Research On Quality-of-service Guarantee In Bufferless On-chip Networks

Datacenter is the infrastructure of cloud computing, which runs a variety ofapplications including latency-sensitive jobs and long-running off-line jobs. In currenttimes, resource reservation for user-facing workloads in online datacenters is acommon way to meet the Quality of Service constraints, however, leading tounderutilization. Co-locate different kinds of applications on shared machines couldincrease the load of datacenters, while the interference between co-located workloadsthat can lead to severe performance losses. In datacenter environment, concurrentapplications contend on shared resources, such as OS kernel、CPU、Network-on-Chip(NoC), would cause performance of latency-sensitive applications wane. Hence, it isvery important to guarantee the QoS of latency-sensitive jobs through better sharedresources management.Heterogeneous cloud computing is going to be popular in the future, and with thehigh performance, many-core CPU would play an important role in the nextgeneration datacenters. NoC in many-core system is a critical shared componentwhich handles all the inner chip communications. Emerging bufferless NoC hasadvantages of low power consumption and complexity that is suitable forpower-critical datacenters. Nevertheless, it has drawbacks on congestion control andQoS guarantee for latency-sensitive applications give rise to long tail phenomenon. Toaddress these issues, the main works of this paper are as follows:(1) We propose DATwhich is an efficient deflection aware congestion control mechanism for bufferlessNoC. By analysis of the deflection information which are recorded in each router,DAT can preciously seek out those injections which cause network congestion. Andthen DAT throttles these injections properly, improving the overall system throughput.As evaluated in a modeled16-core system, DAT reduced the deflection rate by8.6%compared to the baseline.(2) We implement QBLESS architecture which is a QoSaware bufferless NoC design. QBLESS is composed of two parts: a QoS awarerouting algorithm that can substantially reduce flit deflections of high-priorityapplications; and a congestion control mechanism that can achieve the desiredperformance for high-priority applications. In a modeled64-core system, compared toprevious state-of-the-art bufferless NoC design, QBLESS promotes the performanceof high-priority applications by33.2%and reduces network hops by42.8%.