Ulf Bodin

Load-tolerant differentiation with active queue management

Current work in the IETF aims at providing service differentiation on the Internet. One proposal is to provide loss differentiation by assigning levels of drop procedence to IP packets. In this paper, we evaluate the active queue management (AQM) mechanisms RED In and Out (RIO) and Weighted RED (WRED) in providing levels of drop precedence under different loads. For low drop precedence traffic, FIO and WRED can be configured to offer sheltering (i.e., low drop precedence traffic is protected from losses caused by higher drop precedence traffic). However, if traffic control fails or is inaccurate, such configurations can cause starvation of traffic at high drop precedence levels. Configuring WRED to instead offer relative differentiation can eliminate the risk of starvation. However, WRED cannot, without reconfiguration, both offer sheltering when low drop precedence traffic is properly controlled and avoid starvation at overload of low drop precedence traffic. To achieve this, we propose a new AQM mechanism, WRED with Thresholds (WRT). The benefit of WRT is that, without reconfiguration, it offers sheltering when low drop precedence traffic is properly controlled and relative differentiation otherwise. We present simulations showing that WRT has these properties.

On Creating Proportional Loss-Rate Differentiation: Predictability and Performance

Recent extensions to the Internet architecture allow assignment of different levels of drop precedence to IP packets. This paper examines differentiation predictability and implementation complexity in creation of proportional lossrate (PLR) differentiation between drop precedence levels. PLR differentiation means that fixed loss-rate ratios between different traffic aggregates are provided independent of traffic loads. To provide such differentiation, running estimates of loss-rates can be used as feedback to keep loss-rate ratios fixed at varying traffic loads. In this paper, we define a loss-rate estimator based on average drop distances (ADDs). The ADD estimator is compared with an estimator that uses a loss history table (LHT) to calculate loss-rates. We show, through simulations, that the ADD estimator gives more predictable PLR differentiation than the LHT estimator. In addition, we show that a PLR dropper using the ADD estimator can be implemented efficiently.

Effects on TCP from Radio-Block Scheduling in WCDMA High Speed Downlink Shared Channels

Avoiding delay jitter is essential to achieve high throughput for TCP. In particular, delay spikes can cause spurious timeouts. Such timeouts force TCP into slow-start, which may reduce congestion window sizes drastically. Consequently, there may not always be data available for transmission on bottleneck links. For HS-DSCH, jitter can occur due to varying interference. Also, properties of the radio-block scheduling influence the jitter. We evaluate, through simulations, effects on TCP from scheduling. Our evaluation shows that round-robin (RR) schedulers can give more jitter than SIR schedulers. SIR schedulers discriminates low SIR users to improve spectrum utilization while RR schedulers distribute transmission capacity fairly. The high jitter with RR scheduling cause however both lower utilization and decreased fairness in throughput among users than with SIR scheduling. The Eifel algorithm makes TCP more robust against delay spikes and reduces thereby these problems.

Drop strategies and loss-rate differentiation

When offering loss-rate differentiation in IP networks, the drop strategy used can have a considerable influence on packet loss and delay. In particular, a strategy of dropping packets only as they arrive can cause bursty loss patterns and high jitter. When only arriving packets are dropped, the router may need to wait for low priority packets to arrive before dropping any packet. This results in larger queue oscillation than if low priority packets were dropped immediately from the queue. Queue oscillation gives bursty loss patterns and delay jitter. We present simulations showing that dropping packets from the queue gives smoother loss patterns and less jitter than if packets are dropped only as they arrive. These simulations cover both TCP Sack and TCP Reno. WRED with and without the gentle modification is used to make drop decisions.

Comparison of wireless network simulators with multihop wireless network testbed in corridor environment

This paper presents a comparative study between results of a single channel multihop wireless network testbed and the network simulators ns-2 and ns-3.We explore how well these simulators reflect reality with their standard empirical radio modeling capabilities. The environment studied is a corridor causing wave-guiding propagation phenomena of radio waves, which challenges the radio models used in the simulators. We find that simulations are roughly matching with testbed results for single flows, but clearly deviate from testbed results for concurrent flows. The mismatch between simulations and testbed results is due to imperfect wireless propagation channel modeling. This paper reveals the importance of validating simulation results when studying single channel multihop wireless network performance. It further emphasizes the need for validation when using empirical radio modeling for more complex environments such as corridors.

