Previous | Next --- Slide 18 of 46
Back to Lecture Thumbnails
wcrichto

C++ (or pseudocode in this case) can be kind of ugly, so it's nice to look at how channels/message-passing is implemented in other languages. In particular, I find Go and Rust to have quite lovely support for this sort of things.

Channels in Go:

channel := make(chan int)

go func() {
for {
fmt.Println(<-channel)
}
}()

i := 0
for {
channel <- i
i += 1
}


Channels in Rust:

let (port, chan) = Chan::new();
do spawn {
while (true) {
println!("{:d}", port.recv())
}
}

let mut i = 0;
while (true) {
i = i + 1;
chan.send(i)
}

mpile13

My understanding of this code is that each processor iterates over the part of the array that they are assigned to, and there is some communication between processors so that each processor sends its top row to the processor above it and its bottoms row to the processor below it, and some more computation is performed.

(I'm a little bit fuzzy on what happens while the processors are exchanging rows. I know they become dead-locked while waiting for a response, but I'm not 100% sure why they're exchanging rows in the first place)

Then eventually every processor sends their results to Processor 0, which adds up all the results and sends a message to all the other processors saying that the computation is finished.

Because information is only shared through sends and receives, there are implicit locks around the parts of the array each processor is working on.

If anyone wants to fill in anything I'm missing, I'd appreciate it.

mchoquet

The reason they need to send rows around is that the value of each point in the grid is a function of its four neighboring points (a weighted average maybe, or else something more complicated). If we're dividing work up by rows, then the last row a processor is responsible for computing depends on the values of the row after it, so the processor has to get the values of the row after it from the next processor in line. That's where the message-passing comes in.

kayvonf

@mpile13: This illustration on slide 17 might be helpful to understanding why the previous and next rows are communicated.

Question: Why is the program sending entire rows in a single message, and not sending one message per grid element?

Q_Q

Maybe the program sends entire rows because that minimizes the amount of overhead caused by sending a message? Or is there something deeper than that?

rokhinip

Sending a message, as we reviewed earlier is expensive since it requires copying information over from the send buf to the network buffer and again to the receive buf. Not to mention that if it's blocking, we also have latency as a result. So we want to maximize the information we can send as a message.