private long _weightBalanceThreshold = 0 ; //at what weight should the node be treated as contributor to incoming nodes.
public DistributionMatrix(ArrayList weightIdList, Address address, DistributionData distData, ILogger NCacheLog)
{
_address = address;
_distData = distData;
_filteredWeightIdList = new ArrayList();
_itemsCount = weightIdList.Count;
_totalWeight = 1;
_weightToSacrifice = 0;
_cushionFactor = 10;
_percentWeightToSacrifice = 0;
_weightBalanceThreshold = Convert.ToInt32((_maxCacheSize * WeightBalanceThresholdPercent) / 100); //10%, threshold at which we feel to balance weight for incoming nodes. its value is percent of MaxCacheSize
if (NCacheLog.IsInfoEnabled) NCacheLog.Error("DistributionMatrix.ctor", "Address->" + address.ToString() + ", DistributionData->" + distData.ToString());
//muds:
//this is the temp code just to put some trace...
int bucketCount = 0;
foreach (WeightIdPair wiPair in weightIdList)
{
if (wiPair.Address.compare(address) == 0)
{
if(NCacheLog.IsInfoEnabled) NCacheLog.Info("DistributionMatrix.ctor", "waitPair" + wiPair.Address.ToString() + ", wiPait->" + wiPair.BucketId);
_filteredWeightIdList.Add(wiPair);
bucketCount++;
}
}
if (NCacheLog.IsInfoEnabled) NCacheLog.Info("DistributionMatrix..ctor", address + " owns " + bucketCount + " buckets");
_filteredWeightIdList.Sort();
if (NCacheLog.IsInfoEnabled) NCacheLog.Info("DistributionMatrix.ctor", "_filterWeightIdList.Count:" + _filteredWeightIdList.Count + ", distData.BucketPerNode: " + distData.BucketsPerNode);
//Current bucket count - bucketss count after division gives buckets count to be sacrificed.
_bucketsToSacrifice = _filteredWeightIdList.Count - distData.BucketsPerNode;
if (_bucketsToSacrifice <= 0)
{
NCacheLog.Error("DistributionMatrix", "Address::" + address.ToString() + " cant sacrifice any bucket. Buckets/Node = " + distData.BucketsPerNode + " My Buckets Count = " + _filteredWeightIdList.Count);
return;
}
int rows = Convert.ToInt32(Math.Ceiling((double)((decimal)_filteredWeightIdList.Count /(decimal)_bucketsToSacrifice)));
int cols = _bucketsToSacrifice;
InitializeMatrix(rows, cols);
}
public static ArrayList BalanceBuckets(DistributionInfoData distInfo, ArrayList hashMap, Hashtable bucketStats, ArrayList members,long cacheSizePerNode ,ILogger NCacheLog)
{
DistributionData distData = new DistributionData(hashMap, bucketStats, members, NCacheLog, cacheSizePerNode);
Boolean bShouldBalanceWeight = false;
if (distInfo.DistribMode == DistributionMode.AvgWeightTime) //If weight and time to move has to be avg. Cut the weight to half.
{
if (NCacheLog.IsInfoEnabled) NCacheLog.Info("DistributionImpl.BalanceBuckets()", "Request comes with DistributionMode.AvgWeightTime");
distData.WeightPerNode/= 2;
}
ArrayList distMatrix = distData.DistributionMatrixForNodes;
ArrayList finalBuckets = new ArrayList();
//We need to cater for the cases where we dont need to actually balance the data over nodes, as cluster itself is starting
//and no actual load is present within a cluster and on each node.
foreach (DistributionMatrix dMatrix in distMatrix)
{
if (dMatrix.DoWeightBalance == true)
{
bShouldBalanceWeight = true;
break;
}
}
//If cluster is not loaded only shuffled disribution is required. No need to balance any weight.
if (bShouldBalanceWeight == false)
{
if (NCacheLog.IsInfoEnabled) NCacheLog.Info("DistributionImpl.BalanceBuckets()", "Cluster is not loaded only shuffled disribution is required. No need to balance any weight.");
distInfo.DistribMode = DistributionMode.ShuffleBuckets;
}
//For cases below we also need to calculate Weight to be balanced along with buckets sacrifices.
