[wip] IncreaseGranularity strategy

Author: franzpoeschel

src/binding/python/openpmd_api/pipe/__main__.py

@@ -128,8 +128,56 @@ def assign(self, assignment, *_):
             assignment.assigned[self.rank].append(element)
         return assignment
 
+class IncreaseGranularity(io.PartialStrategy):
 
-# Example how to implement a simple strategy in Python
+    def __init__(self,
+                 granularity,
+                 inner_distribution=io.ByHostname(io.RoundRobin())):
+        super().__init__()
+        self.inner_distribution = inner_distribution
+        self.granularity = granularity
+
+    def assign(self, assignment, in_ranks, out_ranks):
+        if "in_ranks_inner" in dir(self):
+            return self.inner_distribution.assign(assignment,
+                                                  self.in_ranks_inner,
+                                                  self.out_ranks_inner)
+
+        hosts_in_order = []
+        already_seen = set()
+        for (_, hostname) in in_ranks.items():
+            if hostname not in already_seen:
+                already_seen.add(hostname)
+                hosts_in_order.append(hostname)
+        del already_seen
+        hostname_to_hostgroup = {}  # real host -> host group
+        current_meta_host = 0
+        granularity_counter = 0
+        for host in hosts_in_order:
+            hostname_to_hostgroup[host] = str(current_meta_host)
+            granularity_counter += 1
+            if granularity_counter >= self.granularity:
+                granularity_counter = 0
+                current_meta_host += 1
+        in_ranks_inner = {}
+        for (rank, hostname) in in_ranks.items():
+            in_ranks_inner[rank] = hostname_to_hostgroup[hostname]
+        out_ranks_inner = {}
+        for (rank, hostname) in out_ranks.items():
+            try:
+                out_ranks_inner[rank] = hostname_to_hostgroup[hostname]
+            except KeyError:
+                out_ranks_inner[rank] = hostname
+
+        self.in_ranks_inner = in_ranks_inner
+        self.out_ranks_inner = out_ranks_inner
+
+        return self.inner_distribution.assign(assignment, in_ranks_inner,
+                                              out_ranks_inner)
+
+
+
+#Example how to implement a simple strategy in Python
 class LoadAll(io.Strategy):
 
     def __init__(self, rank):
@@ -168,7 +216,7 @@ def distribution_strategy(dataset_extent,
             strategy_identifier=match.group(2))
         return io.FromPartialStrategy(io.ByHostname(inside_node), second_phase)
     elif strategy_identifier == 'all':
-        return io.FromPartialStrategy(LoadOne(mpi_rank), LoadAll(mpi_rank))
+        return io.FromPartialStrategy(IncreaseGranularity(5), LoadAll(mpi_rank))
     elif strategy_identifier == 'roundrobin':
         return io.RoundRobin()
     elif strategy_identifier == 'binpacking':
@@ -319,6 +367,7 @@ def __copy(self, src, dest, current_path="/data/"):
                 dest.make_constant(src.get_attribute("value"))
             else:
                 chunk_table = src.available_chunks()
+                # todo buffer the strategy
                 strategy = distribution_strategy(shape, self.comm.rank,
                                                  self.comm.size)
                 my_chunks = strategy.assign(chunk_table, self.inranks,