Schedulable lightweight virtual processes and concurrent loops P&L: -11.5 (≃ -138 USD)
Imagine a nested for loop, it is composed of multiple loops. We can simulate concurrency by executing parts of loops and rapidly switch between them.
This is what a kernel scheduler does and I think it's a very interesting part of an operating system. Unfortunately most programs are not written to be multithreaded so computers use very little of resources available to them.
Rewrite Python code in Java so true multithreading can be used Implement loop runner threads that repeatedly try execute Nth product that matches that thread number
2 hours I am fairly happy with what I created. I think the API could be better. Or the implementation of waiting for results could be better. The calculation to see if there are any new results is expensive due to nested loops. I think this could be done better. Ideally nested loops should only take a few divmod instructions and additions. Otherwise they shall dominate performance used.
2-4 hours I cannot remember when I began
I get duplicate entries.
The parent loop looks this:
for letter in letters: for number in numbers: for symbol in symbols:
Then I try extend it with a new concurrent loop inside the innerloop.
But each iteration would look the same! It's not equivalent to this:
for letter in letters: for number in numbers: for symbol in symbols: for nested in codes: for emoji in nested: print(letter + number + symbol + nested + emoji)
I need some way of extending the loop through an arbitrary number of loops, like a tree of loops.
I spent 1 hour on this. I wasn't really in the mood for development but pushed through.
1 hour
I realised my bug in my logic. Thread 0 is responsible for loop iterations 0-8 Thread 1 is responsible for 8-16 Thread 2 is responsible for 16-24 And so on I'm currently executing loop iterations on the wrong thread due to a mathematical logic of not handling wraparound. I use the modulo operator. Ideally each ticker is only called on the thread for its range.
I have a problem I need to solve. Which isn't related to the parallellism.
Imagine I have a nested loop that looks like this
``` For letter In letters:
For number in numbers:
Print(letter + number)
For symbol in symbols:
Print(letter + symbol) ```
I want these three loops to run concurrently. One approach is to pick the function to run based on the indexes! And merge the numbers and letters together round robin. So we get
A1
A÷
A2
A×
The problem with this approach is that the loops aren't separate. They're one loop that has been merged.
I think I can solve this by causing collections to be a multi collection. [Letters, [numbers, symbols] And picking from each sublist round robin and exposing the loops as separate objects.
I created a multithreading version of this without the Joinable loops. But at this time I am taking a break. I need to somehow tie the multithreading to the joinable loop. When Joinable loop received 2 or more values, it continues processing. This allows a split and merge in pipeline processing. I want joined processes to occur as soon as they can and to be multithreaded. In my design - that I am yet to implement in Java - each thread has a scheduler that rechecks what can be ticked and ticks it. The problem is splitting ticks over threads, it's easy if there is only one level of tickers. You can just tick batches of 8 at a time. In separate threads. My example has a nested loop of 3 collections and a split into two tasks and then those two tasks join into one task to print out the results. I want the nested loops, separate tasks and the joined tasks to be executed in batches. Kind of need a way to send concurrent loops to old threads. Could have a shared collection called the tick pool which is checked by all threads. The current product number is associated with a thread. I picked ÷ 8 due to the 64 being a multiple of 8. Clean number of batches per thread.
```
While (true) {
For loop in tick pool:
If current[loop] < thisThreadN[loop]:
current[loop] = current[loop] + 1
NewTickersOrResult = Loop.tick()
If NewTickersOrResult.isTickers:
Else:
}
```
I am thinking of how to use the performance of multiple CPU cores. It requires a drastic rethinking of how we write code!
It occurred to me that loops could be trivially parallelized with my design.
N = 0 .. N = 3×3×3
If you ran the tick method on all threads with every N, you could run the entire loop in one go.
I need to write a Joinable Ticker that waits for inputs from multiple tickers before sending output along. This lets us create pipelines that split and merge.
I shall add that you only need one while (true) loop per thread. Everything else can be concurrently scheduled within a thread.
That's why I call it virtual concurrency. You would need to use threads to provide IO concurrency Anything that loops forever or blocks cannot be composed. So I shall write all my code going forward to try never block and be reetrant. This is an important lesson I learned while developing my multithreaded scheduler.
Blocking is invisible to the program. The program isn't aware it is blocking. You need to know that a method can block to work around it. My other idea is a parallel while do loop which changes blocking to be non blocking through syntactic sugar. It looks like this
A1 = parallel while { Buf = socket.recv(1024) } Do { Parallel for (Socket socket : sockets) { Socket.send(buf) } } Crossmerge A1
This syntax spins up multiple threads to block on receiving data. And for each one it spins up a thread in a thread pool to handle it and send it to every connection in parallel.
Another idea I have is to change the priority execution order. That scheduler in that code round robins the tickers in the same order every time. There's no reason why we cannot execute the loops some more than others.
Interesting. You did not use any native support for concurrency in Python, and used only basic enumeration, indexing, assignment. I guess, that is what it takes to understand. This has educational value.