private DryadQueryNode CreateOffset(bool isLong, Expression queryExpr, DryadQueryNode child)
{
// Count node
DryadQueryNode countNode = new DryadBasicAggregateNode(null, AggregateOpType.LongCount,
true, false, queryExpr, child);
// Apply node for x => Offsets(x)
Type paramType = typeof(IEnumerable<>).MakeGenericType(typeof(long));
ParameterExpression param = Expression.Parameter(paramType, "x");
MethodInfo minfo = typeof(HpcLinqEnumerable).GetMethod("Offsets");
Expression body = Expression.Call(minfo, param, Expression.Constant(isLong, typeof(bool)));
Type type = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
LambdaExpression procFunc = Expression.Lambda(type, body, param);
DryadQueryNode mergeCountNode = new DryadMergeNode(true, true, queryExpr, countNode);
DryadQueryNode offsetsNode = new DryadApplyNode(procFunc, queryExpr, mergeCountNode);
// HashPartition
LambdaExpression keySelectExpr = IdentityFunction.Instance(typeof(IndexedValue<long>));
int pcount = child.OutputPartition.Count;
DryadQueryNode hdistNode = new DryadHashPartitionNode(keySelectExpr, null, null, pcount,
false, queryExpr, offsetsNode);
DryadQueryNode resNode = new DryadMergeNode(false, true, queryExpr, hdistNode);
return resNode;
}
private DryadQueryNode VisitWhere(QueryNodeInfo source,
LambdaExpression predicate,
MethodCallExpression queryExpr)
{
DryadQueryNode child = this.Visit(source);
DryadQueryNode whereNode;
if (predicate.Type.GetGenericArguments().Length == 2 ||
(!child.IsDynamic && child.OutputPartition.Count == 1))
{
whereNode = this.PromoteConcat(source, child, x => new DryadWhereNode(predicate, queryExpr, x));
}
else
{
// The "indexed" version
// Create (x, y) => DryadWhere(x, y, predicate)
Type ptype1 = typeof(IEnumerable<>).MakeGenericType(child.OutputTypes[0]);
Type ptype2 = typeof(IEnumerable<>).MakeGenericType(typeof(IndexedValue<long>));
ParameterExpression param1 = Expression.Parameter(ptype1, HpcLinqCodeGen.MakeUniqueName("x"));
ParameterExpression param2 = Expression.Parameter(ptype2, HpcLinqCodeGen.MakeUniqueName("y"));
string targetMethod = queryExpr.Method.Name + "WithStartIndex";
MethodInfo minfo = typeof(HpcLinqEnumerable).GetMethod(targetMethod);
minfo = minfo.MakeGenericMethod(child.OutputTypes[0]);
Expression body = Expression.Call(minfo, param1, param2, predicate);
Type type = typeof(Func<,,>).MakeGenericType(ptype1, ptype2, body.Type);
LambdaExpression procFunc = Expression.Lambda(type, body, param1, param2);
child.IsForked = true;
bool isLong = (queryExpr.Method.Name == "LongWhere");
DryadQueryNode offsetNode = this.CreateOffset(isLong, queryExpr, child);
whereNode = new DryadApplyNode(procFunc, queryExpr, child, offsetNode);
}
return whereNode;
}
private DryadQueryNode VisitZip(QueryNodeInfo first,
QueryNodeInfo second,
LambdaExpression resultSelector,
MethodCallExpression queryExpr)
{
DryadQueryNode child1 = this.Visit(first);
DryadQueryNode child2 = this.Visit(second);
if (child1.IsDynamic || child2.IsDynamic)
{
// Well, let us for now do it on a single machine
child1 = new DryadMergeNode(true, false, queryExpr, child1);
child2 = new DryadMergeNode(true, false, queryExpr, child2);
// Apply node for (x, y) => Zip(x, y, resultSelector)
Type paramType1 = typeof(IEnumerable<>).MakeGenericType(child1.OutputTypes[0]);
ParameterExpression param1 = Expression.Parameter(paramType1, "s1");
Type paramType2 = typeof(IEnumerable<>).MakeGenericType(child2.OutputTypes[0]);
ParameterExpression param2 = Expression.Parameter(paramType2, "s2");
MethodInfo minfo = typeof(HpcLinqHelper).GetMethod("Zip");
minfo = minfo.MakeGenericMethod(child1.OutputTypes[0]);
Expression body = Expression.Call(minfo, param1, param2, resultSelector);
Type funcType = typeof(Func<,>).MakeGenericType(param1.Type, param2.Type, body.Type);
LambdaExpression procFunc = Expression.Lambda(funcType, body, param1, param2);
return new DryadApplyNode(procFunc, queryExpr, child1, child2);
}
else
{
int parCount1 = child1.OutputPartition.Count;
int parCount2 = child2.OutputPartition.Count;
// Count nodes
DryadQueryNode countNode1 = new DryadBasicAggregateNode(null, AggregateOpType.LongCount,
true, false, queryExpr, child1);
DryadQueryNode countNode2 = new DryadBasicAggregateNode(null, AggregateOpType.LongCount,
true, false, queryExpr, child2);
countNode1 = new DryadMergeNode(true, false, queryExpr, countNode1);
countNode2 = new DryadMergeNode(true, false, queryExpr, countNode2);
// Apply node for (x, y) => ZipCount(x, y)
Type paramType1 = typeof(IEnumerable<>).MakeGenericType(typeof(long));
ParameterExpression param1 = Expression.Parameter(paramType1, "x");
ParameterExpression param2 = Expression.Parameter(paramType1, "y");
MethodInfo minfo = typeof(HpcLinqHelper).GetMethod("ZipCount");
Expression body = Expression.Call(minfo, param1, param2);
Type funcType = typeof(Func<,,>).MakeGenericType(param1.Type, param2.Type, body.Type);
LambdaExpression zipCount = Expression.Lambda(funcType, body, param1, param2);
