How to set progress bar for parallel_invoke? - c++

Question

What I want is how to get the progress data. I can implement the bar whatever I like. And I'm using Visual C++ 2010, so I can use MFC.
Now, I'm writting multithreaded program. And Microsoft already provides PPL lib since VC++ 2010. The Parallel Patterns Library (PPL) provides algorithms that concurrently perform work on collections of data. It's convenient to implement multithreaded app, but I encounter progress bar problem.
How can I set progress bar for parallel_invoke?
The demo code is as follows:
// parallel-invoke-structure.cpp
// compile with: /EHsc
#include <ppl.h>
#include <string>
#include <iostream>
using namespace concurrency;
using namespace std;
// Returns the result of adding a value to itself.
template <typename T>
T twice(const T& t) {
return t + t;
}
int wmain()
{
// Define several values.
int n = 54;
double d = 5.6;
wstring s = L"Hello";
// Call the twice function on each value concurrently.
parallel_invoke(
[&n] { n = twice(n); },
[&d] { d = twice(d); },
[&s] { s = twice(s); }
);
// Print the values to the console.
wcout << n << L' ' << d << L' ' << s << endl;
}

Answer 1

Maybe this is off topic, but I would recommend checking out Cilk instead of PPL.
Here is a course at MIT (from one of the main creators of Cilk) that outlines its pros, cons and specific usage. It's very easy to use and claims to be much more performant than PPL (in case you're concerned about that).
As for your specific question about progress bars, coming up with a percentage is the first task. If you use timers to map which work costs you what (in terms of wall time), you can use that (via a callback from each worker to register work done) to drive the percentage. Once you have your percentage down, hooking it up to a bar (in the shell) or a GUI progress bar is trivial.
Depending on how much work your worker processes do (how granular they are), you may want to consider different methods of instrumentation. If you look at this wikipedia page, it'll outline many types of profiling (statistical, event based, etc.).
One last note: I find that progress bars are difficult to get right. Anyone can make a bar move. But having it accurately represent a true indication of how much time it will take involves a great understanding of what factors are affecting your performance. As is the case with any computer program, your performance can change based on which system your run it on (cpu architecture, IO bandwidth, network bandwidth). So even if you get it predictable on your box, your clients' boxes may be drastically different--so much so that your progress bar doesn't yield good indications anymore.
Hope that helps...