A Research On TLB Performance Optimization For Exclusive Hierarchy

The Translation Look-aside Buffer(TLB)is critical to the performance of modern virtual storage systems,because it stores the mapping from virtual addresses to physical addresses.Every time when processor accesses the memory,it needs to access the TLB first.In order to reduce the access latency of the TLB and expand its address mapping coverage,current memory systems widely adopt an inclusive hierarchy to manage two-level TLB.However,with the extensive use of memory-intensive applications,the memory footprint of the program has become larger,and the mapping from virtual address to physical address needs to be performed frequently,which puts much pressure on TLB performance.In response to this challenge,the exclusive hierarchy has attracted much attention due to its capacity advantages.With this advantage,it can not only improve the utilization of TLB space,but also increase its mapping coverage.Therefore,this thesis exploits an exclusive hierarchy to manage two-level TLB to improve the TLB hit rate.However,in order to take full advantage of the capacity advantages of the exclusive hierarchy,two major challenges need to be addressed.One is the problem of wasting TLB space caused by "dead" data(which will not be accessed before eviction);the other is the problem of hot data replacement caused by frequent data swap.In response to the above two problems,this thesis has done the following work:(1)This thesis qualitatively and quantitatively compares the hit rate between exclusive TLB and inclusive TLB,fully verifies the potential advantages of exclusive TLB in increasing address coverage,and demonstrates the two issues(i.e.,dead data and frequent data swap)that need to be resolved when using exclusive hierarchy through experiments.(2)Aiming at the dead data problem faced by exclusive TLBs,this thesis proposes a dead data management strategy based on temporal locality and spatial locality.By recording the number of data "trip" between the two levels of TLB,and combining the temporal locality of data access and the spatial locality of adjacent data,dead data can be accurately predicted and deleted,thereby improving the utilization of TLB cache space.(3)Aiming at the problem of frequent data swap,this thesis proposes a low-overhead data retention strategy.The core of this strategy is to use the number of trips and temporal locality to distinguish data reuse,and keep the data with long reuse distance in a specific buffer,which is the "shadow" way in this thesis.This thesis selects ten typical workloads from SPEC CPU2006,PARSEC 3.0 and Graph Big to verify the proposed strategies and methods.The experimental results show that,compared with the two-level TLB with an inclusive hierarchy,the dead data prediction and reused data retention scheme in this thesis can reduce the access miss and latency.Specifically,the number of misses was reduced by 9.08% on average,and the highest was 29.38%;the latency was reduced by an average of 8.84%,up to29.47%.