DryadQueryNode indexedCountNode = new DryadApplyNode(zipCount, queryExpr, countNode1, countNode2);
// HashPartition(x => x.index, parCount2)
ParameterExpression param = Expression.Parameter(body.Type.GetGenericArguments()[0], "x");
Expression keySelectBody = Expression.Property(param, "Index");
funcType = typeof(Func<,>).MakeGenericType(param.Type, keySelectBody.Type);
LambdaExpression keySelectExpr = Expression.Lambda(funcType, keySelectBody, param);
DryadQueryNode distCountNode = new DryadHashPartitionNode(keySelectExpr,
null,
parCount2,
queryExpr,
indexedCountNode);
// Apply node for (x, y) => AssignPartitionIndex(x, y)
param1 = Expression.Parameter(body.Type, "x");
Type paramType2 = typeof(IEnumerable<>).MakeGenericType(child2.OutputTypes[0]);
param2 = Expression.Parameter(paramType2, "y");
minfo = typeof(HpcLinqHelper).GetMethod("AssignPartitionIndex");
minfo = minfo.MakeGenericMethod(child2.OutputTypes[0]);
body = Expression.Call(minfo, param1, param2);
funcType = typeof(Func<,,>).MakeGenericType(param1.Type, param2.Type, body.Type);
LambdaExpression assignIndex = Expression.Lambda(funcType, body, param1, param2);
DryadQueryNode addIndexNode = new DryadApplyNode(assignIndex, queryExpr, distCountNode, child2);
// HashPartition(x => x.index, x => x.value, parCount1)
param = Expression.Parameter(body.Type.GetGenericArguments()[0], "x");
body = Expression.Property(param, "Index");
funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
keySelectExpr = Expression.Lambda(funcType, body, param);
body = Expression.Property(param, "Value");
funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
LambdaExpression resultSelectExpr = Expression.Lambda(funcType, body, param);
DryadQueryNode newChild2 = new DryadHashPartitionNode(keySelectExpr,
resultSelectExpr,
null,
parCount1,
false,
queryExpr,
addIndexNode);
newChild2 = new DryadMergeNode(true, true, queryExpr, newChild2);
// Finally the zip node
return new DryadZipNode(resultSelector, queryExpr, child1, newChild2);
}
}
private DryadQueryNode VisitSelect(QueryNodeInfo source,
QueryNodeType nodeType,
LambdaExpression selector,
LambdaExpression resultSelector,
MethodCallExpression queryExpr)
{
DryadQueryNode selectNode;
if (selector.Type.GetGenericArguments().Length == 2)
{
// If this select's child is a groupby node, push this select into its child, if
// 1. The groupby node is not tee'd, and
// 2. The groupby node has no result selector, and
// 3. The selector is decomposable
if (!source.IsForked &&
IsGroupByWithoutResultSelector(source.queryExpression) &&
Decomposition.GetDecompositionInfoList(selector, m_codeGen) != null)
{
MethodCallExpression expr = (MethodCallExpression)source.queryExpression;
LambdaExpression keySelectExpr = HpcLinqExpression.GetLambda(expr.Arguments[1]);
// Figure out elemSelectExpr and comparerExpr
LambdaExpression elemSelectExpr = null;
Expression comparerExpr = null;
if (expr.Arguments.Count == 3)
{
elemSelectExpr = HpcLinqExpression.GetLambda(expr.Arguments[2]);
if (elemSelectExpr == null)
{
comparerExpr = expr.Arguments[2];
}
}
else if (expr.Arguments.Count == 4)
{
elemSelectExpr = HpcLinqExpression.GetLambda(expr.Arguments[2]);
comparerExpr = expr.Arguments[3];
}
// Construct new query expression by building result selector expression
// and pushing it to groupby node.
selectNode = VisitGroupBy(source.children[0].child, keySelectExpr,
elemSelectExpr, selector, comparerExpr, queryExpr);
if (nodeType == QueryNodeType.SelectMany)
{
Type selectorRetType = selector.Type.GetGenericArguments()[1];
LambdaExpression id = IdentityFunction.Instance(selectorRetType);
selectNode = new DryadSelectNode(nodeType, id, resultSelector, queryExpr, selectNode);
}
}
else
{
DryadQueryNode child = this.Visit(source);
selectNode = this.PromoteConcat(
source, child,
x => new DryadSelectNode(nodeType, selector, resultSelector, queryExpr, x));
}
}
else
{
// The "indexed" version
DryadQueryNode child = this.Visit(source);
if (!child.IsDynamic && child.OutputPartition.Count == 1)
{
selectNode = this.PromoteConcat(
source, child,
x => new DryadSelectNode(nodeType, selector, resultSelector, queryExpr, x));
}
else
{
child.IsForked = true;
// Create (x, y) => Select(x, y, selector)
Type ptype1 = typeof(IEnumerable<>).MakeGenericType(child.OutputTypes[0]);
Type ptype2 = typeof(IEnumerable<>).MakeGenericType(typeof(IndexedValue<long>));
ParameterExpression param1 = Expression.Parameter(ptype1, HpcLinqCodeGen.MakeUniqueName("x"));
ParameterExpression param2 = Expression.Parameter(ptype2, HpcLinqCodeGen.MakeUniqueName("y"));
string methodName = queryExpr.Method.Name;
Type[] selectorTypeArgs = selector.Type.GetGenericArguments();
Type typeArg2 = selectorTypeArgs[selectorTypeArgs.Length - 1];
if (nodeType == QueryNodeType.SelectMany)
{
if (resultSelector != null)
{
methodName += "Result";
}
typeArg2 = typeArg2.GetGenericArguments()[0];
}
string targetMethodName = methodName + "WithStartIndex";
MethodInfo minfo = typeof(HpcLinqEnumerable).GetMethod(targetMethodName);
