CreateUpdateDesc(
KMeansOptions options,
KMeansMapPartition[] workers,
IEnumerable <Vector> vectors,
int[] clusterIds,
int nCenters,
int nRank,
int nIteration
)
{
KMeansUpdateCenters update = new KMeansUpdateCenters();
update.m_evtMapWorkComplete = m_evtMapWorkComplete;
update.m_evtMapWorkAvailable = m_evtMapWorkAvailable;
update.m_evtUpdateWorkComplete = m_evtUpdateWorkComplete;
update.m_evtUpdateWorkAvailable = m_evtUpdateWorkAvailable;
update.m_barrier = m_barrier;
update.m_nIterations = nIteration;
update.m_nCenters = nCenters;
update.m_nRank = nRank;
update.m_nPoints = vectors.Count();
update.m_vectors = vectors;
update.m_sharedCenters = null;
update.m_sharedGroupCounts = null;
update.m_counts = null;
update.m_clusterIds = clusterIds;
update.m_fDelta = 0.0f;
update.m_workers = workers;
update.m_options = options;
return(update);
}
CreateMapPartitionDesc(
KMeansOptions options,
Vector[] vectors,
Vector[] initCenters,
int[] clusterIds,
int nCenters,
int nRank,
int nId,
int nStartIndex,
int nPartitionSize
)
{
KMeansMapPartition partition = new KMeansMapPartition();
partition.m_evtMapWorkComplete = m_evtMapWorkComplete;
partition.m_evtMapWorkAvailable = m_evtMapWorkAvailable;
partition.m_evtUpdateWorkComplete = m_evtUpdateWorkComplete;
partition.m_evtUpdateWorkAvailable = m_evtUpdateWorkAvailable;
partition.m_barrier = m_barrier;
partition.m_nId = nId;
partition.m_nStartIndex = nStartIndex;
partition.m_nPartitionSize = nPartitionSize;
partition.m_nCenters = nCenters;
partition.m_nRank = nRank;
partition.m_nPoints = vectors.Count();
partition.m_oldCenters = initCenters;
partition.m_newCenters = null;
partition.m_newGroupCounts = null;
partition.m_clusterIds = clusterIds;
partition.m_fDelta = 0.0f;
partition.m_vectors = vectors;
return(partition);
}
SelectVersionsToCompare(
KMeansOptions options
)
{
List <String> runnableVersions = new List <string>();
List <String> specifiedVersions = options.m_vImplementations.Keys.ToList();
foreach (String sVersion in specifiedVersions)
{
if (IsParallelizableVersion(sVersion))
{
int procs = options.m_nMinimumThreadsSweep;
while (procs <= System.Environment.ProcessorCount &&
procs <= options.m_nMaximumThreadsSweep)
{
runnableVersions.Add(sVersion + "-" + procs);
if (procs < 16)
{
procs *= 2;
}
else
{
procs += 8;
}
}
}
else
{
runnableVersions.Add(sVersion);
}
}
return(runnableVersions);
}
execute(
KMeansOptions options,
IEnumerable <Vector> vectors,
ref Vector[] centers,
out int[] clusterIds,
int nParallelDegree = 1,
bool bAvoidLazyEval = true
)
{
IEnumerable <Vector> result = null;
m_pStopwatch.Start();
m_nStartParallelPhase = m_pStopwatch.ElapsedMilliseconds;
result = LINQKMeans.Steps(vectors,
centers,
options.m_nMaxIterations,
nParallelDegree,
bAvoidLazyEval);
result.Count();
m_pStopwatch.Stop();
m_nEndParallelPhase = m_pStopwatch.ElapsedMilliseconds;
clusterIds = null;
centers = result.ToArray();
return(options.m_nMaxIterations);
}
execute(
KMeansOptions options,
IEnumerable <Vector> vectors,
ref Vector[] centers,
out int[] clusterIds,
int nMaxParallelism = 1,
bool bAvoidLazyEval = true
);
CheckResult(
KMeansOptions options,
IEnumerable <Vector> centers,
IEnumerable <Vector> refnewcenters,
bool bVerbose = true,
bool bDumpLists = false
)
{
List <int> diffIndexes = new List <int>();
if (centers.Count() != refnewcenters.Count())
{
if (bVerbose)
{
Console.WriteLine("different lengths!");
