Looking back at this note in 2022, it’s amazing to see how time flies. I didn’t have the formal training in software engineering and had to learn about such concepts by myself through books and articles. It was tough but fun!
What is a thread?
According to Beginning Linux Programming, “Multiple strands of execution in a single program are called threads”.
A strand of execution in a single program can be thought of a sequence
of controls that the program operate on its memory, or variable. Think
of a simple “Hello World” program that asks user for an integer
N “Hello world!” to the screen. The sequence of controls
is relatively simple:
- define a new integer variable;
- get a value from user input, and assign it to the integer variable;
Ntimes and print out the sentence ach time.
The “multiple strands” are then interleaving sequences operating on
the same memory space, or the same set of variables. For instance, one
thread can ask user for a value
N, while the other thread do some
computation on this input and assign the output value to yet another
variable. A third thread then periodically look at the value of this
variable and print to the screen.
Synchronizing multiple threads
The question of synchronization can be stated informally as to how to make multiple threads work nicely with each other? Alternatively, what can happen if there are multiple threads in a single process and how can we address these issues?
Problem number 1: Conflicts in accessing a common resource
One issue with multi-threaded programs is handling the conflict between different threads in accessing a common resource. In another word, there are variables that are accessible by all threads, and if two threads attempt to change the variables’ values at the same time, the end behavior is undefined. This is undesirable.
The universally accepted solution for this problem is actually quite straightforward. And yes it is mutex, if you are wondering. But it is important to understand, not what is a mutex (it actually is a simple integer-valued variable with atomic operations), but how to think about it.
Let’s be more specific. Suppose now you have a set of variables that might be accessed by several threads. Near this set, there is a switch. (Yes, a hypothetical switch!). This switch is initially on. Now, we will make the following rule:
- any thread that wants to read or change the values of any variable in this set of variables must wait until the switch is on;
- if the switch is on, turn it off, then do your stuffs on the variable;
- when the editing is finished, turn on the switch and continue.
If the switching mechanism can be made such that in step 2, “check if it is on, then turn it off” is atomic. That i,s if there are multiple threads attempt to run this atomic instruction at the same time, it will always happen sequentially. Then whenever a thread manages to turn off the switch, it can freely access the variable without worrying about undefined behaviors!
Notice that here the switch needs not be associated with the set of variable. Rather, the threads must be designed such that they respect the switching rule. Whenever a thread wants to access a common resource, it should make sure to check the switch before editing.
Now, the switch is basically a mutex. Turning on the switch equals lock the mutex, and so on.
A common question: if a thread lock a mutex, can another thread unlock it?
Answer: The answer is: Yes, but you shouldn’t do it. Let’s return to the switch analogy. Suppose a thread has just turned on the switch and is doing its editing. Clearly, another thread can just turn off the switch, then turn on again and do its editing. Better yet, it does not even need to care about the switch at all: just go in and edit. Now, while turning off the switch is possible technically, it is a serious breach of the switch-rule, making the whole switching strategy fails defeating the purpose of the synchronization strategy in the first place.
Problem number 2: Synchronizing sequence of computations
A second problem in multi-threading is making sure that the threads process data in order. As an example, consider a simple computational graph:
A ---> B1 ---> C
\-> B2 -/
Suppose that we have a very complex computation that involes several steps. First, the main thread A receives an input from users. It then processes the input and sends the result to two threads (thread B1 and thread B2). After both threads have finished, another thread (thread C) takes the results from thread B1 and thread B2 and do some touching up.
Clearly synchronization is required between the computations.
To achive synchronization, a possible strategy is as follows:
- A does her computation, and finishes it;
- A waits for B1 and B2 to come over and take the result;
- B1 tells A: “hey, I got it, you are done with this batch”;
- B2 tells A: “hey, I got it, you are done with this batch”;
A starts working on a new computation.
- B1 and B2 starts computing stuffs.
- When they are done, they let C know; only when C has confirmations from both B1 and B2 it starts its computation;
- After C1 takes the results, B1 and B2 return to A to get things to work on.
This strategy is precisely what can be achieved with semaphores.
Questions and Answers
If I have two threads that share the same variable, one thread only read the variable value and does not write, do I need thread synchronization?
The short answer is: yes, synchronization is still needed. The reason is not because when two threads access the variable at the same time the net behavior is undefined. This is not a problem because reading operation is “atomic”. The main problem is that the compiler might rearrange the code before compiling, and the mutexes and semaphores are needed to create memory barriers to prevent this re-organization.