Expression body;
if (resultSelector == null)
{
minfo = minfo.MakeGenericMethod(child.OutputTypes[0], typeArg2);
body = Expression.Call(minfo, param1, param2, selector);
}
else
{
minfo = minfo.MakeGenericMethod(child.OutputTypes[0], typeArg2, resultSelector.Body.Type);
body = Expression.Call(minfo, param1, param2, selector, resultSelector);
}
Type type = typeof(Func<,,>).MakeGenericType(ptype1, ptype2, body.Type);
LambdaExpression procFunc = Expression.Lambda(type, body, param1, param2);
bool isLong = methodName.StartsWith("Long", StringComparison.Ordinal);
DryadQueryNode offsetNode = this.CreateOffset(isLong, queryExpr, child);
selectNode = new DryadApplyNode(procFunc, queryExpr, child, offsetNode);
}
}
return selectNode;
}
private DryadQueryNode VisitSlidingWindow(QueryNodeInfo source,
LambdaExpression procLambda,
Expression windowSizeExpr,
Expression queryExpr)
{
// var windows = source.Apply(s => HpcLinqHelper.Last(s, windowSize));
// var slided = windows.Apply(s => HpcLinqHelper.Slide(s)).HashPartition(x => x.Index);
// slided.Apply(source, (x, y) => HpcLinqHelper.ProcessWindows(x, y, procFunc, windowSize));
DryadQueryNode child = this.Visit(source);
if (child.IsDynamic)
{
throw new DryadLinqException("SlidingWindow is only supported for static partition count");
}
ExpressionSimplifier<int> evaluator = new ExpressionSimplifier<int>();
Expression windowSize = Expression.Constant(evaluator.Eval(windowSizeExpr), typeof(int));
child.IsForked = true;
// Apply node for s => Last(s, windowSize)
Type paramType = typeof(IEnumerable<>).MakeGenericType(child.OutputTypes[0]);
ParameterExpression param = Expression.Parameter(paramType, HpcLinqCodeGen.MakeUniqueName("s"));
MethodInfo minfo = typeof(HpcLinqHelper).GetMethod("Last");
minfo = minfo.MakeGenericMethod(child.OutputTypes[0]);
Expression body = Expression.Call(minfo, param, windowSize);
Type funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
LambdaExpression procFunc = Expression.Lambda(funcType, body, param);
DryadQueryNode lastNode = new DryadApplyNode(procFunc, queryExpr, child);
lastNode = new DryadMergeNode(true, true, queryExpr, lastNode);
// Apply node for s => Slide(s)
param = Expression.Parameter(body.Type, HpcLinqCodeGen.MakeUniqueName("s"));
minfo = typeof(HpcLinqHelper).GetMethod("Slide");
minfo = minfo.MakeGenericMethod(child.OutputTypes[0]);
body = Expression.Call(minfo, param);
funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
procFunc = Expression.Lambda(funcType, body, param);
DryadQueryNode slideNode = new DryadApplyNode(procFunc, queryExpr, lastNode);
// Hash partition to distribute from partition i to i+1
int pcount = child.OutputPartition.Count;
param = Expression.Parameter(body.Type.GetGenericArguments()[0], "x");
Expression keySelectBody = Expression.Property(param, "Index");
funcType = typeof(Func<,>).MakeGenericType(param.Type, keySelectBody.Type);
LambdaExpression keySelectExpr = Expression.Lambda(funcType, keySelectBody, param);
DryadQueryNode hdistNode = new DryadHashPartitionNode(keySelectExpr,
null,
pcount,
queryExpr,
slideNode);
// Apply node for (x, y) => ProcessWindows(x, y, proclambda, windowSize)
Type paramType1 = typeof(IEnumerable<>).MakeGenericType(body.Type);
ParameterExpression param1 = Expression.Parameter(paramType1, HpcLinqCodeGen.MakeUniqueName("x"));
Type paramType2 = typeof(IEnumerable<>).MakeGenericType(child.OutputTypes[0]);
ParameterExpression param2 = Expression.Parameter(paramType2, HpcLinqCodeGen.MakeUniqueName("y"));
minfo = typeof(HpcLinqHelper).GetMethod("ProcessWindows");
minfo = minfo.MakeGenericMethod(child.OutputTypes[0], procLambda.Body.Type);
body = Expression.Call(minfo, param1, param2, procLambda, windowSize);
funcType = typeof(Func<,,>).MakeGenericType(param1.Type, param2.Type, body.Type);
procFunc = Expression.Lambda(funcType, body, param1, param2);
return new DryadApplyNode(procFunc, queryExpr, hdistNode, child);
}
// Phase 1 of the query optimization
internal void GenerateQueryPlanPhase1()
{
if (this.m_queryPlan1 != null) return;
// Apply some simple rewrite rules
SimpleRewriter rewriter = new SimpleRewriter(this.m_exprNodeInfoMap.Values.ToList());
rewriter.Rewrite();
// Generate the query plan of phase1
var referencedNodes = this.m_referencedQueryMap.Values;
this.m_queryPlan1 = new DryadQueryNode[this.m_queryExprs.Length + referencedNodes.Count];
for (int i = 0; i < this.m_queryExprs.Length; i++)
{
this.m_queryPlan1[i] = this.Visit(this.m_queryNodeInfos[i].children[0].child);
}
int idx = this.m_queryExprs.Length;
foreach (QueryNodeInfo nodeInfo in referencedNodes)
{
// Add a Tee'd Merge
this.m_queryPlan1[idx] = this.Visit(nodeInfo.children[0].child);
DryadQueryNode mergeNode = new DryadMergeNode(true, false, nodeInfo.queryExpression,
this.m_queryPlan1[idx]);
this.m_queryPlan1[idx] = new DryadTeeNode(mergeNode.OutputTypes[0], true,
mergeNode.QueryExpression, mergeNode);
nodeInfo.queryNode = this.m_queryPlan1[idx];
idx++;
}
// Finally, add the output nodes.
