7. Races and Parallelism

[@leixun]

*My takeaways:* 1. Determinacy races 4:05 - Two types of determinacy races 10:26 2. Avoiding races 12:10 3. What is parallelism 18:15 - Computation Dag 24:10 - How much parallelism we can get 26:35 4. The Cilkscale scalability analyzer 42:25 5. Scheduling theory - Greedy scheduler 53:25 - Cilk work-stealing scheduler 1:02:50 6. The Cilk runtime system 1:04:27

[@abdulshabazz8597]

@00:23:42 A pleasing representation of the Fibonacci algorithm & branching semantics

[@tomekczajka]

13:45 This is not true, C11 specs guarantee that modifying x.a and x.b in parallel is fine.

[@juanmamani2110]

lo hicieron pebre al pobre teacher MIT, pasando a todo nivel.. Pero hay que darle su crédito en la lecture.

[@gregchen4023]

In 1:07, he actually does not know cilk at all. Actually in java library there is similar implementation inspired by cilk, so I understand the theory. Each deque stores the tasks(refrences to the function calls and the arguments that need to be passed to the functions), they are waiting to be spawned and be executed. A spawned task is not in the deque, it has already be fetched or stolen from the dequeue. Here spawn and execution of the function call are the same meaning. Once a taske is spawned, no core can steal it because it is already in execution on a core. It can only be stolen when it is in the deque. There is a dedicated thread responsible for management of the set of deques

[@davidprock904]

I'm sorry teacher but I cant submit my homework because it's soo good, I need to get my intellectual property patented first. No really, I've come up with something that is going to change EVERYTHING!

[@jameschichi59]

Youve lost it havent you mate?

[@davidprock904]

@@jameschichi59 no I haven't, still fine tuning it actually, once in a blue moon, I come up with an improvement to it

[@firasismail69]

any updates mate?

[@gregchen4023]

I would like to ask the teacher himself whether he understands what he says, as I do not . I think the teacher does not understand why Tp > = T1/P. I think it is because the ideal case is that every core gets exactly the same amount of work evenly, which is T1/P resulting all the cores finish work at exactly the same time, which is Tp = T1/P. But in fact some of the cores get more workload and some get less(the sync command is required for a core to wait for another one to finish if it gets less workload than the other). When all the cores finishe the work respectively , it should be more than T1/P because of the waiting time for the sync commands.

[@poketopa1234]

He's talking about in theory, not in practice

[@davidprock904]

What if your computer architecture is soooo different, by definition your computer doesn't have a processor/ cores, threads, pipeline, stacks, etc... but it can have unlimited level of Parallelism?

[@shivanshmishra752]

If it has a 1 processo and your program has unlimited parelism it it will just execute it sequentially

[@gregchen4023]

48：00--Does the instructor understand the quick sort? The partition is not done sequentially, it is always split it into 2 partitions in the middle of the array.

[@kiki-yg9nv]

it's not

[@poketopa1234]

He explicitly states that the algorithm is randomized quicksort

[@gregchen4023]

The speech is so messy, because he is teaching something he does not understand.

[@spartacusche]

are you sure he does not understand?