| With rapid growth of Internet technology and cloud computing services,both data and traffic volume in datacenter have increased explosively.These facts boost datacenter to expend its number of servers to hundred-thousand and raise port rate of NICs from 1 Gbps to 40 Gbps.It is a great challenge to design a high-speed datacenter network with such huge scale.Load balancing is very crucial in improving datacenter network efficiency,as it can assign the network resources in an optimal way.Unfortunately,traditional load balancing schemes are not suit for the high-speed datacenter network.it is because of their control granularities are coarse,balancing mechanisms are too simple,and they are also lack of network informa-tion awareness.To makeup these flaws,researchers have designed some new schemes with OpenFlow protocol and flowlet technology.Although these new schemes have improved the datacenter network performance,it still faces many issues like cannot implement into a large scale network and not support RoCEv2 protocol.Therefore,this thesis proposed a flowlet-based load balancing scheme with switch-based congestion information awareness,and designed a flowlet load balancing scheme with an out-of-order packets receiving mech-anism for RoCEv2 protocol.To implement the load balancing scheme into a large scale,this thesis designs a flowlet-based load balancing scheme with switch-based congestion information awareness for high-speed datacenter network.The scheme quantifies the egress port queue status of switch to link congestion factor.As the value of this factor exceeds a certain threshold,switch will send a congestion notification,called SQCN,to the edge switches.In traditional schemes,an edge switch needs k~4 entries to record congestion information for all paths from one edge switches.This scheme uses a multi-layer linked list to store congestion information,which records the overlap link between paths only once and reduces the number of entries to k~3.As a result,the scheme becomes much scalable.By analyzing both in theoretical way and simulation,the results shown that,compare to equal cost multi-path routing(ECMP),the scheme has a 20%increase in throughput,60%decrease in small flow completion time(FCT)and 80%decrease in large FCT.These improvements are much better than CONGA,which is a similar load balancing scheme,and the storage cost is also reduced by 80%compare to CONGA.As RoCEv2 protocol not support out-of-order packet reordering,it will have high retrans-mission cost when using current load balancing schemes in datacenter network.This thesis designs an end-to-end flowlet-based load balancing scheme and a receive mechanism for RoCEv2 protocol to handle out-of-order packets.This scheme is based on DCQCN,a con-gestion control protocol for RoCEv2.When the receiver receives out-of-order packets,it no longer discards them directly,but stores them into a new data segment,which avoids the performance loss caused by retransmission of data.To reduces the reordering processing overhead,when flowlets arrive continuously,the buffer will not submit the data segments to the protocol stack,instead,it will expand its space and receive the subsequent flowlets.Some RDMA applications will generate a large number of small requests concurrently,in this case,traditional schemes will treat each request as a flow,causing traffic load unbalance between paths.This scheme aggregates these small requests into many flowlets with a fixed length,similar to the large flow flowlets'length.Therefore,the switch can just send flowlets randomly and traffic load will be well-balanced.From the simulation results,the scheme has a very similar performance with the packet-level load balancing in FCTs and through-put,which is much better than the flow-level.Compared with the Presto,a scheme of the same type,the protocol stack processing overhead is reduced by 90%. |
PDF Full Text Request