Dictionary<DryadQueryNode, int> forkCounts = new Dictionary<DryadQueryNode, int>();
for (int i = 0; i < this.m_queryExprs.Length; i++)
{
DryadQueryNode queryNode = this.m_queryPlan1[i];
int cnt;
if (!forkCounts.TryGetValue(queryNode, out cnt))
{
cnt = queryNode.Parents.Count;
}
forkCounts[queryNode] = cnt + 1;
}
for (int i = 0; i < this.m_queryExprs.Length; i++)
{
HpcClientSideLog.Add("Query " + i + " Output: " + this.m_outputDatapaths[i]);
DryadQueryNode queryNode = this.m_queryPlan1[i];
if (TypeSystem.IsAnonymousType(queryNode.OutputTypes[0]))
{
throw new DryadLinqException(HpcLinqErrorCode.OutputTypeCannotBeAnonymous,
SR.OutputTypeCannotBeAnonymous);
}
// Add dummy Apply to make Dryad happy (it doesn't like to hook inputs straight to outputs)
if ((queryNode is DryadInputNode) || (forkCounts[queryNode] > 1))
{
// Add a dummy Apply
Type paramType = typeof(IEnumerable<>).MakeGenericType(queryNode.OutputTypes[0]);
ParameterExpression param = Expression.Parameter(paramType, "x");
Type type = typeof(Func<,>).MakeGenericType(paramType, paramType);
LambdaExpression applyExpr = Expression.Lambda(type, param, param);
DryadQueryNode applyNode = new DryadApplyNode(applyExpr, this.m_queryExprs[i], queryNode);
applyNode.OutputDataSetInfo = queryNode.OutputDataSetInfo;
queryNode = applyNode;
}
if (queryNode is DryadConcatNode)
{
// Again, we add dummy Apply in certain cases to make Dryad happy
((DryadConcatNode)queryNode).FixInputs();
}
// Add the output node
DscCompressionScheme outputScheme = this.m_context.Configuration.OutputDataCompressionScheme;
DryadOutputNode outputNode = new DryadOutputNode(this.m_context,
this.m_outputDatapaths[i],
this.m_isTempOutput[i],
outputScheme,
this.m_queryExprs[i],
queryNode);
this.m_queryPlan1[i] = outputNode;
if (this.m_outputUriMap.ContainsKey(this.m_outputDatapaths[i].ToLower()))
{
throw new DryadLinqException(HpcLinqErrorCode.MultipleOutputsWithSameDscUri,
String.Format(SR.MultipleOutputsWithSameDscUri, this.m_outputDatapaths[i]));
}
this.m_outputUriMap.Add(this.m_outputDatapaths[i].ToLower(), outputNode);
this.m_outputTypes[i] = this.m_queryPlan1[i].OutputTypes[0];
// Remove useless Tees to make Dryad happy
if ((queryNode is DryadTeeNode) && (forkCounts[queryNode] == 1))
{
DryadQueryNode teeChild = queryNode.Children[0];
teeChild.UpdateParent(queryNode, outputNode);
outputNode.UpdateChildren(queryNode, teeChild);
}
}
}
// Basic plan: (reverse all partitions) then (reverse data in each partition)
// The main complication is to perform the first step.
// Approach:
// - tee the input.
// - have a dummy apply node that produces the singleton {0} at each partition
// - merge to get a seq {0,0,..} whose length = nPartition.
// - convert that seq to { (0,n), (1,n), ...}
// - hash-partition to send one item to each of the n workers.
// - use binary-apply to attach targetIndex to each source item
// Apply( seq1 = indexCountPair, seq2 = original data) => ({tgt, item0}, {tgt, item1}, .. )
// - hash-partition to move items to target partition.
// - use local LINQ reverse to do the local data reversal.
private DryadQueryNode VisitReverse(QueryNodeInfo source, Expression queryExpr)
{
DryadQueryNode child = this.Visit(source);
if (child.IsDynamic)
{
throw new DryadLinqException("Reverse is only supported for static partition count");
}
child.IsForked = true;
// Apply node for s => ValueZero(s)
Type paramType = typeof(IEnumerable<>).MakeGenericType(child.OutputTypes[0]);
ParameterExpression param = Expression.Parameter(paramType, "s");
MethodInfo minfo = typeof(HpcLinqHelper).GetMethod("ValueZero");
minfo = minfo.MakeGenericMethod(child.OutputTypes[0]);
Expression body = Expression.Call(minfo, param);
Type funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
LambdaExpression procFunc = Expression.Lambda(funcType, body, param);
DryadQueryNode valueZeroNode = new DryadApplyNode(procFunc, queryExpr, child);
// Apply node for s => ReverseIndex(s)
paramType = typeof(IEnumerable<>).MakeGenericType(typeof(int));
param = Expression.Parameter(paramType, "s");
minfo = typeof(HpcLinqHelper).GetMethod("MakeIndexCountPairs");
body = Expression.Call(minfo, param);
funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
procFunc = Expression.Lambda(funcType, body, param);
DryadQueryNode mergeZeroNode = new DryadMergeNode(true, true, queryExpr, valueZeroNode);
DryadQueryNode indexCountNode = new DryadApplyNode(procFunc, queryExpr, mergeZeroNode);
// HashPartition to distribute the indexCounts -- one to each partition.
