sensor networks. In proceedings of SenSys.

[16] Bao, L. and Garcia-Luna-Aceves, J. J. 2001. A new approach to channel access scheduling for ad hoc networks. In proceedings of MobiCom.

[17] Chipara, O., Lu C., and Stankovich, J. A. 2006. Dynamic conflict free query scheduling for wireless sensor networks. In proceedings of ICNP.

[18] Chipara, O., Lu, C., and Roman G-C, 2007. Real-time query scheduling for wireless sensor networks. In proceedings of 28th IEEE International Real-Time Systems Symposium.

[19] Madden, S., Franklin, M. J., Hellerstein, J. M., and Hong TAG, W. 2002. A tiny aggregation service for ad-hoc sensor networks. In proceedings of OSDI.

[20] Zhou, G., He, T., Stankovic, J. A., and Abdelzaher, T. F. 2005. RID: radio interference detection in wireless sensor networks. In proceedings of IEEE INFOCOM.

Figure 3. NQS Response Time

Figure 4. PQS Response Time

Figure 5. AQS Response Time

239