}
return(false);
}
bool bResult = true;
int nIndex = 0;
foreach (Vector center in centers)
{
if (!ValuePresent(options, center, refnewcenters))
{
bResult = false;
diffIndexes.Add(nIndex);
}
nIndex++;
}
if (!bResult && bDumpLists)
{
if (bVerbose)
{
Console.WriteLine("center lists differ!");
}
Console.WriteLine("REF:");
foreach (Vector center in refnewcenters)
{
Console.WriteLine("\t{0}", center);
}
Console.WriteLine("CAND: could not match:");
foreach (int index in diffIndexes)
{
Console.WriteLine("\t{0}", centers.ElementAt(index));
}
}
return(bResult);
}
ValuePresent(
KMeansOptions options,
Vector center,
IEnumerable <Vector> refcenters
)
{
foreach (Vector rc in refcenters)
{
if (Vector.Dist(center, rc) < options.m_fEpsilon)
{
return(true);
}
}
return(false);
}
execute(
KMeansOptions options,
IEnumerable <Vector> vectors,
ref Vector[] centers,
out int[] clusterIds,
int nMaxParallelism,
bool bAvoidLazyEval
)
{
int nCenters = options.m_nClusters;
int nIterations = 0;
clusterIds = null;
centers = ComputeNewCenters(options, vectors, centers, out clusterIds, out nIterations);
return(nIterations);
}
SelectImplementation(
KMeansOptions options,
String sVersion,
out int nMaxParallelism,
out bool bAvoidLazyEval
)
{
if (sVersion.StartsWith("PLINQ") ||
sVersion.StartsWith("threaded") ||
sVersion.StartsWith("tasks"))
{
int nDashIndex = sVersion.IndexOf('-');
String strMaxParallelism = sVersion.Substring(nDashIndex + 1);
nMaxParallelism = Int32.Parse(strMaxParallelism);
bAvoidLazyEval = true;
if (sVersion.StartsWith("PLINQ"))
{
return(new PLINQKMeans(options));
}
if (sVersion.StartsWith("threaded"))
{
return(new ThreadedKMeans(options));
}
if (sVersion.StartsWith("tasks"))
{
return(new TaskingKMeans(options));
}
return(null);
}
else
{
nMaxParallelism = 1;
bAvoidLazyEval = true;
if (sVersion == "seq")
{
return(new SequentialKMeans(options));
}
if (sVersion == "LINQ")
{
return(new LINQKMeans(options));
}
}
return(null);
}
execute(
KMeansOptions options,
IEnumerable <Vector> vectors,
ref Vector[] centers,
out int[] clusterIds,
int nMaxParallelism = 1,
bool bAvoidLazyEval = true
)
{
int nIterations = 0;
clusterIds = null;
centers = Compute(vectors,
centers,
options.m_nMaxIterations,
bAvoidLazyEval).ToArray();
nIterations = options.m_nMaxIterations;
return(nIterations);
}
execute(
KMeansOptions options,
IEnumerable <Vector> vectors,
ref Vector[] centers,
out int[] clusterIds,
int nMaxParallelism = 1,
bool bAvoidLazyEval = true
)
{
int nIterations = 0;
options.m_nCurrentThreadsSweep = nMaxParallelism;
centers = ComputeNewCenters(options,
vectors,
centers,
options.m_nClusters,
out clusterIds,
out nIterations);
return(nIterations);
}
///-------------------------------------------------------------------------------------------------
/// <summary> Main entry-point for this application. </summary>
///
/// <remarks> crossbac, 8/6/2013. </remarks>
///
/// <param name="args"> Array of command-line argument strings. </param>
///-------------------------------------------------------------------------------------------------
static void Main(string[] args)
{
KMeansOptions options = KMeansOptions.getOptions(args);
if (options == null)
{
return;
}
if (options.m_bGenerateData)
{
GenerateRandomInput(options.m_strFileName,
options.m_nGenerateElems,
options.m_nGenerateDims);
return;
}
Vector[] attributes =
options.m_bBinaryInput ?
