Research And Implement Of Resource Placement Technology Based On CC-NUMA

Nowadays, NUMA (Non-Uniform Memory Access) becomes the mainstream architecture in the high performance server. Compare with SMP, NUMA has a characteristic of different memory access latency, and has a special optimize requirement to operation system.Linux operation system's work contentrates on the kernel in order to optimize the performance of NUMA, for instance, the support of discontig memory, the support of ACPI, Page Migrate, Page Replicate and Page Placement. From the point of view of usr-level, our work puts forword the conception of Resource Placement, it places task scheduling and memory allocation onto specific system resources (processors and memory blocks etc.) for root user and application.First of all, this paper introduces the correlative background knowledge of our work and all kinds of memory optimization technology for NUMA, then brings forword the conception of Resource Placement and its fundamental, analyzes the implement mechanism of CpuMemSets on Linux operation system, and observes CpuMemSets's fault of lack of the system resources topology information. Thus, our work commences on kernel-level and user-level respectively for improves the performance of system.On kernellevel, we present a Hierarchical Scheduling algorithm base on NUMA topology, which is strengthens the operation system's NUMA support from the aspect of task scheduling in kernel, and realizes NUMA affinity scheduling, improves data access locality and system performance.On the other hand, we provide usable CpuMemSets user-level library, and present a multi-decision-making information guided resource placement policy, this policy resolves the problem of improper placement because of ignorant of topology information efficiently.Finally, validates the usability and validity of Resource Placement technology and NUMA affinity scheduling by simulates NUMA environment on a SMP machine.