In the chaos of the lock-free revolution, and the return to control under the dictator of mutually exclusive ordering, and, finally, the aristocracy of fine-grained and read-write-based bureaucracy, the citizens have grown discontent. The shared memory of society is powerful, but we require some means for providing citizens an easier way to interact with it! The final nail in the coffin was the specter of deadlock. Normal citizens can no longer manage and rely on the deep bureaucracy of the government.
There is hope!
The aristocracy of fine-grained and read-write locks are ceding power to a democratically elected representative! Among those candidates are the exotic -- peddling rare animal and tiger zoos that have free publicity from Netflix, and professional fake tanners that distinguish themselves with an orange glow and a dislike of those that are not orange, and most importantly, a most pawticularly adorable candidate.
Penny's running on a non-traditional platform of hiding the complexity of locks from the citizen application! Her key initiatives:
- Pets-care for all.
- Shed the red tape of mutual exclusion and fine-grained locking.
- forkjoin parallelism for freedom
- mapreduce processing toward a better society
Forkjoin parallelism enables data parallel access to data.
Instead of a normal fork loop in which each iteration executes sequentially, what if we could execute each iteration in a separate thread, on a separate core!?
It is called "forkjoin" as when we hit the forkjoin in the code, we "fork out parallel computation", and when the "loop" is done, we "join together".
Note that read forks are not used here, and the word "fork" means something close to a "fork in the road".
Question 1: Will this work for any program, and any fork loop? For which will it work?
Mapreduce is a technique common in distributed systems whereby we define a set of functions, most notably map and reduce.
Question 2: Have you heard of these functions from other languages? What do they often do?
Mapreduce will call the map function on each item in the input, thus generating a set of output items. Then we will combine all of these to get a final results, a process called "reducing them".
Question 3: What types of problems could be computed using these primitives?
Read through main.c and answer the following:
Question 4: What is more complex about the forkjoin parallelism versus mapreduce?
- don't need the context/memset variable
Question 5: What is more complex about the mapreduce parallelism versus forkjoin?
- requires more helper functions
Question 6:
Run sh test.sh.
Anything surprising or unexpected?
- take about the same time instead of one being overly better than the other, even with a large data set
Question 7:
Check out the implementation of forkjoin in forkjoin.c.
Can you think of any way this could be optimized around the thread operations?
- when we wanna gather all data back together, it needs to be sequential. Heavy duty computation will be very expensive. Reduce looks complicated. Specified so we can run multiple reduce's at the same time. Problem for in parallel: trying to modufy aggregate in parallel
Penny promises bringing parallel programming to all! Unfortunately, she doesn't provide abstractions that everyone can use for every problem. But she does bring abstractions that everyone can understand, and can use without the pains of history -- locks, deadlock, and lock-free programming.
Higher-level primitives for parallel programming exist, but they are all limited in their programming model in some way.