Lan Wang - Academia.edu (original) (raw)

Papers by Lan Wang

Task allocation systems in the Cloud have been recently proposed so that their performance is opt... more Task allocation systems in the Cloud have been recently proposed so that their performance is optimised in real-time based on reinforcement learning with spiking Random Neural Networks (RNN). In this paper, rather than reinforcement learning, we suggest the use of multi-layer neural network architectures to infer the state of servers in a dynamic networked Cloud environment, and propose to select the most adequate server based on the task that optimises Quality of Service. First, a procedure is presented to construct datasets for state classification by collecting time-varying data from Cloud servers that have different resource configurations, so that the identification of server states is carried out with supervised classification. We test four distinct multi-layer neural network architectures to this effect: multi-layer dense clusters of RNNs (MLRNN), the hierarchical extreme learning machine (HELM), the multi-layer perceptron, and convolutional neural networks. Our experimental results indicate that server-state identification can be carried out efficiently and with the best accuracy using the MLRNN and HELM .

—Cloud systems include both locally based servers at user premises and remote servers and multipl... more —Cloud systems include both locally based servers at user premises and remote servers and multiple Clouds that can be reached over the Internet. This paper describes a smart distributed system that combines local and remote Cloud facilities. It operates with a task allocation system that takes decisions to allocate tasks dynamically to the service that offers the best overall Quality of Service and a routing overlay which optimizes network delay for data transfer between clouds. Experimental results are conducted at the global intercontinental level, both to collect data for decision making and to illustrate the effectiveness of our approach.

This paper uses big data and machine learning for 3 the real-time management of Internet scale qu... more This paper uses big data and machine learning for 3 the real-time management of Internet scale quality-of-service 4 (QoS) route optimization with an overlay network. Based on the 1 5 collection of data sampled every 2 min over a large number of 6 source–destinations pairs, we show that intercontinental Internet 7 protocol (IP) paths are far from optimal with respect to QoS met-8 rics such as end-to-end round-trip delay. We, therefore, develop 9 a machine learning-based scheme that exploits large scale data 10 collected from communicating node pairs in a multihop overlay 11 network that uses IP between the overlay nodes, and selects paths 12 that provide substantially better QoS than IP. Inspired from cog-13 nitive packet network protocol, it uses random neural networks 14 with reinforcement learning based on the massive data that is col-15 lected, to select intermediate overlay hops. The routing scheme is 16 illustrated on a 20-node intercontinental overlay network that col-17 lects some 2 × 10 6 measurements per week, and makes scalable 18 distributed routing decisions. Experimental results show that this 19 approach improves QoS significantly and efficiently. 20 Index Terms—The Internet, big data, network quality of service 21 (QoS), smart overlays, random neural network, cognitive packet 22 network. 23

The increasingly wide application of Cloud Computing enables the consolidation of tens of thousan... more The increasingly wide application of Cloud Computing enables the consolidation of tens of thousands of applications in shared infrastructures. Thus, meeting the QoS requirements of so many
diverse applications in such shared resource environments has become a real challenge, especially since the characteristics and workload of applications differ widely and may change over time. This paper presents
an experimental system that can exploit a variety of online QoS aware adaptive task allocation schemes, and three such schemes are designed and compared.
These are a measurement driven algorithm that uses reinforcement learning, secondly a ``sensible'' allocation algorithm that assigns tasks to sub-systems that are observed to provide a
lower response time, and then an algorithm that splits the task arrival stream into sub-streams
at rates computed from the hosts' processing capabilities. All of these schemes are compared via measurements
among themselves and with a simple round-robin scheduler, on two experimental test-beds with homogenous and heterogenous hosts having different processing capacities.