/// <summary>
/// Creates the scheduler state
/// </summary>
public SchedulerState(Scheduler scheduler)
{
// If the filter is not applied, rootFilter and filterPassingNodes are null.
m_rootFilter = scheduler.RootFilter;
m_filterPassingNodes = scheduler.FilterPassingNodes;
m_incrementalSchedulingState = scheduler.IncrementalSchedulingState;
}
private static void AddIncidentNodes(
IReadonlyDirectedGraph graph,
RangedNodeSet walkFromNodes,
Func <IReadonlyDirectedGraph, NodeId, IEnumerable <Edge> > getEdges,
Func <NodeId, NodeId, bool> validateEdgeTopoProperty,
NodeRange incidentNodeFilter,
bool skipOutOfOrderNodes,
RangedNodeSet addTo)
{
Contract.Requires(!incidentNodeFilter.IsEmpty);
foreach (NodeId existingNode in walkFromNodes)
{
IEnumerable <Edge> edges = getEdges(graph, existingNode);
foreach (Edge edge in edges)
{
NodeId other = edge.OtherNode;
if (skipOutOfOrderNodes && !validateEdgeTopoProperty(existingNode, other))
{
continue;
}
if (incidentNodeFilter.Contains(other))
{
addTo.Add(edge.OtherNode);
}
}
}
}
public void AddInterior()
{
var rangedSet = new RangedNodeSet();
rangedSet.ClearAndSetRange(new NodeRange(new NodeId(123), new NodeId(223)));
rangedSet.Add(new NodeId(200));
rangedSet.Add(new NodeId(201));
XAssert.IsFalse(rangedSet.Range.IsEmpty);
XAssert.AreEqual(new NodeRange(new NodeId(123), new NodeId(223)), rangedSet.Range);
XAssert.IsFalse(rangedSet.Contains(NodeId.Min));
XAssert.IsFalse(rangedSet.Contains(new NodeId(122)));
XAssert.IsFalse(rangedSet.Contains(new NodeId(123)));
XAssert.IsFalse(rangedSet.Contains(new NodeId(124)));
XAssert.IsTrue(rangedSet.Contains(new NodeId(200)));
XAssert.IsTrue(rangedSet.Contains(new NodeId(201)));
XAssert.IsFalse(rangedSet.Contains(new NodeId(223)));
XAssert.IsFalse(rangedSet.Contains(new NodeId(224)));
XAssert.IsFalse(rangedSet.Contains(NodeId.Max));
ExpectContentsFromEnumeration(rangedSet, new NodeId(200), new NodeId(201));
}
/// <summary>
/// Deserializes an instance of <see cref="DirtyNodeTrackerSerializedState"/> using an instance of <see cref="BuildXLReader"/>.
/// </summary>
/// <param name="reader"></param>
public static DirtyNodeTrackerSerializedState Deserialize(BuildXLReader reader)
{
Contract.Requires(reader != null);
RangedNodeSet dirtyNodes = RangedNodeSet.Deserialize(reader);
RangedNodeSet materializedNodes = RangedNodeSet.Deserialize(reader);
RangedNodeSet perpetualDirtyNodes = RangedNodeSet.Deserialize(reader);
return(new DirtyNodeTrackerSerializedState(dirtyNodes, materializedNodes, perpetualDirtyNodes));
}
private DirtyNodeTrackerSerializedState(RangedNodeSet dirtyNodes, RangedNodeSet materializedNodes, RangedNodeSet perpetualDirtyNodes)
{
Contract.Requires(dirtyNodes != null);
Contract.Requires(materializedNodes != null);
Contract.Requires(perpetualDirtyNodes != null);
DirtyNodes = dirtyNodes;
MaterializedNodes = materializedNodes;
PerpetualDirtyNodes = perpetualDirtyNodes;
}
public void Empty()
{
var rangedSet = new RangedNodeSet();
XAssert.IsTrue(rangedSet.Range.IsEmpty);
XAssert.IsFalse(rangedSet.Contains(NodeId.Min));
XAssert.IsFalse(rangedSet.Contains(new NodeId(123)));
XAssert.IsFalse(rangedSet.Contains(NodeId.Max));
ExpectContentsFromEnumeration(rangedSet);
}
/// <summary>
/// Creates an instence of <see cref="DirtyNodeTracker"/>.
