National Energy Research Scientific Computing Center (NERSC) Codee Analyzer Training, April 2022, 12 May 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: 08 Putting it all together in real codes using LULESHmk

Description

No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).

A

So, finally, in order to try a first half a first sense of the complexity of real applications, we have selected the new snk example and we proposed just to produce the open scc version, because we have seen that openhcc typically produces a bit faster code at openmp for the same purpose, so to simplify.

A

Let's use openscc as a reference for this more realistic code. Example lulac electrode, dynamic application from physics. In the end, it's all about the same. Let's locate the profiling, which is the hotspot function here you can see. The code consists of more than 11 functions, not one single one, and you have at least four functions that consume more than 10 percent of the runtime. So eventually or potentially, you will like. You will want to optimize the performance of these four functions.

A

Again, you need to know how to verify correctness. There was a question about this yesterday in p, in pi is simple, just one floating point number in matrix matrix multiplication, let's compare element by element and see that the error is not greater than a given threshold, but real applications.

A

Typically, measuring correctness is not that simple and it's it requires knowledge about the scientific domain and producing by the scientists and the computational scientists, some figures of merit or some metrics that we can use to verify that the code is correct. Okay, so verifying correctness is not easy, but here in lulac we just need to check that it is producing typically, this same output that we can see here.

A

uh We have seen this so essentially what we propose you from the real application. One file 15 functions, almost 20 loops, 500 lines of code, something a bit bigger than four lines of code in one single loop in one function in the pi example.

A

So here you will also be addressing the complexity of real applications. You have a loop that calls and other functions and also it is not a straightforward, massively parallel computation, like pi or math, matrix matrix multiplication.

A

Here we have something that we call sparse reductions, but in the end it's a type of computation that appears, if you are using molecular dynamics evaluating finite elements, finite differences, particle methods, you will find this kind of computation in the code. That is having indirections on an array that is a reduction array and you need to protect those interactions from introducing risk conditions and runtime.

A

This is what we call sparse reductions. You have more extensive documentation about these compute patterns that are at the heart of kodi in previous uh materials published in in nurse website. We can point you to those materials if you are interested, but it should be enough to use the current capabilities of kodi. It should not be necessary for you to dig into the theory or the or the rationale behind these computation patterns and starting from this, let's follow the same path.

A

Report evaluation, enable gpu and multi-threading focus on the loops, particularly in the low hanging, fruit, low difficulty, loops actions, level. Two actions you can see here. These sparse computation patterns that are highlighted here and finally, just following instructions open a directives will produce the code. You will see that the sparse reductions have been optimized for parallel execution, protecting with atomic instructions, the the indirections to avoid risk conditions, and also you will see that kodi produces incomplete, open sec code.

A

This is helping you to really know how to fill in some information that not may not be available through static code analysis. Okay, in particular, you will need to fill in some data.

A

The data is pointed out here, so here you have the template that is produced these brackets columns brackets. If you try to compile this openhc application, that opens the compiler could report an error. This is intentional so that you know exactly where the error is, and you can fix it just by introducing the ray shape. Lmk is using only 1d array, so it should be fairly just the maximum size of the vectors. These are 1d vectors.

A

Once you introduce that information make those changes, you will see how mvc, if you enable minus capital m info flag, you will see all the information reported by the compiler, how it used this information to implement the data transfers and to actually manage the complexity of the function, calls through implicit, open, fcc, routine, sequential to annotate the functions, and this is done automatically by the nvidia compiler.

A

Finally, these are the numbers, and this is what you get so here we are seeing that the code runs correctly on the gpu. We have the first starting version for the gpu and, as usually in the first version running on gpu, it doesn't run faster. It runs more or less with the same performance or a bit slower, because remember for real applications is not enough to handle this. You will need to address additional tpu challenges that we will see in the more advanced courses.

A

Okay, so we expect this lewis mk example to be further optimized in the upcoming course in september and october. So this is the typical situation you you will see in when starting on the gpu with real applications. Okay and finally, there's the typical remarks remember to load the module help. Is your friend all the command line tools run the scripts in the appropriate reservations and to understand the issues. Please review read: go through the knowledge base, a lot of information illustrated by example, is there and that's it?

A

That's what I wanted, or we wanted to cover right after the break in just 10 minutes or 15 minutes any questions or any remarks. uh Helen about the this example codon. What comes next free hands-on for the people to play with lulu's mk finish the previous laps or why not try to start with their own application.

B

Yeah a free, hands-on time and mostly to work on this lurch codes and not just blindly apply those directives, but look into the code to see what it has done to the code and then, if you wanna and work on your own code, that's also a good opportunity, because the developers are here to answer your questions, good um check out the q, a I like one of the questions yesterday, it was um answered today whether this code is unique.

B

I think I, like this code, just my my opinion. um First of all, it helps novice programmers to to just insert all these directives. For you, this is a big help uh on other. One of the other tools I know about is the create review tool that one has you have to use a create compiler only and then, of course, after the general, you can use any other compiler after that, um but for the kodi tool uh you can just run it on a logging node with I think and backhand.

B

It does gcc and the static cone analysis, but it's it's it's simpler, so that you're ready to get the directives inserted for your code.

B

The other advantage of this tool is that it has lots of knowledge, base and recommendations of performance, optimization suggestions so take good advantage of these. I think that's also a big plus. I hope our attendees get learned something and apply this tool and thank you very much for the training for us, the petra team. The tour is now called kodi. The company is a pentra, it's a rebranding.

B

So if you have used this before it was a pw analyzer tool, I think- and I I appreciate all the users attending um the training and maybe spread the knowledge to your colleagues and friends, and thank you for nerds and oakridge users and organizers, suzanne and we've been here today.