| 1 |
|
| 2 |
Assign elements to processors evenly #elem/#proc -- |
| 3 |
|
| 4 |
std::vector<std::vector<int>> node_to_proc(#node) |
| 5 |
blk = 1 |
| 6 |
st_blk = 0 |
| 7 |
en_blk = st_blk + #blk-elem |
| 8 |
|
| 9 |
for each (proc) { |
| 10 |
node_on_proc(#node) = false |
| 11 |
ste = start element |
| 12 |
ene = end element |
| 13 |
|
| 14 |
end = min(ene, en_blk) |
| 15 |
for each element(ste..ene) { |
| 16 |
foreach node in element { |
| 17 |
if (!node_on_proc(node)) { |
| 18 |
node_to_proc(node).push_back(proc); |
| 19 |
node_on_proc(node) = true; |
| 20 |
} |
| 21 |
} |
| 22 |
} |
| 23 |
if (end < ene) get next block |
| 24 |
else, continue with same block |
| 25 |
} |
| 26 |
|
| 27 |
... All interior nodes will have node_to_proc(node).size() == 1 |
| 28 |
... All unconnected nodes will have node_to_proc(node).size() == 0 |
| 29 |
|
| 30 |
... Add the granularity of element blocks |
| 31 |
Each block has beg, end range, and connectivity size. |
| 32 |
|
| 33 |
... Better to iterate by block: |
| 34 |
|
| 35 |
foreach (element block) |
| 36 |
-- get connectivity |
| 37 |
find st_proc |
| 38 |
find end_proc |
| 39 |
|
| 40 |
proc for an element (0-based): |
| 41 |
|
| 42 |
proc = elem / elem-per-proc + |
| 43 |
|
| 44 |
elem-per-proc = #elem / #proc |
| 45 |
extra = #elem % #proc |
| 46 |
|
| 47 |
elementCount = elem_per_proc + (proc < extra ? 1 : 0); |
| 48 |
if (proc < extra) { |
| 49 |
startElement = (el_per_proc+1) * (proc); |
| 50 |
} else { |
| 51 |
startElement = (el_per_proc) * (proc) + remain; |
| 52 |
} |