The Research And Implementation Of Support And Optimization Of Kylin Memory Management System For NUMA

NUMA (Non-Uniform Memory Access) architecture has good scalability, which's SMP's(Symmetric Multiple Processor) disadvantage. Also, it keeps the strongpoints of SMP, such as symmetric, coherent highspeed cache, single address space and single system image, etc. The prospect of NUMA's application could be foreseen in the fields of science computing and business service.Memory is the most important resource. All physical memory in NUMA system are addressed uniformly, but distributed in the whole system. That leads to the variable access latency of portions of memory from other system components. So improving the data locality becomes the key to enhance system performance. Also, the discontiguous memory complicates the design and implementation of memory management. Based on Kylin, we focus on the technology of memory management to support and optimize NUMA.This thesis firstly studies the discovery and representation of system topology. With the complete supports for ACPI specification, Kylin OS provides a group of interfaces, with which memory management could consciously allocate resources according to the topology.Secondly, because the flat structure of physical memory management in Kylin can not meet the need of NUMA, we expand it to a node-zone-page 3-level frame structure. This structure can support discontiguous memory well, and map the topology information neatly.Thirdly, to improve the data locality, we propose and implement a memory allocation algorithm aware of nodes affinity. It can balance the different latency between local and remote memory accesses, thus maximize the system performance. The simulation result proves its serviceability.Lastly, we study and implement the replica optimization technology in Kylin, including kernel and user data replication. These technics can reduce the amount of remote memory accesses, and enhance the system performance greatly.