KMeansCalculator.ReadBinaryInput(options.m_strFileName) :
KMeansCalculator.ReadTextInput(options.m_strFileName);
ComparePerformance(options, attributes);
}
///-------------------------------------------------------------------------------------------------
/// <summary> Constructor. </summary>
///
/// <remarks> Chris Rossbach ([email protected]), 8/2/2012. </remarks>
///
/// <param name="options"> Options for controlling the operation. </param>
///-------------------------------------------------------------------------------------------------
public ThreadedKMeans(KMeansOptions options)
{
m_options = options;
}
ComputeNewCenters(
KMeansOptions options,
IEnumerable <Vector> vectors,
Vector[] centers,
out int[] clusterIds,
out int nIterations
)
{
float delta = 0.0f;
int iterations = 0;
int nRank = vectors.ElementAt(0).Rank;
int nPoints = vectors.Count();
int nCenters = centers.Length;
clusterIds = new int[nPoints];
for (int i = 0; i < nPoints; i++)
{
clusterIds[i] = -1;
}
Vector[] newCenters = new Vector[nCenters];
Vector[] oldCenters = new Vector[nCenters];
int[] groupCounts = new int[nCenters];
for (int i = 0; i < nCenters; i++)
{
groupCounts[i] = 0;
newCenters[i] = Vector.ZeroVector(nRank);
oldCenters[i] = new Vector(centers[i]);
}
m_pStopwatch.Start();
m_nStartParallelPhase = m_pStopwatch.ElapsedMilliseconds;
do
{
delta = 0.0f;
for (int i = 0; i < nPoints; i++)
{
Vector vec = vectors.ElementAt(i);
int nIndex = FindNearestCenter(vec, oldCenters);
if (clusterIds[i] != nIndex)
{
delta += 1.0f;
}
clusterIds[i] = nIndex;
newCenters[nIndex] += vec;
groupCounts[nIndex] += 1;
}
for (int i = 0; i < nCenters; i++)
{
newCenters[i] /= groupCounts[i];
}
delta /= nPoints;
iterations++;
Vector[] tCenters = oldCenters;
oldCenters = newCenters;
newCenters = tCenters;
for (int i = 0; i < nCenters; i++)
{
groupCounts[i] = 0;
newCenters[i].clear();
}
} while((delta > options.m_fConvergenceThreshold) && (iterations < options.m_nMaxIterations));
for (int i = 0; i < nCenters; i++)
{
centers[i] = oldCenters[i];
}
nIterations = iterations;
m_nEndParallelPhase = m_pStopwatch.ElapsedMilliseconds;
return(centers);
}
///-------------------------------------------------------------------------------------------------
/// <summary> Compare performance of several implementations. </summary>
///
/// <remarks> Chris Rossbach ([email protected]), 8/7/2012. </remarks>
///
/// <param name="options"> Options for controlling the operation. </param>
/// <param name="attributes"> The attributes. </param>
///-------------------------------------------------------------------------------------------------
static void ComparePerformance(
KMeansOptions options,
IEnumerable <Vector> attributes
)
{
List <String> runnableVersions = SelectVersionsToCompare(options);
Dictionary <String, Sample[]> perfdata = new Dictionary <string, Sample[]>();
Dictionary <String, Performance> stats = new Dictionary <string, Performance>();
foreach (String version in runnableVersions)
{
Sample[] vdata = new Sample[options.m_nTotalRuns];
for (int i = 0; i < options.m_nTotalRuns; i++)
{
vdata[i] = new Sample();
}
perfdata[version] = vdata;
}
IEnumerable <Vector> goldcenters =
SelectInitialCenters(attributes,
options.m_nClusters,
options.m_nRandomSeed);
for (int i = 0; i < options.m_nTotalRuns; i++)
{
long lReferenceImplTime = 0;
IEnumerable <Vector> refnewcenters = null;
if (options.m_bCheckResult)
{
Vector[] refcenters = DuplicateCenters(goldcenters);
ReferenceKMeans refkmeans = new ReferenceKMeans();
refnewcenters = refkmeans.Compute(attributes, refcenters, options.m_nMaxIterations, true);
lReferenceImplTime = refkmeans.RuntimeMilliseconds;
}
foreach (String sVersion in runnableVersions)
{
bool bAvoidLazyEval = true;
int nMaxParallelism = 1;
int[] rClusterIds = null;
Vector[] newcenters = DuplicateCenters(goldcenters);
KMeansCalculator kmeans = SelectImplementation(options, sVersion, out nMaxParallelism, out bAvoidLazyEval);
int nIterations = kmeans.execute(options, attributes, ref newcenters, out rClusterIds, nMaxParallelism, bAvoidLazyEval);
Sample isample = perfdata[sVersion][i];
isample.m_impltime = kmeans.RuntimeMilliseconds;
isample.m_success = true;
isample.m_reftime = 0;
if (options.m_bCheckResult)
{
isample.m_reftime = lReferenceImplTime;
isample.m_success = KMeansCalculator.CheckResult(options,
newcenters,
refnewcenters,
options.m_bVerbose,
options.m_bVerbose);
if (!isample.m_success)
{
Console.WriteLine("FAILED");
return;
}
}
}
}
Console.WriteLine("SUCCEEDED");
foreach (String v in perfdata.Keys)
{