// each partition will receive (myPartitionID, pcount).
int pcount = child.OutputPartition.Count;
param = Expression.Parameter(body.Type.GetGenericArguments()[0], "x");
Expression keySelectBody = Expression.Property(param, "Index");
funcType = typeof(Func<,>).MakeGenericType(param.Type, keySelectBody.Type);
LambdaExpression keySelectExpr = Expression.Lambda(funcType, keySelectBody, param);
DryadQueryNode hdistNode = new DryadHashPartitionNode(keySelectExpr,
null,
pcount,
queryExpr,
indexCountNode);
// Apply node for (x, y) => AddIndexForReverse(x, y)
ParameterExpression param1 = Expression.Parameter(body.Type, "x");
Type paramType2 = typeof(IEnumerable<>).MakeGenericType(child.OutputTypes[0]);
ParameterExpression param2 = Expression.Parameter(paramType2, "y");
minfo = typeof(HpcLinqHelper).GetMethod("AddIndexForReverse");
minfo = minfo.MakeGenericMethod(child.OutputTypes[0]);
body = Expression.Call(minfo, param1, param2);
funcType = typeof(Func<,,>).MakeGenericType(param1.Type, param2.Type, body.Type);
LambdaExpression addIndexFunc = Expression.Lambda(funcType, body, param1, param2);
DryadQueryNode addIndexNode = new DryadApplyNode(addIndexFunc, queryExpr, hdistNode, child);
// HashPartition(x => x.index, x => x.value, pcount)
// Moves all data to correct target partition. (each worker will direct all its items to one target partition)
param = Expression.Parameter(body.Type.GetGenericArguments()[0], "x");
body = Expression.Property(param, "Index");
funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
keySelectExpr = Expression.Lambda(funcType, body, param);
body = Expression.Property(param, "Value");
funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
LambdaExpression resultSelectExpr = Expression.Lambda(funcType, body, param);
DryadQueryNode reversePartitionNode = new DryadHashPartitionNode(
keySelectExpr, resultSelectExpr, null,
pcount, false, queryExpr, addIndexNode);
// Reverse node
paramType = typeof(IEnumerable<>).MakeGenericType(reversePartitionNode.OutputTypes[0]);
param = Expression.Parameter(paramType, "x");
minfo = typeof(HpcLinqVertex).GetMethod("Reverse");
minfo = minfo.MakeGenericMethod(child.OutputTypes[0]);
body = Expression.Call(minfo, param);
funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
procFunc = Expression.Lambda(funcType, body, param);
DryadQueryNode resNode = new DryadMergeNode(true, true, queryExpr, reversePartitionNode);
resNode = new DryadApplyNode(procFunc, queryExpr, resNode);
return resNode;
}
private DryadQueryNode VisitMultiApply(QueryNodeInfo source,
LambdaExpression procLambda,
bool perPartition,
bool isFirstOnly,
MethodCallExpression queryExpr)
{
DryadQueryNode[] childs = new DryadQueryNode[source.children.Count];
for (int i = 0; i < source.children.Count; ++i)
{
childs[i] = this.Visit(source.children[i].child);
}
bool isDynamic = childs.Any(x => x.IsDynamic);
if (perPartition && !isDynamic)
{
// Homomorphic case.
if (isFirstOnly)
{
for (int i = 1; i < childs.Length; ++i)
{
childs[i] = new DryadTeeNode(childs[i].OutputTypes[0], true, queryExpr, childs[i]);
childs[i].ConOpType = ConnectionOpType.CrossProduct;
childs[i] = new DryadMergeNode(childs[0].OutputPartition.Count, queryExpr, childs[i]);
}
}
else
{
int count = childs[0].OutputPartition.Count;
for (int i = 1; i < childs.Length; ++i)
{
if (childs[i].OutputPartition.Count != count)
{
throw DryadLinqException.Create(HpcLinqErrorCode.HomomorphicApplyNeedsSamePartitionCount,
SR.HomomorphicApplyNeedsSamePartitionCount,
queryExpr);
}
}
}
}
else
{
// Non-homomorphic case.
for (int i = 0; i < childs.Length; ++i)
{
if (childs[i].IsDynamic || childs[i].OutputPartition.Count > 1)
{
childs[i] = new DryadMergeNode(true, false, queryExpr, childs[i]);
}
}
}
DryadQueryNode applyNode = new DryadApplyNode(procLambda, true, queryExpr, childs);
return applyNode;
}
private DryadQueryNode VisitPartitionOp(string opName,
QueryNodeInfo source,
QueryNodeType nodeType,
Expression controlExpr,
MethodCallExpression queryExpr)
{
DryadQueryNode resNode;
if (nodeType == QueryNodeType.TakeWhile &&
controlExpr.Type.GetGenericArguments().Length != 2)
{
// The "indexed" version.
resNode = this.Visit(source);
// The following block used to be skipped for resNode.OutputPartition.Count == 1,
// which causes compilation error (bug 13593)
// @@TODO[p3] : implement a working optimization for nPartition==1 that calls
// directly to Linq TakeWhile.
