Page 1 of 1

Graph Manipulations Optimal Memory Transfer

PostPosted: Fri Oct 21, 2016 9:13 pm
by dekanutepederaz
Hello,

I am doing operations over a graph with the Parallella board. Currently, I have a working algorithm which colors the graph, but with a very slow speed. I think that the memory transfers this, so this is how I use the memory.
- I insert the list of neighbors and a table V*V in the shared memory from the host
- then I use read from the shared memory each time from the eCore when I need that information; most times, the neighbors and the square table are too big in order to be fitted into the local memory of the core
- I generate a solution and read it from the host from the core memory

First, is this a good memory distribution?

Second, I think that the second operations slows the process the most and would it be better if I use DMA for this?

Third, would it become faster if I do not stop the core each time I stop using it, but instead have an endless loop?

Thank you in advance!!!