2.4 Green Networking
2.4.3 Sleep Mode Operation Algorithms
The objective of the thesis is to exploit appropriate traffic information for sleep mode oper- ation. Thus, different sleep mode operation algorithms are reviewed and classified.
For a sleep mode operation algorithm, it can be manually configured as a static scheme (or called an offline algorithm) for a statistical traffic cycle based on pre-measured information. This type of algorithms such as [50, 51] performs well when the traffic pattern is relatively static and is accurately captured and modelled. They may be applied with different variations at different locations of a network, where the superiority of low complexity and low processing overhead should be guaranteed. Otherwise, it is better to realise the sleep mode operation in an adaptive way based on some monitored information related to the required metrics because of unusual events and unpredicted traffic variations that are especially common in SCNs. This kind of algorithms refers to online algorithms, which can be further classified into slow reaction algorithms and fast reaction algorithms depending on the frequency of making network adjustments [23, 48]. For BSs with a large wake-up time or large wake-up energy consumption, long-term measurements can be performed, with a statistical traffic perception as a result. The long-term information can be taken as the decision making material when controlling the sleep mode operation to reduce the frequency of state transition. This complies with slow reaction algorithms, e.g. [52, 53]. Conversely, a network can track the real-time traffic and react more timely to the sudden changes if the system hardware can support a swifter and more energy-efficient wake-up without letting users perceive the adjustments. In this case, fast reaction algorithms like [54, 55] can be applied. As the number of users in a conventional large macrocell is usually big, which yields statistical characteristics, the traffic volumes in different periods usually conform to a long-term distribution. For this scenario, static or slow reaction sleep mode operation schemes may be enough to respond to the traffic variations. While in SCNs, fast varying traffic volumes in each cell can be observed due to the significant reduction in the number of users in each cell, which results in a much more unbalanced traffic distribution among the small cells. This engenders the need of fast reaction algorithms to control BS state transitions, which is expected to be allowed by light-weight small cell BSs in SCNs.
In terms of the locations of decision making and control engines, sleep mode operation schemes can be categorised into centralised, distributed and partially distributed schemes. In a centralised scheme like [56,57], a central network controller is responsible to collect informa- tion measured at other network elements, i.e. BSs or users. The information is analysed and
46 Chapter 2. Literature Review
decisions of BS state transitions are made by the central controller. The centralised architec- ture enables centralised algorithms, which optimise networks from a global view. However, the information collection may yield more system overheads in backhaul links. In contrast, distributed algorithms such as [54, 55] are suitable for a flat architecture, where local infor- mation from users or other BSs is collected by each BS. BSs then reason the local traffic conditions and make decisions. Although this approach loses the global view and may under- optimise the BS state transitions, it has the merits of fewer overheads and lower latency. In partially distributed schemes such as [57,58], some of the BSs in a network are given priorities to be active and control the rest of the BSs falling into their respective coverage areas. With a better view of the network conditions, the priority BSs can therefore make more reliable decisions. The first two types are usually used to describe a single-tier sleep mode operation algorithm while the partially distributed schemes can also be implemented in a HetSNet with a hierarchical architecture. The lower tier BSs with larger coverage areas can play as priority BSs, which control the state transitions of higher tier smaller BSs. Beyond these types of schemes, operators can also cooperate with each other to share the infrastructures and save energy cost.
From another point of view, sleep mode operation algorithms can also be classified de- pending on what kind of network elements control the operation. This classification focuses on the types of decision making engines and the functionality required by a sleep mode, which directly links to the actual power consumption of the sleep mode. The state transitions of a BS can be controlled by itself, meaning that it requires some processing hardware modules and RF modules to be active for the purposes of measuring even when they are at sleep modes, collecting and processing information [59]. These require more BS components to operate as usual or close to full power. BS state transitions can also be determined by other active BSs. This method only needs sleep modes to have a limited processing capability and lower power consumption. These kinds of sleep modes just have to be able to receive and recognise activa- tion messages from other active BSs for wake-up [60]. Similarly to this, state transitions may be controlled by users. The difference is that the computation responsibility is distributed among users [61]. For a centralised architecture, a decision making engine can be located at a centralised site, where information is collected from the governed BSs which take the mea- surements. A sleep mode in this case requires fewer functions available and fewer hardware modules operating because the sleeping BSs can be controlled by the centralised controller via backhaul links. The exception to centralised architectures are the ones involving wireless
Chapter 2. Literature Review 47
backhaul links. This case requires some hardware modules to be active similarly as the case where the state transitions of a BS are controlled by other network elements. For different BS types, appropriate strategies can be chosen. In a more comprehensive network architecture with different network conditions (e.g. various BS types, centralisation/decentralisation), mixed strategies can be taken to accommodate different requirements.
In this fast developing field, some sleep mode operation algorithms have been proposed. Some example algorithms are cited when introducing the taxonomy. A more comprehensive summary about the existing on-demand sleep mode operation algorithms is included in [48].