Efficient cross-layer routing and congestion control architectures for distributed systems

Cross-layer routing and congestion control;Adaptive rate allocation;Server selection;Load balancing;Generalized efficient sharing (GES);Fractional (partial) flows;Network level max-min fairness;Network level service level agreement;Network level quality of service;Hybrid peer to peer networks;Per packet prices;Efficient peer incentives;Meta-data management;Active/passive content;Content index management;Secure content exchange;Power-efficient datacenter;Efficient Openflow;Software defined networking;Surrogate servers;Low content transfer times;High throughput architectures

Kassa, Debessay

关键词: Cross-layer routing and congestion control;    Adaptive rate allocation;    Server selection;    Load balancing;    Generalized efficient sharing (GES);    Fractional (partial) flows;    Network level max-min fairness;    Network level service level agreement;    Network level quality of service;    Hybrid peer to peer networks;    Per packet prices;    Efficient peer incentives;    Meta-data management;    Active/passive content;    Content index management;    Secure content exchange;    Power-efficient datacenter;    Efficient Openflow;    Software defined networking;    Surrogate servers;    Low content transfer times;    High throughput architectures;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/45430/Debessay_Kassa.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Major distributed systems such as the Internet, datacenter and hybrid P2P networks share a common knownchallenge of finding an optimal path to transfer content from a source to a destination and the optimal rateat which content is transmitted. In general networks such as the Internet, per user, there is usually onepossible content source/destination such as a web server. There can be multiple possible paths to/from thedestination/source (server).In datacenter networks which usually have a tree structure and in hybrid Peer-to-Peer (P2P) networks,there can be multiple possible servers at which content can be stored and from which content can be re-trieved. Multiple possible servers (sources/destinations) translates into multiple possible paths to/from acontent destination/source. Finding an optimal path to/from a destination/source requires efficient conges-tion control and routing schemes.The transmission control protocol (TCP) is the major congestion control protocol in the Internet. TCPand its variants have the drawback of not accurately knowing rate share of flows at bottleneck links. Someprotocols proposed to address these drawbacks are not fair to short flows, which are the majority of theInternet traffic. Other protocols result in high queue length and packet drops which translate into a highaverage flow completion time (AFCT).The currently major deployed intra-domain routing algorithm is the Open Shortest Path First (OSPF)[1]. OSPF uses a simple heuristic routing metric (link weight). The routing metric used doesn’t properlytake into account the latest status of the network. Other traffic engineering schemes such as the TeXCPproposed to address the routing issues of existing schemes also fail to find a globally (domain level) optimalroute. Besides, they incur additional probing and path reservation packet overheads.Recently deployed datacenter network architectures rely on random server and hence path selection inthe attempt to ease congestion. However such random selection of paths can result in serious congestionand content transfer delay. This can for instance be caused by large content transfers (elephant flows) whichtake a long time to finish. In this case a random path selection can add to the congestion caused by elephantflows. Existing cloud datacenter architectures such as the Google File System (GFS) and the Hadoop FileiiSystem (HDFS) rely on a single name node server (NNS) to manage metadata information of which contentis stored in which server. A single NNS can be a potential bottleneck and a single point of failure.Hybrid P2P content sharing can result in significant scalability gains in addition to assisting contentdistribution networks (CDNs). However, currently proposed CDN and P2P hybrid schemes do not provideaccurate, fair and efficient incentives to attract and maintain more peers. Besides, they do not use efficientprioritized congestion control and content source selection mechanisms to reduce content transfer times.In this thesis, we present the design and analysis of cross-layer congestion control and routing protocolsto address the above challenges of major distributed systems. Our schemes derive an efficient rate metricwhich sources use to set their sending rates and which routers/switches use as a link weight to computean optimal path. Among other contributions our rate and path computation schemes achieve network levelmax-min fairness where available resources are quickly utilized as long as there is demand for them. Ourschemes have prioritized rate allocation mechanisms to satisfy different network level service level agreements(SLA)s on throughput and delays.For cloud datacenter networks, our scheme uses a light weight front end server (FES) to allow the useof multiple NNS and there by mitigate the shortcomings of existing architectures. For hybrid P2P networksour schemes ensure high and accurate incentives to participating peers. Such fair incentives attract morepeers which securely download and distribute contents. The thesis also presents efficient content indexmanagement schemes for the hybrid P2P networks with robust incentive implementation mechanisms.We have implemented our protocols for general networks (the Internet), for cloud datacenter and hybridP2P networks in the well known NS2 simulator. We have conducted detailed packet level and trace-basedexperiments. Simulation results show that our protocol for general networks can result in the reduction of theaverage file completion time (AFCT) by upto 30% when compared with well known existing schemes. Ourcross-layer design for cloud datacenter networks can achieve a content transfer time which is about 50% lowerthan the existing schemes and a throughput which is higher than existing approaches by upto than 60%. Ourdetailed trace-based experiments also shows that our hybrid P2P protocol outperforms existing schemes interms of file download time and throughput by up to 30% on average. The results also demonstrate that ourhybrid P2P scheme obtains fair uplink prices for the uploaders and fair cost for the downloaders maintainingan overall system fairness. Besides, the results show the efficient enforcements of the prioritized allocations.Our implementation of the hybrid P2P protocol using an Apache SQL Server with PHP in Linux virtualmachines demonstrates that content index management mechanisms of our protocol are scalable.