Research On Performance Optimization And Quality Control In Approximate Networks-on-Chip

Thanks to the error forgiveness of some applications(multi-media processing,machine learning,etc.),the components running these applications can improve system performance or reduce energy consumption through mitigating computation effort.Since Networks-on-Chip(No Cs)are the main contributor to power consumption and performance of a many-core system,researchers have proposed approximate No Cs,which use lossy compressors to make data compression for on-chip data communications.They are designed to reduce network workloads,and thus improve system performance.When designing an approximate approach,since users have their quality requirements to application outputs,the quality loss and benefit caused by approximation should be taken into account when making approximation decisions.Therefore,it is crucial to build an accurate quality loss model and to maximize the approximation benefits under the quality loss constraint.We observed that there are several problems existing in the state-of-the-art approximate No Cs: 1)the global network status is not considered,and 2)a quality model is needed to guarantee quality safety when making approximation decisions at runtime.Therefore,in this paper,the quality control method for approximate No Cs is studied.First,we proposed a quality model,which takes the compressing method,program,input data and approximation level into account.Then,other important metrics are modeled,such as network performance and communication volume.Next,an optimization problem for approximate No Cs is formulated,which makes approximation decisions under the quality loss constraint.Lastly,to solve the problem,we proposed a method which is based on global network status and error budgeting.The main idea of this method is to first minimize network congestion,followed by reducing packet zero-load latency.Compared with two recent works,the proposed method speeds up application execution for as much as 29.42%.The proposed method is able to mitigate network congestion,and it controls quality loss accurately.