Top performance version 2 reading notes (IV) -- memory monitoring

《System Performance: Enterprise and the Cloud, 2nd Edition (2020)》 Reading notes and brief notes

7. Memory memory

7.1 The term

The term	chinese
main memory	Main memory	Physical memory
virtual memory	Virtual memory	Abstract main memory concept , Virtual memory is not real memory
resident memory	memory-resident	Memory currently in main memory
anonymous memory	Anonymous memory	Memory without file system location or pathname . Including the working data of the process address space , It's called a heap
address space	address space	Memory context
segment	paragraph	An area of memory marked for special use , For example, it is used to store executable or writable pages
instruction text	Instruction text	In memory cpu Instructions , Usually in segments
OOM	OOM	out of memory
page	page	Operating system and CPU The unit of memory used .4KB or 8KB, Modern processing supports multiple page sizes to support larger page sizes
page fault	Page missing	Invalid memory access
paging	Change the page	Exchange pages between main memory and storage devices
swapping	In exchange for	It refers to transferring the whole process from the main memory to the switching device .Linux Medium exchange refers to the transfer of pages to the exchange device （ Migration Exchange page ）
swap	In exchange for （ Space ）	Disk space for storing anonymous data of page change and exchange process . It can be a piece of space for a storage device , Also known as a physical switching device , Or file system , Called swap file .

7.2 Concept

Basic concepts related to memory and memory performance .

7.2.1 Virtual memory virtualmemory

The process address space is mapped by the virtual memory subsystem to main memory and the physical swap device.

swap Partitions are often called swap partitions , This is a special hard disk space , That is, when the actual memory is not enough , The operating system will take out some temporarily unused data from memory , In the swap partition , So as to make enough memory space for the currently running program .

7.2.2 Change the page paging

Page change is to change pages into and out of main memory , They are called page in and page out respectively . allow ：

• Run a partially loaded program
• Run programs larger than main memory
• Efficient migration between main memory and storage devices

Two kinds of ：file system paging 、 anonymous paging.

paging
file system paging	File system page change	File system paging is caused by the reading and writing of pages in memory-mapped files. File system paging is triggered by reading and writing mapped file pages located in memory . If a file system page has been modified in main memory （dirty）, Page out page-out You need to change this page write todisk; contrary , If it hasn't been modified （clean）, Because the disk already has a copy , therefore page-out You can just free up this memory for reuse .
anonymous paging	Change pages anonymously	Anonymous paging involves data that is private to processes: the process heap and stacks. Anonymous page feed involves process private data ： Heap and stack . Why is it called anonymous page change ？ Because in OS It lacks a named address （ If there is no file system path ）. Anonymous page feed requires that data be migrated to physical exchange devices or exchange files ,linux In exchange for （swap） To order this type of page change . Anonymous page feed affects performance , When app When accessing the paged out page , Will be read I/O Operation blocking , This is the process of anonymous page changing , to app Bring synchronization delay . Switching out anonymous page feeds does not directly affect performance , Because it is executed asynchronously by the kernel

7.2.3 Page change on demand demanding paging

7.2.4 Over submission overcommit

Linux The system allows excessive submission , Allow allocation of more than the system can store ---- More than the sum of physical memory and switched devices .

With overcommit,malloc() You will be successful , Otherwise failure , Developers can allocate memory generously and use it sparsely on demand , Instead of taking care to distribute .

7.2.5 In exchange for process swaping

Process swapping is the movement of entire processes between main memory and the physical swap device or swap file.

Move the entire process between main memory and the physical exchange device or exchange file .

7.2.6 File system cache usage file system cache usage

Process swapping is the movement of entire processes between main memory and the physical swap device or swap file

7.2.7 Utilization and saturation utilization and saturation

Main memory utilization can be calculated as used memory versus total memory.

The memory occupied by the file cache can be treated as unused , Because it can be reused by applications .

When the demand for memory exceeds the main memory, it is called main memory saturation , Now OS Can use page feed 、 Exchange or in Linux Use in OOM To free up memory .

7.2.8 distributor allocators

When virtual memory handles multitasking physical memory , The actual allocation of virtual address space and memory stacking are usually handled by allocators .API Interface is malloc(),free() And so on .

7.2.9 Shared memory shared memory

Whether the shared memory between processes should be counted into the total memory occupied by a process ？proportional set size (PPS) Solve the problem , Contains private memory （ Unshared ） Separate from shared memory .pmap Tools .

7.2.10 Working set size work set size

7.2.11 The word is long word size

7.3 framework

7.3.1 Hardware

Main memory

Time delay

Time available to access main memory CAS Time delay （column address strobe, Column address controller ） To measure ： Send the address to be read from （ Column ） Give a memory module , Time until data can be read .

Main memory architecture

UMA

Above picture , Through the shared bus , Every CPU Access memory has uniform access delay . If a single operating system instance is running above and can run uniformly on all processors , Also called symmetric multiprocessor architecture （SMP）.

NUMA

CPU1 It can be directly connected to... Through its memory bus DRAM A launch I/O operation , This is called local memory .CPU1 adopt CPU2 as well as CPU interconnection （ Two hop ） Yes DRAM B launch I/O operation , This is due to higher remote memory access latency .

Bus

Shared bus	shared system bus	Such as UMA
Direct connection	direct	processor Direct connection memory
interconnection	interconnect	multiprocessor in , Every processor Directly connected to the respective memory , The processors are connected by a CPU Interconnection connection , Such as NUMA

DDR SDRAM

For any architecture , The speed of the memory bus often depends on the memory interface standards supported by the processor and the motherboard .

1996 Year of double data rate synchronous dynamic random-access memory(DDR SDRAM)( Double rate synchronous dynamic random access memory ).

Multichannel multichannel

Multiple memory buses in parallel , Increase the speed . There are usually two channels （dual-）, Three channels （triple-）, Four channels （quad-）. such as inter-i7 The processor is a four channel DDR3-1600, The maximum memory bandwidth is 51.2Gbytes/s.

CPU cache

L1/L2/L3 cache .

MMU Memory management unit

Manage address mapping from virtual memory to physical memory .