Network IO, disk IO

The Internet Io, Sending a network request is common 1 Milliseconds to hundreds of milliseconds . disk io, commonly 5 Milliseconds to hundreds of milliseconds . Two kinds of io Every time the time is not fixed , Even if you take the same data . The Internet Io There may be network fluctuations leading to .

io intensive .cpu Keep making requests , It takes tens or hundreds of milliseconds to wait each time . After getting the data, the processing speed is 1 The subtlety is solved . So most of the time cpu All idling .

One revolution of the local disk is about 5 millisecond , If the data taken is large , Just turn around a lot . There is a lot of time .

before cpu Fetching data from memory is 20 nanosecond , Now it has been raised to the millisecond level . A million fold increase in time .

How to improve cpu Utilization ratio , For example, the network requests a large amount of data ,1 copies . Then create 100 Threads , When the first thread makes a request , Switching thread . Other threads start sending requests in turn . Don't open too many threads . Thread switching required 1 Haos second , It also takes time , And when switching cpu Not working , Threads also increase memory overhead . Creating a thread generally considers cpu The number of nuclear ,cpu Task execution time vs. wasted time . The bigger the number is , The fewer threads you need to open . And because of the thread switching time , So the final result should be tested .

10 Ten thousand data , open 100 Threads , Every thread 1000 A mission . Some threads are fast , Some threads are slow . The fast thread is destroyed after it completes the task . There are many final tasks , Fewer threads . amount to cpu Performance is reduced again . So this allocation of threads is flawed .

What we want is that threads always exist , Then the efficiency will not be low . You can do this , Just don't divide the tasks equally , Each thread handles one task at a time , When it's done , Just go here 10 Take tasks from ten thousand data . Until the task is processed . This is the judgment condition of each thread 10 Whether 10000 data are empty , When it is empty, the thread is destroyed . The thread pool is just used . Thread pools help us manage these threads .

Thread pools generally use bounded queues , This can trigger thread resizing . Then the number of threads in the thread pool when there is no task is 0, When dealing with a task , The number becomes the number of core threads , When the task is finished , Nor will it fall to 0, It also reduces the number of core threads . Then there is the thread pool .

Unbounded queues can be used if there is no time requirement . Like web crawlers , Data processing