Load Balancing Scheme Design Based On Erasure Code Offloading

With the rapid growth of online services and cloud computing,large-scale data centers(DCs)are being built around the world.Because of the high CPU overhead and transmission latency,traditional TCP/IP stack feels powerless when facing new generation data center workloads.The data center networks(DCNs)usually become the bottlenecks of application performance.The rising of RDMA introduces a new direction of speeding transmissions in the DCNs.Some large-scale DCs have deployed RDMA as transmission protocol for a long time.But load balancing scheme ECMP which is widely deployed in the DCs can't balance load perfectly.For RDMA networks with high throughput and low latency,load balancing schemes become more important.The existing RDMA works don't fully utilize the multi-path feature of the DCNs.Fully utilizing multiple paths to transmit data can not only improve the utilization of the entire network,but also decrease the delay of network transmissions and improve the throughput of applications.At the same time,the existing works usually use RDMA for specific applications by tightly coupled developing.Developers need to understand RDMA well,and then use sufficient hardware knowledge to set proper RDMA parameters to provide high performance network transmissions for specific applications.These bring users who want to use RDMA to accelerate their applications troubles.To address troubles when deploying RDMA in the DCNs,this thesis designs and implements a load balancing scheme based on RNIC erasure code offloading and its successor a general multi-path RDMA transmission framework.The main contributions of this thesis include:(1)We investigate state-of-the-art load balancing works and their merits and demerits.Combining the erasure code(EC)offloading supported by RDMA NIC with the multi-path feature of the hierarchical DCNs,we investigate the feasibilities of load balancing and multi-path transmission using the EC offloading.(2)For problems the ECMP suffers,we designed a load balancing scheme,which provides good traffic load balancing support for the DCNs with RDMA deployed.And simulations show the scheme can effectively improve application performance and improve the DCNs' utilization.(3)With the multi-path feature in the DCNs,a RDMA multi-path transmission system is provided.By using this system,users can easily use RDMA for network transmission which providing high throughput,low latency,low CPU overhead transmission services for the DCNs.We verified our designs through experiments.