// Note: the test is: if (resNode.IsDynamic || resNode.OutputPartition.Count > 1)
{
resNode.IsForked = true;
bool isLong = (queryExpr.Method.Name == "LongTakeWhile");
DryadQueryNode offsetNode = this.CreateOffset(isLong, queryExpr, resNode);
// Create (x, y) => GroupIndexedTakeWhile(x, y, controlExpr)
Type ptype1 = typeof(IEnumerable<>).MakeGenericType(resNode.OutputTypes[0]);
Type ptype2 = typeof(IEnumerable<>).MakeGenericType(typeof(IndexedValue<long>));
ParameterExpression param1 = Expression.Parameter(ptype1, HpcLinqCodeGen.MakeUniqueName("x"));
ParameterExpression param2 = Expression.Parameter(ptype2, HpcLinqCodeGen.MakeUniqueName("y"));
string methodName = "GroupIndexed" + queryExpr.Method.Name;
MethodInfo minfo = typeof(HpcLinqEnumerable).GetMethod(methodName);
minfo = minfo.MakeGenericMethod(resNode.OutputTypes[0]);
Expression body = Expression.Call(minfo, param1, param2, controlExpr);
Type type = typeof(Func<,,>).MakeGenericType(ptype1, ptype2, body.Type);
LambdaExpression procFunc = Expression.Lambda(type, body, param1, param2);
resNode = new DryadApplyNode(procFunc, queryExpr, resNode, offsetNode);
}
}
else if (!source.IsForked &&
(nodeType == QueryNodeType.Take || nodeType == QueryNodeType.TakeWhile) &&
(source.OperatorName == "OrderBy" || source.OperatorName == "OrderByDescending"))
{
resNode = this.Visit(source.children[0].child);
bool isDescending = (source.OperatorName == "OrderByDescending");
MethodCallExpression sourceQueryExpr = (MethodCallExpression)source.queryExpression;
LambdaExpression keySelectExpr = HpcLinqExpression.GetLambda(sourceQueryExpr.Arguments[1]);
Expression comparerExpr = null;
if (sourceQueryExpr.Arguments.Count == 3)
{
comparerExpr = sourceQueryExpr.Arguments[2];
}
resNode = this.PromoteConcat(
source.children[0].child,
resNode,
delegate(DryadQueryNode x) {
DryadQueryNode y = new DryadOrderByNode(keySelectExpr, comparerExpr,
isDescending, sourceQueryExpr, x);
return FirstStagePartitionOp(opName, nodeType, controlExpr, queryExpr, y);
});
if (resNode.IsDynamic || resNode.OutputPartition.Count > 1)
{
// Need a mergesort
resNode = new DryadMergeNode(keySelectExpr, comparerExpr, isDescending, false, sourceQueryExpr, resNode);
}
}
else
{
resNode = this.Visit(source);
if (nodeType == QueryNodeType.Take || nodeType == QueryNodeType.TakeWhile)
{
resNode = this.PromoteConcat(
source, resNode,
x => FirstStagePartitionOp(opName, nodeType, controlExpr, queryExpr, x));
}
}
resNode = new DryadPartitionOpNode(opName, nodeType, controlExpr, false, queryExpr, resNode);
return resNode;
}
private DryadQueryNode VisitConcat(QueryNodeInfo source, MethodCallExpression queryExpr)
{
DryadQueryNode[] childs = new DryadQueryNode[source.children.Count];
for (int i = 0; i < source.children.Count; ++i)
{
childs[i] = this.Visit(source.children[i].child);
}
DryadQueryNode resNode = new DryadConcatNode(queryExpr, childs);
int parCount = resNode.OutputPartition.Count;
if (!resNode.IsDynamic && parCount > StaticConfig.MaxPartitionCount)
{
// Too many partitions, need to repartition
int newParCount = parCount / 2;
DryadQueryNode countNode = new DryadBasicAggregateNode(null, AggregateOpType.LongCount,
true, false, queryExpr, resNode);
DryadQueryNode mergeCountNode = new DryadMergeNode(true, false, queryExpr, countNode);
// Apply node for s => IndexedCount(s)
Type paramType = typeof(IEnumerable<>).MakeGenericType(typeof(long));
ParameterExpression param = Expression.Parameter(paramType, "s");
MethodInfo minfo = typeof(HpcLinqHelper).GetMethod("IndexedCount");
minfo = minfo.MakeGenericMethod(typeof(long));
Expression body = Expression.Call(minfo, param);
Type funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
LambdaExpression indexedCountFunc = Expression.Lambda(funcType, body, param);
DryadQueryNode indexedCountNode = new DryadApplyNode(indexedCountFunc, queryExpr, mergeCountNode);
// HashPartition(x => x.index, parCount)
param = Expression.Parameter(body.Type.GetGenericArguments()[0], "x");
Expression keySelectBody = Expression.Property(param, "Index");
funcType = typeof(Func<,>).MakeGenericType(param.Type, keySelectBody.Type);
LambdaExpression keySelectExpr = Expression.Lambda(funcType, keySelectBody, param);
DryadQueryNode distCountNode = new DryadHashPartitionNode(keySelectExpr,
null,
parCount,
queryExpr,
indexedCountNode);
// Apply node for (x, y) => AddPartitionIndex(x, y, newParCount)
ParameterExpression param1 = Expression.Parameter(body.Type, "x");
Type paramType2 = typeof(IEnumerable<>).MakeGenericType(resNode.OutputTypes[0]);
ParameterExpression param2 = Expression.Parameter(paramType2, "y");
minfo = typeof(HpcLinqHelper).GetMethod("AddPartitionIndex");
minfo = minfo.MakeGenericMethod(resNode.OutputTypes[0]);
body = Expression.Call(minfo, param1, param2, Expression.Constant(newParCount));
funcType = typeof(Func<,,>).MakeGenericType(param1.Type, param2.Type, body.Type);
LambdaExpression addIndexFunc = Expression.Lambda(funcType, body, param1, param2);
DryadQueryNode addIndexNode = new DryadApplyNode(addIndexFunc, queryExpr, distCountNode, resNode);
// HashPartition(x => x.index, x => x.value, newParCount)