switch (distInfo.DistribMode)
{
case DistributionMode.OptimalTime:
foreach (DistributionMatrix dMatrix in distMatrix)
{
int [,] IdMatrix = dMatrix.IdMatrix;
for (int i = 0; i < dMatrix.MatrixDimension.Cols; i++)
finalBuckets.Add(IdMatrix[0, i]); //Always first row of the matrix to be given
}
if (NCacheLog.IsInfoEnabled)
{
NCacheLog.Info("DistributionImpl.BalanceBuckets()", "Request is DistributionMode.OptimalTime");
NCacheLog.Info("Selected Buckets are: -");
for (int i = 0; i < finalBuckets.Count; i++)
NCacheLog.Info(finalBuckets[i].ToString());
}
return finalBuckets;
case DistributionMode.ShuffleBuckets: //Although code replication is observed here. Still I prefer to make its copy rather puting fewer if-else to control. I need some time efficiency here.
foreach (DistributionMatrix dMatrix in distMatrix)
{
int[,] IdMatrix = dMatrix.IdMatrix;
int[] resultIndices;
RowsBalanceResult rbResult = DistributionCore.ShuffleSelect(dMatrix);
resultIndices = rbResult.ResultIndicies;
for (int i = 0, j = 0; i < resultIndices.Length; i++)
{
int index = resultIndices[i]; //Index would never be zero, rather the value corresponding in the Matrix be zero.
//Get row and col on the basis of matrix index (index of one-D array).
int row = index / dMatrix.MatrixDimension.Cols;
int col = index % dMatrix.MatrixDimension.Cols;
if (IdMatrix[row, col] == -1) //dealing with exceptional case when last row is selected and it got few non-indices.So replace those with lowest most indices in the matrix.
{
finalBuckets.Add(IdMatrix[0, j]);
j++;
}
else
{
finalBuckets.Add(IdMatrix[row, col]);
}
}
}
if (NCacheLog.IsInfoEnabled )
{
NCacheLog.Info("DistributionImpl.BalanceBuckets()", "Request is DistributionMode.ShuffleBuckets");
NCacheLog.Info("Selected Buckets are: -");
for (int i = 0; i < finalBuckets.Count; i++)
NCacheLog.Info(finalBuckets[i].ToString());
}
return finalBuckets;
case DistributionMode.OptimalWeight: //For both same code works. Change is only in weight that is modified above . it is called FallThrough in switch statements.
case DistributionMode.AvgWeightTime:
foreach (DistributionMatrix dMatrix in distMatrix)
{
int[,] IdMatrix = dMatrix.IdMatrix;
int[] resultIndices;
RowsBalanceResult rbResult = DistributionCore.CompareAndSelect(dMatrix);
resultIndices = rbResult.ResultIndicies;
for (int i = 0,j=0; i < resultIndices.Length; i++)
{
int index = resultIndices[i]; //Index would never be zero, rather the value corresponding in the Matrix be zero.
//Get row and col on the basis of matrix index (index of one-D array).
int row = index / dMatrix.MatrixDimension.Cols;
int col = index % dMatrix.MatrixDimension.Cols;
if (IdMatrix[row, col] == -1) //dealing with exceptional case when last row is selected and it got few non-indices.So replace those with lowest most indices in the matrix.
{
finalBuckets.Add(IdMatrix[0,j]);
j++;
}
else
{
finalBuckets.Add(IdMatrix[row, col]);
}
}
}
if (NCacheLog.IsInfoEnabled )
{
NCacheLog.Info("DistributionImpl.BalanceBuckets()", "Request is DistributionMode.AvgWeightTime/ DistributionMode.OptimalWeight");
NCacheLog.Info("Selected Buckets are: -");
for (int i = 0; i < finalBuckets.Count; i++)
NCacheLog.Info(finalBuckets[i].ToString());
}
return finalBuckets;
default:
break;
} //end switch
return null;
} //end func.
public static ArrayList BalanceBuckets(DistributionInfoData distInfo, ArrayList hashMap, Hashtable bucketStats, ArrayList members, long cacheSizePerNode, ILogger NCacheLog)
{
DistributionData distData = new DistributionData(hashMap, bucketStats, members, NCacheLog, cacheSizePerNode);
Boolean bShouldBalanceWeight = false;
if (distInfo.DistribMode == DistributionMode.AvgWeightTime) //If weight and time to move has to be avg. Cut the weight to half.
