Trade-off Between Memory and Time

Based on the results, we observed some trade-offs between space and speed. Among three data structures, Hash method has the best performance in exact string search.

But a simple hash table does not contain sufficient information to perform substring or super string match. If the given sensor data forwarding system has to process prefix suffix or substring matching, other mechanisms should be imported. TST performs averagely in the middle of the three algorithms in each test but have long search time.

ST has long insertion time and occupies large memory space. On the other hand, it also has a short exact string search time and an even better search time for non-match cases. The weakness of ST is definitely blamed to the large memory space. For small word sets, ST, TST and Hash are similar in memory size, but with the word set grows, the space requirement of ST dramatically goes up. The memory efficient version of ST can reduce 20%-24% memory space and still keeps the search time in our experiments when the training word set contains more than 10000 strings, but the reduced memory size still large compared to TST and Hash.

However, when we move our test environment to sensor-based platforms, the search time rank inverses. Due to the instruction set dependency, modular-based instructions takes much longer time on sensor-based platforms. To have better performance, hash-based matching method should carefully evaluate the properties of the given platform before choosing hash functions. In our experiences, rotation-based hash function has the least conflict in average cases, but it takes almost two times longer to finish a search than TST.

5.8. CONCLUSION

We study and evaluate the use of ST for efficient predicate matching for data-centric sensor networks. With the proposed memory optimization scheme, we are able to implement ST on a sensor network development platform. The experimental results

show that ST performs significantly better than TST and hash which suggest ST being the most promising solution for fast sensor data forwarding.

六、分析與討論之三 — Sensor Network for Everyday Use: The BL-Live Testbed Experience

6.1. Introduction

There have been avid research activities on sensor networks worldwide. Research labs, such as Berkeley WEBS and UCLA CENS, have initiated research projects and

related hardware/software platform development early in 1999-2000. NSF of the United States started to call for research proposals in the area of Sensors and Sensor Networks in 2003. Early vision on the military use of the sensor networks also prompted the establishment of sensor and sensor network related programs in DARPA.

It is now 2006. After over 7 years of R&D in sensors and sensor networks, we see now sporadic reports of sensor networks for short-term experimental purpose [72][73][74]. There lacks still any long-term, everyday-use deployment. We wonder why. Is the deployment of sensor networks too difficult or practically impossible?

Motivated to address this question, we deploy a 30+ node wireless sensor network in a university campus building, the Barry Lam (BL) Hall of Electrical Engineering at the National Taiwan University main campus.

This project is referred to as BL-Live for that the seemingly cold concrete BL Hall is transformed into a lively smart office building. The main objective is to obtain practical experience and to discover problems that otherwise will be difficult to observe in small-scale test-beds or in simulations.

The sensor network in the BL Hall facilitates two everyday services: 1) Elevator Report and 2) Smart Office. The first service reports the status of the slow-paced elevators located on two opposite sides of the building. This service allows the building residents to select an elevator that will arrive earlier at the floor they desire.

The second service detects the presence meeting participants in an office. This allows automated control of the camera to broadcast publicly the progress of a meeting in the office whose door is better off kept closed to conserve energy in the summer. The two services, although casual, address real needs of the BL Hall residents. More

importantly, the two services function as the applications that drive the use of the sensor network infrastructure.

To support the Elevator Report and Smart Office services, there involve three major system and network components, hardware, sensor networking, and sensor signal

processing. Hardware is the most fundamental element. We have successfully cloned and manufactured 40 pieces of an ultra low-power wireless

sensor node named Telos [75] from scratch. Most of these Telos clones are deployed in the building to form the sensor network infrastructure for the Elevator Report service. The rest are deployed in a number of offices and carried by the volunteers for the Smart Office service.

Each node on the sensor network infrastructure runs the Magnetic Diffusion [76], a routing protocol that enables the collection of the sensor data from the elevators to the data sink in one of the student laboratories in BL Hall. The nodes deployed in the offices runs a simple protocol that periodically probes for the existence of any sensor nodes alive within the radio range. Each node carried by a volunteer answers the probes from the office nodes. This enables the detection of the volunteer’s presence in the offices.

The sensor nodes placed in the elevators are integrated with accelerometers. The accelerometers sense the acceleration and breaking patterns of the elevators. From these motion data, we derive the status of the elevators on-node. These status reports are then relayed by the sensor network to the data sink. The data sink is connected to a Web server and the statuses of the elevators are made available as a public Web page.

The office sensor nodes deliver the presence information via Internet to the Web server and the presences of the volunteers are made available as well as a public Web page.

After more than 9 months of experience building, deploying and operating such a wireless sensor network, we come to a number of observations and suggestions on the cost, deployment, placement, data delivery, energy efficiency, and usability issues towards practical, commercialize-able sensor networks in the future.