param = Expression.Parameter(body.Type.GetGenericArguments()[0], "x");
body = Expression.Property(param, "Index");
funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
keySelectExpr = Expression.Lambda(funcType, body, param);
body = Expression.Property(param, "Value");
funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
LambdaExpression resultSelectExpr = Expression.Lambda(funcType, body, param);
resNode = new DryadHashPartitionNode(keySelectExpr,
resultSelectExpr,
null,
newParCount,
false,
queryExpr,
addIndexNode);
resNode = new DryadMergeNode(true, true, queryExpr, resNode);
}
return resNode;
}
private DryadQueryNode VisitApplyWithPartitionIndex(
QueryNodeInfo source,
LambdaExpression procLambda,
Expression queryExpr)
{
// var indices = source.Apply(s => ValueZero(s)).Apply(s => AssignIndex(s));
// indices.Apply(source, (x, y) => ProcessWithIndex(x, y, procFunc));
DryadQueryNode child = this.Visit(source);
if (child.IsDynamic)
{
throw new DryadLinqException("ApplyWithPartitionIndex is only supported for static partition count");
}
child.IsForked = true;
// Apply node for s => ValueZero(s)
Type paramType = typeof(IEnumerable<>).MakeGenericType(child.OutputTypes[0]);
ParameterExpression param = Expression.Parameter(paramType, "s");
MethodInfo minfo = typeof(HpcLinqHelper).GetMethod("ValueZero");
minfo = minfo.MakeGenericMethod(child.OutputTypes[0]);
Expression body = Expression.Call(minfo, param);
Type funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
LambdaExpression procFunc = Expression.Lambda(funcType, body, param);
DryadQueryNode valueZeroNode = new DryadApplyNode(procFunc, queryExpr, child);
valueZeroNode = new DryadMergeNode(true, true, queryExpr, valueZeroNode);
// Apply node for s => AssignIndex(s)
paramType = typeof(IEnumerable<>).MakeGenericType(typeof(int));
param = Expression.Parameter(paramType, "s");
minfo = typeof(HpcLinqHelper).GetMethod("AssignIndex");
body = Expression.Call(minfo, param);
funcType = typeof(Func<,>).MakeGenericType(param.Type, body.Type);
procFunc = Expression.Lambda(funcType, body, param);
DryadQueryNode assignIndexNode = new DryadApplyNode(procFunc, queryExpr, valueZeroNode);
// HashPartition to distribute the indices -- one to each partition.
int pcount = child.OutputPartition.Count;
param = Expression.Parameter(body.Type, "x");
funcType = typeof(Func<,>).MakeGenericType(param.Type, param.Type);
LambdaExpression keySelectExpr = Expression.Lambda(funcType, param, param);
DryadQueryNode hdistNode = new DryadHashPartitionNode(keySelectExpr,
null,
pcount,
queryExpr,
assignIndexNode);
// Apply node for (x, y) => ProcessWithIndex(x, y, procLambda));
Type paramType1 = typeof(IEnumerable<>).MakeGenericType(child.OutputTypes[0]);
ParameterExpression param1 = Expression.Parameter(paramType1, HpcLinqCodeGen.MakeUniqueName("x"));
Type paramType2 = typeof(IEnumerable<>).MakeGenericType(typeof(int));
ParameterExpression param2 = Expression.Parameter(paramType2, HpcLinqCodeGen.MakeUniqueName("y"));
minfo = typeof(HpcLinqHelper).GetMethod("ProcessWithIndex");
minfo = minfo.MakeGenericMethod(child.OutputTypes[0], procLambda.Body.Type);
body = Expression.Call(minfo, param1, param2, procLambda);
funcType = typeof(Func<,,>).MakeGenericType(param1.Type, param2.Type, body.Type);
procFunc = Expression.Lambda(funcType, body, param1, param2);
return new DryadApplyNode(procFunc, queryExpr, child, hdistNode);
}
private DryadQueryNode VisitApply(QueryNodeInfo source1,
QueryNodeInfo source2,
LambdaExpression procLambda,
bool perPartition,
bool isFirstOnly,
Expression queryExpr)
{
DryadQueryNode child1 = this.Visit(source1);
DryadQueryNode applyNode;
if (source2 == null)
{
// Unary-apply case:
if (perPartition)
{
//homomorphic
applyNode = this.PromoteConcat(source1, child1, x => new DryadApplyNode(procLambda, queryExpr, x));
}
else
{
//non-homomorphic
if (child1.IsDynamic || child1.OutputPartition.Count > 1)
{
child1 = new DryadMergeNode(true, false, queryExpr, child1);
}
applyNode = new DryadApplyNode(procLambda, queryExpr, child1);
}
}
else
{
// Binary-apply case:
DryadQueryNode child2 = this.Visit(source2);
if (perPartition && isFirstOnly)
{
// The function is left homomorphic:
if (!child2.IsForked && (child1.IsDynamic || child1.OutputPartition.Count > 1))
{
// The normal cases..
if (IsMergeNodeNeeded(child2))
{
if (child1.IsDynamic)
{
child2 = new DryadMergeNode(true, false, queryExpr, child2);
child2.IsForked = true;
}
else
{
// Rather than do full merge and broadcast, which has lots of data-movement
// 1. Tee output2 with output cross-product
// 2. Do a merge-stage which will have input1.nPartition nodes each performing a merge.
// This acheives a distribution of the entire input2 to the Apply nodes with least data-movement.
child2 = new DryadTeeNode(child2.OutputTypes[0], true, queryExpr, child2);
child2.ConOpType = ConnectionOpType.CrossProduct;
child2 = new DryadMergeNode(child1.OutputPartition.Count, queryExpr, child2);
}
}
else
{
// the right-data is alread a single partition, so just tee it.
// this will provide a copy to each of the apply nodes.
child2 = new DryadTeeNode(child2.OutputTypes[0], true, queryExpr, child2);
}
}
else
{
// Less common cases..
