Iterating host code: Parallella restarts

[gordon]

[sebraa]

Open the Epiphany once before the first iteration, and close it after the last one. Opening/Closing the Epiphany repeatedly may not work (or may stop working after a while). It is better to structure your code so that it can do multiple runs without a full system reset.

[gordon]

[sebraa]

[gordon]

I can't iterate the cores itself because in my case, there are two different device programs to be executed on two different set of cores i.e the diagonal and non diagonal cores. Unless the diagonal cores program finishes its execution I can't start the other program because it takes in the values returned by the first program. So for the second iteration to start on the non diagonal cores, I somehow need to run the second iteration of the diagonal cores first and then send the values to the non diagonal cores via the host and then begin the execution. Can you help me in this direction?

[sebraa]

[gordon]

Actually, after one iteration the host code appends all results into a single 8X8 matrix and then there is swapping of rows and columns.The new matrix is again split and then written to the 16 cores. So actually if we take the example of a diagonal core, the result of first iteration can't be again used for 2nd iteration because it few elements of it need to be used by some other core and few by other.

[sebraa]

Still, the barriers are a useful synchronization primitive.
- Step 1: diagonal cores do math
- Step 2: swapping phase
- Step 3: non-diagonal cores do math
- Step 4: return result to host, let host send new data
- Repeat.

In any case, you do not want to restart the Epiphany system for every invocation of your program, but instead have the cores handle multiple iterations. For good performance, you want to avoid any host-communication anyway.

[gordon]

[sebraa]

Well, as I've told you, access to the host is very slow, especially if your access pattern is not optimal.

Use consecutive addresses (burst transfers): Otherwise you lose 75% throughput.
Use 64-bit accesses: Otherwise you lose 50% throughput.
Use writes: Reads incur high latency; read requests travel at 12.5% speed only; reads do not allow bursts.

The theoretical maximum throughput on the Parallella is 600 MB/s (with the Epiphany running at 600 MHz), but I know if anyone has ever achieved this. 150 MB/s are a more realistic estimate. As far as my experience goes, the internal communication speed was never an issue.