数据中心中路由编码的可行性研究

Feasibility Study of Routing Codes in Datacenters

修复带宽最优并不代表修复通信量也是最优的,后者与物理网络拓扑有着密切联系.本文基于路由编码的思想减少修复通信量.首先,基于信息流图对物理网络中数据的传递过程进行建模,证明得出了满足路由编码可行的充要条件,并发现路由编码可以基于再生码实现.然后,针对数据中心网络设计的特点,为Fat-tree设计了一个工作在应用层的协议来生成修复树,为CamCube设计了一个启发式算法来生成修复树.关于最小存储再生码的数据修复过程的仿真实验表明,路由编码可以有效地降低修复通信量,2种修复树生成方案在各自适合的网络中均有较好性能:在帮助节点数较小时,Fat-tree和CamCube中的修复通信量分别降低了大约50%和30%.

Repair traffic is not always optimal when repair bandwidth is optimal.The former is relative to physical network topology.This paper aimed at reducing repair traffic based on routing codes.First,we modeled data transmission in physical networks based on information flowgraph so that we could get the necessary and sufficient condition to feasibility of routing codes.And we found that routing codes could be realized based on regenerating codes.Then,we designed a protocol working on application layers to generate repair trees in Fat-tree,and a heuristic algorithm to generate repair trees in CamCube,which were both in agreement with their own design features of datacenter networks.Simulations about data-repair processes in systems using minimum-storage regenerating codes showthat routing codes can reduce repair traffic efficiently,and performance of the two generation schemes of repair trees are both good in their own adapted networks.In fact,repair traffic had about 50% and 30% reductions in Fat-tree and CamCube respectively when the number of providers was small.