QAT Hardware Accelerated Compression In Block Layer Storage

Data compression can enhance system I/O performance and storage efficiency,but applied on file system and application layer results in the disadvantage of deployment because the modifications of specific demands are complicated.VDO(Virtual Data Optimizer)is a block layer compression technology,can provide transparent compression to upper layer as a kernel module,in order to be easily applied on different scenarios.In this paper,we investigate VDO technology,and summarize the problems of its design are:(1)block layer compression have demands of high compression ratio algorithm,low ratio algorithm is not suit for the scenario.(2)high compression ratio algorithms have large consumption of CPU computation resources,and software algorithm library cannot afford it.Prior studies focus on using general-purposed hardware to implement high compression ratio algorithms,but ASIC(Application Specific Integrated Circuit)has better efficiency and lower resource consumption compared with general-purposed hardware.Nevertheless,there are three challenges when introduce QAT hardware into VDO technology:(1)Compression module does not support selecting algorithms,need to introduce an algorithm management.(2)Synchronous mechanism in compression logic is not suit for QAT hardware,need to introduce asynchronous mechanism.(3)Packer module only use a single Packer and influence efficiency,need to introduce parallel mechanism.In this paper,we propose QVDO(QAT accelerated VDO)block layer compression technology,integrating QAT ASIC technology into VDO technology,using hardware implemented compression algorithm with high compression ratio,providing transparent and efficient data compression to upper layer.In the design of QVDO system,we mainly implement and optimize these three modules:(1)introduce an algorithm controller into compression module to support different compression algorithms,and provide an exception handling mechanism to ensure data compression and decompression is available when hardware exception occurs.These can enhance scalability and availability of system.(2)implement a QAT hardware interaction module,using asynchronous mechanism,letting calling routine and callback routine to perform sending and receiving data works separately,and providing a memory management mechanism to allocate and free memory,and also a context transmission mechanism to prevent information lost.These can leverage QAT hardware accelerator processing ZLIB compression efficiently.(3)introduce a parallel processing mechanism to packer module,using dispatcher to distribute unhandled compressed fragments into several Packers and processes packing works separately,and also provide a packing checker to check packing effect.These can accelerate packing workflow and enhance system efficiency.We use Calgary Corpus and FIO to show that QVDO system can enhance compression ratio up to 1.93 x,enhance read and write throughput to 1.23 x and 1.71 x separately,and CPU utilization decreases 19% compared with VDO technology.we establish QVDO system on Ceph storage system,and use Filebench as macro-benchmark,shows that QVDO system can improve throughput to 1.1x,and performs well when running several applications such as file server.