/// </summary>
/// <param name="graph">the node dataflow graph.</param>
/// <param name="dirtyNodes">the set of dirty nodes.</param>
/// <param name="perpetualDirtyNodes">the set of nodes that stay dirty even after execution or running from cache.</param>
/// <param name="dirtyNodesChanged">flag indicating whether dirty nodes have changed</param>
/// <param name="materializedNodes">the set of nodes that have materialized their outputs</param>
public DirtyNodeTracker(IReadonlyDirectedGraph graph, RangedNodeSet dirtyNodes, RangedNodeSet perpetualDirtyNodes, bool dirtyNodesChanged, RangedNodeSet materializedNodes)
{
Contract.Requires(graph != null);
Contract.Requires(dirtyNodes != null);
Contract.Requires(materializedNodes != null);
Contract.Requires(perpetualDirtyNodes != null);
Graph = graph;
DirtyNodes = dirtyNodes;
PerpetualDirtyNodes = perpetualDirtyNodes;
DirtyNodesChanged = dirtyNodesChanged;
MaterializedNodes = materializedNodes;
}
private void ExpectContentsFromEnumeration(RangedNodeSet set, params NodeId[] expected)
{
var expectedSet = new HashSet <NodeId>(expected);
foreach (NodeId expectedNode in expectedSet)
{
XAssert.IsTrue(set.Contains(expectedNode), "Contains() didn't find an expected node");
}
foreach (NodeId actualNode in set)
{
bool found = expectedSet.Remove(actualNode);
XAssert.IsTrue(found, "Enumeration returned an unexpected node");
}
XAssert.AreEqual(0, expectedSet.Count, "Enumeration didn't return one or more expected nodes (expectation set nonempty)");
}
public void NonemptyRangeButEmptySet()
{
var rangedSet = new RangedNodeSet();
rangedSet.ClearAndSetRange(new NodeRange(new NodeId(123), new NodeId(223)));
XAssert.IsFalse(rangedSet.Range.IsEmpty);
XAssert.AreEqual(new NodeRange(new NodeId(123), new NodeId(223)), rangedSet.Range);
XAssert.IsFalse(rangedSet.Contains(NodeId.Min));
XAssert.IsFalse(rangedSet.Contains(new NodeId(122)));
XAssert.IsFalse(rangedSet.Contains(new NodeId(123)));
XAssert.IsFalse(rangedSet.Contains(new NodeId(124)));
XAssert.IsFalse(rangedSet.Contains(new NodeId(223)));
XAssert.IsFalse(rangedSet.Contains(NodeId.Max));
ExpectContentsFromEnumeration(rangedSet);
}
private static bool IsReachableFromInternal(
IReadonlyDirectedGraph graph,
NodeId from,
NodeId to,
RangedNodeSet pooledSetA,
RangedNodeSet pooledSetB,
RangedNodeSet pooledSetC,
bool skipOutOfOrderNodes)
{
// This implementation attempts to efficiently traverse a graph without any precomputed indices or labeling beyond topologically ordered node values.
// Index-based approaches are tricky for the expected usage (the BuildXL scheduler) in which the underlying graph is dynamic.
// Instead, we traverse the graph with no prior information in hand, with a careful traversal order and some pruning.
// The thinking on pruning / use of topological labels is not new; for some more robust examples see e.g. FELINE:
// Veloso, Renê Rodrigues, et al. "Reachability Queries in Very Large Graphs: A Fast Refined Online Search Approach." EDBT. 2014.
// Figure 6 in particular gives some geometric insight, though the pruning here is less effective.
// Now, let's build some intuition about this implementation. We begin from a naive approach and will refine to what's actually implemented.
// First, consider the problem of traversing an _undirected_ graph to determine reachability from some point M to N.
