The Research And Implementation Of Metadata And Data Classified Cache Strategy

Metadata and common data in the file system have different access characteristics.In current storage system,all requests are encapsulated into unified commands and sent to the device.According to the traditional I/O path model,all requests are delivered into the storage device without distinction.All data would be managed uniformly in the cache system because the information of data request type cannot be obtained.The scattered metadata requests take more time to access the disk,so metadata should be retained in the cache preferentially.Apparently the current single cache system cannot satisfy this point.Based on the analysis of two types of data I/O access features,this thesis proposes a method to separate metadata from common data in the cache system.Firstly,add the flag of data type to requests issued by the file system.With the delivery of I/O requests,the flag information is passed to the underlying storage system.Extract the flag information in the storage system and then cache common data and metadata separately.This approach can avoid common data occupying metadata cache space.The cache blocks are organized by the radix-tree structure.Two caches belong to different trees.According to the different access characteristics,two kinds of data will be scheduled differently.The common data with greater size is replaced with a radix-tree node(multiple and contiguous cache blocks)while the metadata is replaced with a cache block.The access pattern of common data typically possesses spatial locality,so the cache blocks on the same radix-tree node will be pre-fetched.In order to ensure the hit rate of metadata,it is necessary to adjust the proportion of metadata cache capacity.The dirty blocks in the cache will be flushed to the disk by the asynchronous thread.The test results show that the I/O performance of cache system that separates metadata from common data is improved by 18.7% compared to the unified cache system.