Validated cost models for sensor network queries

Abstract

Generating a good execution plan for a declarative query has long been a central problem in data management research. With the rise in interest in wireless sensor networks (WSNs) as query processing platforms, it was quickly noticed that the corresponding optimization problem is even more challenging than the classical one, since, in comparison to classical platforms, a WSN is a very constrained computational infrastructure (in terms of memory, processing, and communication capabilities, and, crucially, depletable energy stocks). Optimizing a declarative query for execution in WSNs is thereby made both more important and more challenging. One of the requirements for effective query optimization is the availability of effective models for estimating the cost of alternative execution plans. This paper describes how query cost models for space, time and energy were methodically derived and validated for an expressive algebra for continuous queries over sensor streams.

Resilient sensor network query processing using logical overlays

“…The compilation/optimization process takes as input a SNEEql query (as exemplified in Fig. 1), QoS expectations (not shown in the figure) in the form of a desired acquisition rate (i.e., the frequency at which sensing takes place) and a maximum delivery time (i.e., an upper bound on the acceptable amount of time between data being acquired and being reflected in the emitted results), and the following kinds of metadata: (1) the current connectivity graph, which describes the (cost-assigned) communication edges in the WSN; (2) the logical schema for the query, which describes the available logical extents over the sensing modalities in the WSN; (3) the physical schema for the query, which describes which physical nodes contribute data to which logical extent, and which node acts as base station; (4) statistics about nodes (e.g., available memory and energy stocks); (5) cost-model parameters (e.g., unit costs for sleeping, sensing, processing, and communicating) [1]. The query takes two streams, one stemming from sensors in a field, the other from sensors in a forest.…”

“…By being governed by an agenda, a SNEE QEP implements a simple form of TDMA. Whilst this is often economical provided that the estimation models are accurate (and [1] shows that they are), any changes to the timing of the operators or transmissions requires the agenda to be recomputed and hence the QEP to be recompiled and propagated into the WSN.…”

“…We note that, as described in [1], SNEE makes intensive use of the empirically-validated analytical cost models for energy, memory and time expenditure computed over the RT, DAF and agenda for a query. For example, in SNEE, we can estimate the energy and time cost of running a QEP fragment instance on any given node per agenda execution cycle.…”

“…This is done by computing the energy drain on each node resulting from the execution of one agenda cycle using the energy cost model in [1]. Given the statistics about current energy levels on each node, we can estimate which node will fail first due to energy depletion.…”

“…We assume that all the metadata has been collected and are accurate and, therefore, we exclude the collection time from our measurements. We assume the SNEE energy and time cost estimation models are accurate as shown in [1]. We assume that our cardinality estimation model (which was thoroughly experimentally validated using the Avrora [12] emulator) is also accurate.…”