// o\ /o o\ /o
// M -- o -- o -- N
// o/ \o o/ \o
// We can imagine the graph in some two-dimensional layout. To perform well with M and N fairly close, it would be wise to proceed in a breadth-first fashion:
// ●\ /● o\ /o
// M -- ● -- o -- N
// ●/ \● o/ \o
// On the first iteration, we have all nodes reachable in one hop from M. On the i'th iteration, we have all nodes reachable in 'i' hops. In the example above,
// N would be found on iteration 3. Geometrically, think of the reached set as a circle expanding outward from M. Note that on each iteration i, we only need to
// hold *the nodes i hopes away* (not i - 1 hops etc.) so in fact we are tracking the outer circumference of this circle. This works since there exists some single
// integer i by which N is at least i hops away (if reachable).
// We can leverage the geometric intuition of a circle to improve that approach a bit. Assume that on some iteration i, we have visited all nodes interior to the circle,
// and the discrete nodes are so numerous and dense as to approximate a circle's area. We can then think of i as a radius and the area as a count of nodes visited -
// on the order of i^2. If we instead traverse the same radius via two circles - each of radius (i/2) then the visited area is (i/2)^2 * 2 = i^2 / 2. Intersection of
// the circles implies a path between the two nodes.
// ●\ /● ●\ /●
// M -- ● -- ● -- N
// ●/ \● ●/ \●
// (above: the prior example on iteration 1 when expanding outward from both endpoints; a path will be found on iteration 2 instead of 3).
// We can still track only the outer circumference of the circles, so long as we alternately expand each circle (rather than expanding both instantaneously; each expansion
// increases the effective search radius by one and so there's no way for the circles to skip past each other).
// Now we first leverage the assumption of a *directed* graph. Simply, we can follow edges in opposing directions from each node (now labelling specifically as 'to' {T} and from'{F}),
// which geometrically looks like expanding semi-circles rather than circles (the nodes labeled X were skipped based on direction):
// X\ />● ●>\ />X
// >F --> ● --> ● --> T
// X/ \>● ●>/ \>X
// Finally we consider the usefulness of a topological labeling of the nodes. This is in fact a generalization of following edges in only one direction:
// - Outgoing-incident nodes must have higher labels than the one current.
// - Incoming-incident nodes must by symmetry have lower labels than the one currnet.
// This means that traversing outgoing edges result in a node-set (for the semi-circle's edge) that increases monotonically over iterations. Symmetrically
// the set for incoming edge traversal decreases monotonically. The example below adds bracketed topological labels; note that for outgoing edges we initially
// have [4, 4] and then [5, 7], and for incoming we first have [11, 11] and then [8, 10]. With this in mind we can see it is futile to traverse from node 10 to node 1,
// when expanding the incoming range since that node is on the 'wrong side' of the outgoing range already.
// X[3]\ />●[7] ●[9]>\ />X[12]
// >F[4] --> ●[6] --> ●[8] --> T[11]
// /->X[2]/ \>●[5] ->●[10]>/ \>X[13]
// X[1] -----------------------/
RangedNodeSet incomingRangeSet = pooledSetA;
incomingRangeSet.SetSingular(to);
var outgoingRangeSet = pooledSetB;
outgoingRangeSet.SetSingular(from);
var swap = pooledSetC;
swap.Clear();
bool toggle = false;
// Loop condition is effectively !incomingRangeSet.IsEmpty && !outgoingRangeSet.IsEmpty, but checked as one of the ranges changes.
