问题
So I am working on a C++/cli image processing library and am trying to optimize my code. Basically, I am passed a System::Drawing::Bitmap of the image, which I then need to write to disk, perform complex analysis on, and return the results of the analysis. I thought that I could write the image to disk in parallel to speed up things (my algorithm does not modify the image). However, I have not worked with threads much, so I wanted to get your input on what the best way to do this would be.
string ProcessImage(System::Drawing::Bitmap ^bmp, System::String^ targetFile)
{
bmp->Save(targetFile);
System::Drawing::Bitmap^ bmp8 = BitmapConvertPixelFormat(bmp, 8); //<-- a function I wrote which converts the 32bpp I am passed into an 8bpp one
string results = Analyze(bmp8); //<--- takes a good bit of time
return results;
}
Please let me know your thoughts. Thank you in advance!
回答1:
Writing/reading in parallel to/from a single mechanical disk is not a good idea because the mechanical head needs to spin every time to service an I/O request, so using multiple threads will just bounce it around needlessly and create overhead.
You can try to benchmark a bit, but I'm afraid you'll just have to resort to using a single thread and writing sequentially.
回答2:
Queueing off the disk writes to another thread seems like a qood idea, but only to one writer thread per disk so that the complex analysis can run on without the slow disk-writes holding it up.
回答3:
If it's worth it you can invest some time into understanding how to use Parallel HDF5. It can write to a file in parallel.
来源:https://stackoverflow.com/questions/10254147/writing-data-to-disk-in-parallel