[Linux kernel] NUMA non-uniform memory access structure ( NUMA concept introduction | NUMA architecture advantage analysis | SMP, NUMA, MPP architecture)





1. NUMA non-uniform memory access structure



Non-uniform memory access structure , English name Non Uniform Memory Access , NUMA for short ;

In the system of "non-uniform memory access knot" , there are multiple CPU processors , each processor has its own independent local memory , each CPU processor only supports its own local memory fast access , and there is a connection between each CPU processor. connected via "bus";

The processor can access the "memory" of other processors through the "bus" , but the processor can access its own memory faster than the memory of other processors ;

NUMA non-uniform memory access structure, can support CPU expansion , can expand hundreds of CPU processors;

insert image description here







2. Analysis of NUMA Architecture Advantages



NUMA non-uniform memory access structure (Non Uniform Memory Access) system architecture, can integrate multiple processors , so that the system has high performance in "processing transactions" ;

In the NUMA architecture, the processor accesses its own local memory very fast , but the memory of other processors is slow , so in order to ensure the execution performance of things, it is necessary to reduce the data interaction between the CPU processors. The NUMA architecture is only suitable for OLTP ( On-Line Transaction Processing (On-Line Transaction Processing) transaction processing scenarios;


When using OLTP programs, if users access the central database, the efficiency of using SMP architecture is higher than that of MPP;





2. SMP, NUMA, MPP architecture



The application generally has 3 33 types:

  • Symmetric multi-processor structure , English name is "Symmetrical Multi-Processing" , referred to as SMP;
  • Non-uniform memory access structure , English name Non Uniform Memory Access , NUMA for short ;
  • Massive parallel processing structure , English name "Massive Parallel Processing" , referred to as MPP;

Guess you like

Origin blog.csdn.net/han1202012/article/details/123905502