Sample[] samples = perfdata[v];
Performance perf = new Performance(samples);
Console.WriteLine("{0,-15}: {1}, avg={2}", v, perf.RawRuntimes(), perf.m_impltime.ToString("f1"));
}
}
///-------------------------------------------------------------------------------------------------
/// <summary> Constructor. </summary>
///
/// <remarks> Chris Rossbach ([email protected]), 8/2/2012. </remarks>
///
/// <param name="options"> Options for controlling the operation. </param>
///-------------------------------------------------------------------------------------------------
public TaskingKMeans(KMeansOptions options)
{
m_options = options;
}
public static KMeansOptions getOptions(string[] args)
{
Getopt.Getopt g = new Getopt.Getopt("KMeans", args, "cr:x:i:m:T:e:bR:X:K:I:vV:gd:E:");
KMeansOptions options = new KMeansOptions();
try {
int c;
while ((c = g.getopt()) != -1)
{
switch (c)
{
case 'g':
options.m_bGenerateData = true;
break;
case 'd':
options.m_nGenerateDims = Int32.Parse(g.Optarg);
break;
case 'E':
options.m_nGenerateElems = Int32.Parse(g.Optarg);
break;
case 'v':
options.m_bVerbose = true;
break;
case 'V':
options.m_vImplementations[g.Optarg] = null;
break;
case 'X':
options.m_nMaximumThreadsSweep = Int32.Parse(g.Optarg);
break;
case 'K':
options.m_nPartitionSize = Int32.Parse(g.Optarg);
break;
case 'R':
options.m_nRandomSeed = Int32.Parse(g.Optarg);
break;
case 'c':
options.m_bCheckResult = true;
break;
case 'C':
options.m_bComparePerformance = true;
break;
case 'r':
options.m_nTotalRuns = Int16.Parse(g.Optarg);
break;
case 'T':
options.m_fConvergenceThreshold = float.Parse(g.Optarg);
break;
case 'e':
options.m_fEpsilon = float.Parse(g.Optarg);
break;
case 'x':
options.m_nMinimumThreadsSweep = int.Parse(g.Optarg);
break;
case 'i':
options.m_strFileName = g.Optarg;
break;
case 'I':
options.m_nMaxIterations = int.Parse(g.Optarg);
break;
case 'm':
options.m_nClusters = int.Parse(g.Optarg);
break;
case 'b':
options.m_bBinaryInput = true;
break;
case 'h':
PrintUsage();
return(null);
default:
Console.WriteLine("Unknown option: " + (Char)c);
PrintUsage();
return(null);
}
}
} catch (Exception e) {
Console.WriteLine("Options exception: " + e.Message);
PrintUsage();
return(null);
}
return(options);
}
ComputeNewCenters(
KMeansOptions options,
IEnumerable <Vector> vectors,
Vector[] centers,
int nCenters,
out int[] clusterIds,
out int nIterations
)
{
int nRank = vectors.ElementAt(0).Rank;
int nPoints = vectors.Count();
clusterIds = CreateInitialClusterIds(nPoints);
m_bMapPhasesComplete = false;
m_bUpdatePhasesComplete = false;
m_evtMapWorkAvailable.Reset();
m_evtMapWorkComplete.Reset();
m_barrier = new Barrier(options.m_nCurrentThreadsSweep, (b) => {
m_evtMapWorkAvailable.Reset();
m_evtMapWorkComplete.Set();
if (m_bVerbose)
{
Console.WriteLine("reached barrier!");
}
});
KMeansMapPartition[] partitions;
Thread[] mappers;
int nVectors = vectors.Count();
int nWorkers = options.m_nCurrentThreadsSweep;
int nPartitionSize = (int)Math.Round((double)nVectors / (double)nWorkers);
partitions = new KMeansMapPartition[nWorkers];
mappers = new Thread[nWorkers];
for (int i = 0; i < nWorkers; i++)
{
int nStartIndex = i * nPartitionSize;
int nWorkerPartitionSize = nPartitionSize;
if (nStartIndex + nPartitionSize > nVectors)
{
nWorkerPartitionSize = nVectors - nStartIndex;
}
partitions[i] = CreateMapPartitionDesc(options,
vectors,
centers,
clusterIds,
nCenters,
nRank,
i,
nStartIndex,
nWorkerPartitionSize);
mappers[i] = new Thread(ThreadedKMeans.MapPartition);
mappers[i].Start(partitions[i]);
}
KMeansUpdateCenters update;
update = CreateUpdateDesc(options,
partitions,
vectors,
clusterIds,
nCenters,
nRank,
0);
Thread updater = new Thread(ThreadedKMeans.UpdateCenters);
updater.Start(update);
m_pStopwatch.Start();
m_nStartParallelPhase = m_pStopwatch.ElapsedMilliseconds;
m_evtMapWorkAvailable.Set();
updater.Join();
m_pStopwatch.Stop();
nIterations = update.m_nIterations;
centers = update.m_sharedCenters;
m_nEndParallelPhase = m_pStopwatch.ElapsedMilliseconds;
return(centers);
}
///-------------------------------------------------------------------------------------------------
/// <summary> Constructor. </summary>
///
/// <remarks> Chris Rossbach ([email protected]), 8/2/2012. </remarks>
///
/// <param name="options"> Options for controlling the operation. </param>
///-------------------------------------------------------------------------------------------------
public SequentialKMeans(KMeansOptions options)
{
m_options = options;
}
public LINQKMeans(KMeansOptions options)
{
m_options = options;
}