Will IO always occupy CPU? A good question about concurrent / parallel systems (turn)

This is a good one about concurrency / Problems with parallel systems . The simple answer is ：IO The required CPU Very few resources . Most of the work is assigned to DMA（Direct Memory Access） Direct memory access Accomplished .

Let's not talk about traditional 5 Big IO Model , Let's talk about concurrency first （Concurrencey）. A very loose explanation is to do at the same time A and B Two things . First do the process for a while A, Then context switches , Do it for a while B. After a while, I'll cut back and continue to do A. So it gives us an illusion , We are doing at the same time A and B Two things . This is the famous process model .

It looks cool , But in fact, there are no eggs . because A,B You have to finish both things, no ？ Whether you finish A Do it again B Or switch back and forth , It should take the same time . Even more , Because we also need to consider the overhead of context switching . So when I first learned the concurrent process model , There are 10000 white eyes in my heart .（ I feel the same way ）

however , If there is more inside the computer CPU A part is working ？ If A This matter CPU It can be assigned to other parts to help it complete ？ Is the situation completely different ？ System IO It happens to be such a perfect example .

For disk IO, The real scene may be like this ：

CPU say ： Brother hard disk, help me copy a small movie I want to see to main memory , thank you , Pro - .
 Hard disk said ： well ！ I'll call you after the exam .
CPU say ： mua ！ Then I'll play games ！
...
CPU Fight and fight  （100 Nanoseconds have passed ）
...
 Hard disk said ： Small C I finished the exam .
CPU say ： Teacher Cang, I'm coming ！

Of course, we can also download teacher Cang's works online , This is the Internet IO. But the situation is basically the same ,CPU Children's shoes are waiting for a little movie , Played a round of roll, roll .

therefore , It is precisely because of this assignment , Communications , Waiting process , Concurrent system shows its significance . Of course, the actual process may be 10000 times more complicated than this . such as CPU I won't talk to the hard disk directly , There is a middleman between them , It's called DMA（Direct Memory Access） chip

CPU Calculate file address  ——>  delegate DMA Read the file  ——> DMA Take over the bus  ——> CPU Of A Process blocking , Hang up ——> CPU Switch to B process  ——> 
DMA Notify after reading the document CPU（ An interrupt exception ）——> CPU Switch back to the A Process operation file 

This process , Corresponding to the figure below （ Picture source ：《UNIX Network programming 》）, notice application Is this a timeline ？aio_read After the operation , It's all blank ,CPU No matter the , You can do other things .

Suppose the file was originally read CPU Need to wait 50 nanosecond . Now, although the two context switches need to be consumed separately 5 nanosecond .CPU Still made money 40 Nanosecond time slice . You can see from the above picture , What I just talked about is tradition 5 Big IO In the model “ asynchronous IO” The general process of . Want to know more about , It is recommended to read directly 《UNIX Network programming 》 Volume 1 sockets , Classic ！
although IO It won't take up a lot of CPU Time , But very often IO It will still be very wasteful CPU The time of the , So face a lot IO The task of , Sometimes you need an algorithm to merge IO, Or by cache To relieve IO Stressful .