while (true)
{
Func <IReadonlyDirectedGraph, NodeId, IEnumerable <Edge> > getEdges;
Func <NodeId, NodeId, bool> validateEdgeTopoProperty;
RangedNodeSet walkFromNodes;
RangedNodeSet intersectWith;
NodeRange incidentNodeFilter;
if (toggle)
{
// Decreasing from 'to' to NodeId.Min
getEdges = (g, n) => g.GetIncomingEdges(n);
validateEdgeTopoProperty = (node, other) => node.Value > other.Value;
incidentNodeFilter = NodeRange.CreateLowerBound(outgoingRangeSet.Range.FromInclusive);
walkFromNodes = incomingRangeSet;
intersectWith = outgoingRangeSet;
}
else
{
// Increasing from 'from' to NodeId.Max
getEdges = (g, n) => g.GetOutgoingEdges(n);
validateEdgeTopoProperty = (node, other) => node.Value < other.Value;
incidentNodeFilter = NodeRange.CreateUpperBound(incomingRangeSet.Range.ToInclusive);
walkFromNodes = outgoingRangeSet;
intersectWith = incomingRangeSet;
}
NodeId intersection;
NodeRange range;
if (RangeIncidentNodesAndIntersect(
graph,
walkFromNodes,
getEdges,
validateEdgeTopoProperty,
incidentNodeFilter,
intersectWith,
skipOutOfOrderNodes,
range: out range,
intersection: out intersection))
{
return(true);
}
if (range.IsEmpty)
{
break;
}
swap.ClearAndSetRange(range);
AddIncidentNodes(graph, walkFromNodes, getEdges, validateEdgeTopoProperty, incidentNodeFilter, skipOutOfOrderNodes, swap);
if (toggle)
{
RangedNodeSet temp = incomingRangeSet;
incomingRangeSet = swap;
swap = temp;
}
else
{
RangedNodeSet temp = outgoingRangeSet;
outgoingRangeSet = swap;
swap = temp;
}
toggle = !toggle;
}
return(false);
}
private static bool RangeIncidentNodesAndIntersect(
IReadonlyDirectedGraph graph,
RangedNodeSet walkFromNodes,
Func <IReadonlyDirectedGraph, NodeId, IEnumerable <Edge> > getEdges,
Func <NodeId, NodeId, bool> validateEdgeTopoProperty,
NodeRange incidentNodeFilter,
RangedNodeSet intersectWith,
bool skipOutOfOrderNodes,
out NodeRange range,
out NodeId intersection)
{
// Note that initially, currentMin > currentMax so NodeRange.CreatePossiblyEmpty
// would return an empty range. We return an empty range iff no nodes pass incidentNodeFilter below.
uint currentMin = NodeId.MaxValue;
uint currentMax = NodeId.MinValue;
foreach (NodeId existingNode in walkFromNodes)
{
IEnumerable <Edge> edges = getEdges(graph, existingNode);
foreach (Edge edge in edges)
{
NodeId other = edge.OtherNode;
if (!validateEdgeTopoProperty(existingNode, other))
{
if (skipOutOfOrderNodes)
{
continue;
}
throw new BuildXLException(I($"Topological order violated due to an edge between nodes {existingNode} and {other}"));
}
if (!incidentNodeFilter.Contains(other))
{
continue;
}
if (other.Value > currentMax)
{
currentMax = edge.OtherNode.Value;
Contract.AssertDebug(currentMax <= NodeId.MaxValue && currentMax >= NodeId.MinValue);
}
if (other.Value < currentMin)
{
currentMin = edge.OtherNode.Value;
Contract.AssertDebug(currentMin <= NodeId.MaxValue && currentMin >= NodeId.MinValue);
}
if (intersectWith.Contains(other))
{
intersection = other;
Contract.AssertDebug(currentMin <= NodeId.MaxValue && currentMin >= NodeId.MinValue);
Contract.AssertDebug(currentMax <= NodeId.MaxValue && currentMax >= NodeId.MinValue);
range = NodeRange.CreatePossiblyEmpty(new NodeId(currentMin), new NodeId(currentMax));
return(true);
}
}
}
intersection = NodeId.Invalid;
Contract.AssertDebug(currentMin <= NodeId.MaxValue && currentMin >= NodeId.MinValue);
Contract.AssertDebug(currentMax <= NodeId.MaxValue && currentMax >= NodeId.MinValue);
range = NodeRange.CreatePossiblyEmpty(new NodeId(currentMin), new NodeId(currentMax));
return(false);
}