{
if (NCacheLog.IsInfoEnabled)
{
NCacheLog.Info("DistributionImpl.BalanceBuckets()", "Request comes with DistributionMode.AvgWeightTime");
}
distData.WeightPerNode /= 2;
}
ArrayList distMatrix = distData.DistributionMatrixForNodes;
ArrayList finalBuckets = new ArrayList();
//We need to cater for the cases where we don't need to actually balance the data over nodes, as cluster itself is starting
//and no actual load is present within a cluster and on each node.
foreach (DistributionMatrix dMatrix in distMatrix)
{
if (dMatrix.DoWeightBalance == true)
{
bShouldBalanceWeight = true;
break;
}
}
//If cluster is not loaded only shuffled distribution is required. No need to balance any weight.
if (bShouldBalanceWeight == false)
{
if (NCacheLog.IsInfoEnabled)
{
NCacheLog.Info("DistributionImpl.BalanceBuckets()", "Cluster is not loaded only shuffled distribution is required. No need to balance any weight.");
}
distInfo.DistribMode = DistributionMode.ShuffleBuckets;
}
//For cases below we also need to calculate Weight to be balanced along with buckets sacrifices.
switch (distInfo.DistribMode)
{
case DistributionMode.OptimalTime:
foreach (DistributionMatrix dMatrix in distMatrix)
{
int [,] IdMatrix = dMatrix.IdMatrix;
for (int i = 0; i < dMatrix.MatrixDimension.Cols; i++)
{
finalBuckets.Add(IdMatrix[0, i]); //Always first row of the matrix to be given
}
}
if (NCacheLog.IsInfoEnabled)
{
NCacheLog.Info("DistributionImpl.BalanceBuckets()", "Request is DistributionMode.OptimalTime");
NCacheLog.Info("Selected Buckets are: -");
for (int i = 0; i < finalBuckets.Count; i++)
{
NCacheLog.Info(finalBuckets[i].ToString());
}
}
return(finalBuckets);
case DistributionMode.ShuffleBuckets: //Although code replication is observed here. Still I prefer to make its copy rather putting fewer if-else to control. I need some time efficiency here.
foreach (DistributionMatrix dMatrix in distMatrix)
{
int[,] IdMatrix = dMatrix.IdMatrix;
int[] resultIndices;
RowsBalanceResult rbResult = DistributionCore.ShuffleSelect(dMatrix);
resultIndices = rbResult.ResultIndicies;
for (int i = 0, j = 0; i < resultIndices.Length; i++)
{
int index = resultIndices[i]; //Index would never be zero, rather the value corresponding in the Matrix be zero.
//Get row and col on the basis of matrix index (index of one-D array).
int row = index / dMatrix.MatrixDimension.Cols;
int col = index % dMatrix.MatrixDimension.Cols;
if (IdMatrix[row, col] == -1) //dealing with exceptional case when last row is selected and it got few non-indices.So replace those with lowest most indices in the matrix.
{
finalBuckets.Add(IdMatrix[0, j]);
j++;
}
else
{
finalBuckets.Add(IdMatrix[row, col]);
}
}
}
if (NCacheLog.IsInfoEnabled)
{
NCacheLog.Info("DistributionImpl.BalanceBuckets()", "Request is DistributionMode.ShuffleBuckets");
NCacheLog.Info("Selected Buckets are: -");
for (int i = 0; i < finalBuckets.Count; i++)
{
NCacheLog.Info(finalBuckets[i].ToString());
}
}
return(finalBuckets);
case DistributionMode.OptimalWeight: //For both same code works. Change is only in weight that is modified above . it is called FallThrough in switch statements.
case DistributionMode.AvgWeightTime:
foreach (DistributionMatrix dMatrix in distMatrix)
{
int[,] IdMatrix = dMatrix.IdMatrix;
int[] resultIndices;
RowsBalanceResult rbResult = DistributionCore.CompareAndSelect(dMatrix);
resultIndices = rbResult.ResultIndicies;
for (int i = 0, j = 0; i < resultIndices.Length; i++)
{
int index = resultIndices[i]; //Index would never be zero, rather the value corresponding in the Matrix be zero.
//Get row and col on the basis of matrix index (index of one-D array).
int row = index / dMatrix.MatrixDimension.Cols;
int col = index % dMatrix.MatrixDimension.Cols;
if (IdMatrix[row, col] == -1) //dealing with exceptional case when last row is selected and it got few non-indices.So replace those with lowest most indices in the matrix.
{
finalBuckets.Add(IdMatrix[0, j]);
j++;
}
else
{
finalBuckets.Add(IdMatrix[row, col]);
}
}
}
if (NCacheLog.IsInfoEnabled)
{
NCacheLog.Info("DistributionImpl.BalanceBuckets()", "Request is DistributionMode.AvgWeightTime/ DistributionMode.OptimalWeight");
NCacheLog.Info("Selected Buckets are: -");
for (int i = 0; i < finalBuckets.Count; i++)
{
NCacheLog.Info(finalBuckets[i].ToString());
}
}
return(finalBuckets);
default:
break;
} //end switch
return(null);
} //end func.
