Researches On Fairness Oriented Dynamic Bandwidth Allocation In Datacenters

In recent years,cloud computing is more and more popular along with many cloud providers emerged,and many tenants have deployed their applications on the cloud.Cloud providers lease resources to tenants in the form of Virtual Machines(VMs).Network is critical resource for applications,however,the VMs that current cloud datacenters offer only specify the CPU,memory and storage resources,while do not specify the network resource.Network is shared among untrusted tenants in a best-effort manner,which cannot guarantee fairness among tenants,and can't guaran-tee VM's network performance.So,network sharing is a urgent problem needed to be solved for cloud datacenters.Cloud datacenters can reserve static bandwidths for tenants to guarantee their ap-plications network performance,which will lead to inefficient utilization of network resource.Because the occurrences of VM,s flows are random,when some VMs have no data to transfer,the bandwidths reserved for them cannot be allocated to other VMs,leading to waste of network resource.To improve the network's utilization and VM s bandwidth,we should allocate bandwidth dynamically based on real-time network s-tate.Several solutions have been proposed on the dynamic bandwidth allocation prob-lem.In most solutions,bandwidth allcation is performed in hosts.But hosts only have very little knowledge on the network's state,and they cannot work out each VM's pre-cise share of bandwidth.So they use self-adaptive method to adjust VM's bandwidth,which need many times to converge to fair allocation.The emerging of Software Defined Network(SDN)offers us a new paradigm:we can maintains network's state view in the controller(s),and compute fair bandwidth for each VM based on the view.Given VMs' states and their weights set by tenants,we can work out the fair allocation,this is a max-min fairness allocation problem,which can be solved by a eentralized solution.However,datacenter network,s state changes frequently,and we need dynamically adjust the allocations along the time,while centralized solution suffers serious scalability problem.So in this paper,we propose a Link Max-min Fairness bandwidth allocation method(LMF),which performs bandwidth allocation on one link each time.Because a change of networks state only influence a part of links,so we only need perform LMF on the links which are influenced.LMF is a distributed algorithm,while multiple work pro-cesses can perform LMF on different links simultaneously,which makes LMF more scalable.In addition,we define multiple rules to synchronization alloeation informa-tion among links,to make our allocation converges to global fairness.In this paper,we design an SDN-based bandwidth allocation system FairShare,which adopts LMF to allocate bandwidth.FairShare has three main modules:VM monitor,maintains network's state view;Bandwidth Allocator,performs bandwidth allocation;Rate controller,controls each VM's rate through switches.We verify Fair-Share through experiments,and results shows that it not only maintains fairness among VMs,and improves their network performances.In the current datacenter networks,there are multiple paths between a pair of hosts generally.The previous studies on dynamic bandwidth allocation problem only select one static link for each pair of virtual machines,which may lead to low utilization of cloud network.In order to take the advantage of multipath,we proposed a best-path based system FairShare-MP.When a VM has data to transfer,we will first choose the best path for it,and then allocate bandwidth for it on the optimal path.Our best-path selection algorithm is optimized,so the path selection time is very short.Experiments shows that,best-path based policy improves the VM's bandwidth greatly.