Scatter tutorial - Supercomputing and Parallel Programming in Python and MPI 9

Key Takeaways

hello everyone and welcome to another MPI for Python tutorial video for high performance computing supercomputing and parallel programming so in the last video we covered broadcasting now it's how we could take one variable or object and pass it to every other node in the system so we could share that variable amongst everybody so that was broadcasting I think that was fairly simple to understand next up we have what's known as scatter and gather so what scatter is gonna do is say in the master node we have a group of objects so let's say we have a list what we can do is we can take that list and you do scatter and it's gonna take that list explode the list and send it in parts to each of the nodes it's gonna divide up that list and send it to each of the nodes and then what gather does is basically the opposite of that so it's gonna take all the objects from all of the nodes gather them together and go to the target which usually will be the master node so with that let's go ahead and get started so as usual what we're gonna want to do is from MPI 4 equals MPI comm underscore world size equals get size man all these upper cases are killing me break equals calmed I get underscore rank okay that's all we need for this so then what we're gonna have is just the same kind of thing that we've had for quite a few of these is just when the if rank equals 0 what do we want to do we're gonna say data equals and then we're gonna make a generator here and generate a dynamic list of all of the ranks we could make a simple one but like we've got hard code it but that's not a good idea anyways data is going to equal this generator and what we're going to do is we're gonna say data equals x plus 1 to the power of X for X in range of size for X range even if you want it so that's what data is gonna be so basically if data is gonna be X plus 1 to the power of itself for each of the X's in the range of size which is referencing how big is this how big is this supercomputer in nodes this just basically makes sure that we will have one object per node it's just you know some sort of just arbitrary object that we're gonna give it so anyways mm-hmm once we've done that let's print out what we have just so you understand so you can see what data actually represents just in case you're not totally following what's happening here so we're gonna say we will be scattering and then data simple enough next we're gonna say else data equals none and you should kind of recognize this from before and then we come down here we say data equals calm dot scatter what are we scattering well data and from where that's gonna be Ruth zero now what we're gonna do is print rank rank and data okay so let's save that let me move it to news and then we'll come over to the master node up arrow and now I've changed mind s89 hit enter and again we're using five here Oh what have we done oh we've typo'd fantastic why didn't anybody tell me okay import okay so here we have we will be scattering and then it says we're gonna be scattering a 1/2 964 and 625 so you could hard-code on just a list right of 1/2 964 625 or whatever variables you want but you can see what it's done here is it scattered this stuff so rank 0 has data what data did bring 0 get it got the 1 Rank 1 got what data the 2 ring 2 got the 9 and so on so that should be a pretty easy to understand example of what scatter does it's just gonna kind of take everything and go picking just divert our divers just diversify it but at the same time I guess the word I'm looking for is scatter what if we did for X in range of size +5 right so what if we made that size a little bit bigger here Barrow let's see what happens so what happens in this case is in order to have scattered this stuff we wanted to be able to send multiple variables to some of these right so we were expecting only 5 and we wound up with 10 now first off I wouldn't be surprised if there was some way in this scatter call are the actual scatter command to scatter multiple at the same time you could make the object itself a an array right we've seen that we can pass an array or I guess it's not really a you could pass a numpy array or a Python list but anyway you'll want to just keep it in mind that you can't send more variables then you're ready to accept or more objects then you're ready to accept so if you're going to do this you would need to first divide it up by size so you would take the list maybe div by size in them say put X many variables and something like that you can even use modulo to round it off so let's say you had nine nine objects that you wanted to pass you could use the modulo to put everything in in one group and then modulo the rest so something like that anyways I might have a more in-depth example of scatter later I'll have to look through the documentation that would make sure there is an already like a prebuilt command within scatter that allows to do this before we go coding our on so anyways that's gonna conclude this tutorial video hopefully you guys enjoy in the next video what we talking about is gather so anyways stay tuned to the next video as always thanks for watching thanks for all the support and subscriptions and until next time