Adaptive threshold-based admission control

We present an algorithm for performing dynamic per-link admission control. It is designed to be suitable for deployment using existing quality of service and router techniques. The rationale for the algorithm is to find a bitrate limit for reserved capacity, which enables efficient capacity utilization and statistical multiplexing gain. By using simple traffic meters configured with respect to defined service levels, a network resource controller estimates an admission limit. The estimated limit reflects the ratio between reserved capacity and the aggregate behavior of current flows. Simulations indicate that reliable admission limits can be estimated already at low reservation levels.

Hints and notifications [for wireless links]

With current Internet protocols, users may experience low and unpredictable forwarding quality at wireless links. This is due to varying link properties caused by changing radio conditions. Decreased forwarding quality can cause severe degradation in utilization. This is undesirable since forwarding capacity often is expensive at wireless links because of the limited radio spectrum. Allowing the application and transport layers to communicate with wireless link layers can improve the forwarding quality and utilization. We propose to enable inter-layer communication by adding hints and notifications (HAN) to the Internet architecture. Hints can be introduced and used without notifications, while notifications need hints or a similar mechanism to operate. By using IP options and ICMP messages to implement HAN, a backward-compatible partial deployment is possible. With HAN, the network layer becomes truly wireless friendly and the radio spectrum can be used efficiently while supporting both real-time and traditional data applications.

Machine learning approach to automatic bucket loading

The automation of bucket loading for repetitive tasks of earth-moving operations is desired in several applications at mining sites, quarries and construction sites where larger amounts of gravel and fragmented rock are to be moved. In load and carry cycles the average bucket weight is the dominating performance parameter, while fuel efficiency and loading time also come into play with short loading cycles. This paper presents the analysis of data recorded during loading of different types of gravel piles with a Volvo L110G wheel loader. Regression models of lift and tilt actions are fitted to the behavior of an expert driver for a gravel pile. We present linear regression models for lift and tilt action that explain most of the variance in the recorded data and outline a learning approach for solving the automatic bucket loading problem. A general solution should provide good performance in terms of average bucket weight, cycle time of loading and fuel efficiency for different types of material and pile geometries. We propose that a reinforcement learning approach can be used to further refine models fitted to the behavior of expert drivers, and we briefly discuss the scooping problem in terms of a Markov decision process and possible value functions and policy iteration schemes.

Cognitive load-control for congested Wireless Sensor Network channels

Modern wireless sensors are capable of analyzing measured data and transmit at high rates if needed. They may individually or collectively select how much data shall be sent depending on predefined rules tied to properties of measured data. Sending rates thereby change triggered by events appearing in the sensor environment. With wireless technologies such as 802.15.4 (ZigBee) having limited capacity uncontrolled traffic can easily cause overload resulting in a decreased throughput. This paper presents a load control mechanism that uses multiple inputs, demand for transmission capacity and perceived quality at each sensor node, to properly configure sending rates of these nodes. The mechanism is based on Fuzzy Logic and quality assessed with utility functions for packet loss and throughput. NS-3 simulation results show clear improvements in overall throughput at high loads when the proposed mechanism is enabled.

A Measurement Study for Predicting Throughput from LQI and RSSI

Wireless sensor networks (WSN) commonly use ZigBee to communicate, especially when low power consumption is demanded. ZigBee may however provide unpredictable throughput although transmission distances are short. This is especially evident in difficult environments with complicated reflections and various materials through which radio signals need to pass through. Distributed scheduling based on cognitive networking principles may improve both network predictability and overall throughput. This paper presents measurements of key parameters for such cognitive scheduling, and discusses their potential for predicting suitable per-node transmission rates. Results include variability of throughput, RSSI and LQI observed for different transmission powers, transmission ranges, and number of transmitting nodes.