// a full merge of the right-data may be necessary.
if (child2.IsDynamic || child2.OutputPartition.Count > 1)
{
child2 = new DryadMergeNode(true, false, queryExpr, child2);
if (child1.IsDynamic || child1.OutputPartition.Count > 1)
{
child2.IsForked = true;
}
}
}
applyNode = new DryadApplyNode(procLambda, queryExpr, child1, child2);
}
else if (perPartition && !isFirstOnly && !child1.IsDynamic && !child2.IsDynamic)
{
// Full homomorphic
// No merging occurs.
// NOTE: We generally expect that both the left and right datasets have matching partitionCount.
// however, we don't test for it yet as users might know what they are doing, and it makes
// LocalDebug inconsistent as LocalDebug doesn't throw in that situation.
applyNode = new DryadApplyNode(procLambda, queryExpr, child1, child2);
}
else
{
// Non-homomorphic
// Full merges of both data sets is necessary.
if (child1.IsDynamic || child1.OutputPartition.Count > 1)
{
child1 = new DryadMergeNode(true, false, queryExpr, child1);
}
if (child2.IsDynamic || child2.OutputPartition.Count > 1)
{
child2 = new DryadMergeNode(true, false, queryExpr, child2);
}
applyNode = new DryadApplyNode(procLambda, queryExpr, child1, child2);
}
}
return applyNode;
}
// Creates an "auto-sampling range-partition sub-query"
private DryadQueryNode CreateRangePartition(bool isDynamic,
LambdaExpression keySelectExpr,
LambdaExpression resultSelectExpr,
Expression comparerExpr,
Expression isDescendingExpr,
Expression queryExpr,
Expression partitionCountExpr,
DryadQueryNode child)
{
// Make child a Tee node
child.IsForked = true;
// The partition count
Expression countExpr = null;
if (isDescendingExpr == null)
{
isDescendingExpr = Expression.Constant(false, typeof(bool)); //default for isDescending is false.
}
// NOTE: for MayRTM, isDynamic should never be true
if (!isDynamic)
{
if (partitionCountExpr != null)
{
countExpr = partitionCountExpr;
}
else
{
// If partitionCount was not explicitly set, use the child's partition count.
countExpr = Expression.Constant(child.OutputPartition.Count);
}
}
Type recordType = child.OutputTypes[0];
Type keyType = keySelectExpr.Type.GetGenericArguments()[1];
// Create x => Phase1Sampling(x_1, keySelector, denv)
Type lambdaParamType1 = typeof(IEnumerable<>).MakeGenericType(recordType);
ParameterExpression lambdaParam1 = Expression.Parameter(lambdaParamType1, "x_1");
ParameterExpression denvParam = Expression.Parameter(typeof(HpcLinqVertexEnv), "denv");
MethodInfo minfo = typeof(HpcLinqSampler).GetMethod("Phase1Sampling");
Expression body = Expression.Call(minfo.MakeGenericMethod(recordType, keyType),
lambdaParam1, keySelectExpr, denvParam);
Type type = typeof(Func<,>).MakeGenericType(lambdaParam1.Type, body.Type);
LambdaExpression samplingExpr = Expression.Lambda(type, body, lambdaParam1);
// Create the Sampling node
DryadApplyNode samplingNode = new DryadApplyNode(samplingExpr, queryExpr, child);
// Create x => RangeSampler(x, keySelectExpr, comparer, isDescendingExpr)
Type lambdaParamType = typeof(IEnumerable<>).MakeGenericType(keyType);
ParameterExpression lambdaParam = Expression.Parameter(lambdaParamType, "x_2");
//For RTM, isDynamic should never be true.
//string methodName = (isDynamic) ? "RangeSampler_Dynamic" : "RangeSampler_Static";
Debug.Assert(isDynamic == false, "Internal error: isDynamic is true.");
string methodName = "RangeSampler_Static";
minfo = typeof(HpcLinqSampler).GetMethod(methodName);
minfo = minfo.MakeGenericMethod(keyType);
Expression comparerArgExpr = comparerExpr;
if (comparerExpr == null)
{
if (!TypeSystem.HasDefaultComparer(keyType))
{
throw DryadLinqException.Create(HpcLinqErrorCode.ComparerMustBeSpecifiedOrKeyTypeMustBeIComparable,
string.Format(SR.ComparerMustBeSpecifiedOrKeyTypeMustBeIComparable, keyType),
queryExpr);
}
comparerArgExpr = Expression.Constant(null, typeof(IComparer<>).MakeGenericType(keyType));
}
Expression lastArg;
if (isDynamic)
{
lastArg = denvParam;
}
else
{
lastArg = countExpr;
}
body = Expression.Call(minfo, lambdaParam, comparerArgExpr, isDescendingExpr, lastArg);
type = typeof(Func<,>).MakeGenericType(lambdaParam.Type, body.Type);
LambdaExpression samplerExpr = Expression.Lambda(type, body, lambdaParam);
// Create the sample node
DryadQueryNode sampleDataNode = new DryadMergeNode(false, true, queryExpr, samplingNode);
DryadQueryNode sampleNode = new DryadApplyNode(samplerExpr, queryExpr, sampleDataNode);
sampleNode.IsForked = true;
// Create the range distribute node
DryadQueryNode resNode = new DryadRangePartitionNode(keySelectExpr,
resultSelectExpr,
null,
comparerExpr,
isDescendingExpr,
countExpr,
queryExpr,
child,
sampleNode);
resNode = new DryadMergeNode(false, true, queryExpr, resNode);
// Set the dynamic manager for sampleNode
if (isDynamic)
{
sampleDataNode.DynamicManager = new DynamicRangeDistributor(resNode);
}
return resNode;
}
internal virtual string Visit(DryadApplyNode node,
CodeMemberMethod vertexMethod,
string[] readerNames,
string[] writerNames)
{
return node.AddVertexCode(vertexMethod, readerNames, writerNames);
}