def dataCreationTest(self, totalMB, workers, threadsPerWorker, testName):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
#we wish we could actually test that we achieve saturation here but we can't yet.
text = """size(Vector.range(%s, {_*_}))""" % (totalMB * 1024 * 1024 / 8)
t0 = time.time()
result,simulation = self.computeUsingSeveralWorkers(
"1+1",
s3,
workers,
timeout = 120,
memoryLimitMb = 55 * 1024 / workers,
threadCount = threadsPerWorker,
returnSimulation = True,
useInMemoryCache = False
)
try:
t0 = time.time()
result = simulation.compute(text, timeout=120)
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(testName, totalTimeToReturnResult, None)
finally:
simulation.teardown()
def test_splitToRowMajor(self):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
nRows = 100000
nColumns = 50
result, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
self.setupScript(nRows, nColumns),
s3,
1,
timeout=30,
memoryLimitMb=8 * 1024,
threadCount=4,
returnSimulation=True,
useInMemoryCache=False)
try:
self.assertTrue(result.isResult())
setup = result.asResult.result
t0 = time.time()
result = simulation.compute(self.splitToRowMajorScript(),
timeout=360,
chunks=setup)
totalTimeToReturnResult = time.time() - t0
self.assertTrue(result.isResult())
PerformanceTestReporter.recordTest(
"algorithms.text.splitToRowMajor.%srows_%scolumns" %
(nRows, nColumns), totalTimeToReturnResult, None)
finally:
simulation.teardown()
def dataCreationTest(self, totalMB, workers, threadsPerWorker, testName):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
#we wish we could actually test that we achieve saturation here but we can't yet.
text = """size(Vector.range(%s, {_*_}))""" % (totalMB * 1024 * 1024 /
8)
t0 = time.time()
result, simulation = self.computeUsingSeveralWorkers(
"1+1",
s3,
workers,
timeout=120,
memoryLimitMb=55 * 1024 / workers,
threadCount=threadsPerWorker,
returnSimulation=True,
useInMemoryCache=False)
try:
t0 = time.time()
result = simulation.compute(text, timeout=120)
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(testName,
totalTimeToReturnResult, None)
finally:
simulation.teardown()
def stringCreationAndSumTest(self, totalStrings, workers, threadsPerWorker, testName):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
#we wish we could actually test that we achieve saturation here but we can't yet.
text = """Vector.range(%s, String).sum(size)""" % totalStrings
t0 = time.time()
_, simulation = \
self.computeUsingSeveralWorkers(
"1+1",
s3,
workers,
timeout = 240,
memoryLimitMb = 55 * 1024 / workers,
threadCount = threadsPerWorker,
returnSimulation = True,
useInMemoryCache = False
)
try:
t0 = time.time()
result = simulation.compute(text, timeout=240)
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(testName, totalTimeToReturnResult, None)
finally:
simulation.teardown()
def test_class_sorting_is_fast(self):
sz = 10000000
#burn the calculation in.
self.classSortingTest(sz, useClass=True)
self.classSortingTest(sz, useClass=False)
t0 = time.time()
with PerformanceTestReporter.RecordAsPerfTest("python.datatasks.sort_class_instances"):
self.classSortingTest(sz, useClass=True)
t1 = time.time()
with PerformanceTestReporter.RecordAsPerfTest("python.datatasks.sort_integers"):
self.classSortingTest(sz, useClass=False)
t2 = time.time()
intTime = t2 - t1
classTime = t1 - t0
self.assertTrue(classTime < intTime * 5, "Times (ints) %s and (classes) %s were not very close." % (intTime, classTime))
print intTime, " to sort ints"
print classTime, " to sort class instances"
def test_transposeToColumnMajor(self):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
nRows = 100000
nColumns = 50
result, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
self.transposeSetupScript(nRows, nColumns),
s3, 1, timeout = 300, memoryLimitMb = 45 * 1024, threadCount = 30,
returnSimulation = True, useInMemoryCache = False)
try:
self.assertTrue(result.isResult())
rowMajor = result.asResult.result
t0 = time.time()
result = simulation.compute(
self.transposeRowMajorToColumnMajorScript(nRows, nColumns),
timeout = 500,
rowMajor = rowMajor
)
totalTimeToReturnResult = time.time() - t0
self.assertTrue(result.isResult())
PerformanceTestReporter.recordTest(
"algorithms.text.transposeRowMajorToColumnMajor.%srows_%scolumns" % (nRows, nColumns),
totalTimeToReturnResult, None)
finally:
simulation.teardown()
def stringCreationAndSumTest(self, totalStrings, workers, threadsPerWorker, testName):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
#we wish we could actually test that we achieve saturation here but we can't yet.
text = """Vector.range(%s, String).sum(size)""" % totalStrings
t0 = time.time()
_, simulation = \
self.computeUsingSeveralWorkers(
"1+1",
s3,
workers,
timeout = 240,
memoryLimitMb = 55 * 1024 / workers,
threadCount = threadsPerWorker,
returnSimulation = True,
useInMemoryCache = False
)
try:
t0 = time.time()
result = simulation.compute(text, timeout=240)
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(testName, totalTimeToReturnResult, None)
finally:
simulation.teardown()
def test_byteToStringAndBackInDifferentPatterns(self):
s3 = ActualS3Interface.ActualS3InterfaceFactory()
setupText = (
"""
let ds = Vector.range(3000000000, {UInt8(_%100)});
let dat = Vector.range(100, fun(block) {
Vector.range(1000000, fun(o) { let base = block * 10000000 + o * 10; (base, base + 10) })
});
(ds, dat, dat.sum())
"""
)
setupResults, simulation = self.computeUsingSeveralWorkers(
setupText,
s3,
1,
memoryLimitMb=45 * 1024,
threadCount=30,
returnSimulation=True,
ioTaskThreadOverride=8,
useInMemoryCache=False,
timeout=30,
objectStore=self.createObjectStore(s3)
)
try:
ds, dat, datSum = setupResults.asResult.result
t0 = time.time()
result = simulation.compute(
"size(datSum ~~ { ds[_[0],_[1]].dataAsString }) == size(datSum)",
timeout=120,
ds=ds,
dat=dat,
datSum=datSum
)
PerformanceTestReporter.recordTest(
"python.BigBox.DataAsString.FlatVector",
time.time() - t0,
None
)
t0 = time.time()
result = simulation.compute(
"size(dat ~~ {_ ~~ { ds[_[0],_[1]].dataAsString } }) == size(dat)",
timeout=120,
ds=ds,
dat=dat,
datSum=datSum
)
PerformanceTestReporter.recordTest(
"python.BigBox.DataAsString.NestedVector",
time.time() - t0,
None
)
finally:
simulation.teardown()
def downloadTaxiData(self,
filecount,
parse=False,
workers=1,
threadsPerWorker=30,
downloaderThreads=8):
s3 = ActualS3Interface.ActualS3InterfaceFactory()
bucketName = self.getTestDataBucket()
result, simulation = self.computeUsingSeveralWorkers(
"1+1",
s3,
workers,
memoryLimitMb=45 * 1024 / workers,
threadCount=threadsPerWorker,
returnSimulation=True,
ioTaskThreadOverride=downloaderThreads,
useInMemoryCache=False,
objectStore=self.createObjectStore(s3)
)
try:
dsText = (
"""let ds = """ + "+".join([
'datasets.s3("%s", "taxi_month_%s.csv")' % (bucketName, ix) for ix in range(1, filecount+1)
]) + ";"
)
text = dsText + "(ds, ds.sum(), size(ds))"
downloadTimeStart = time.time()
result = simulation.compute(text, timeout=240)
self.assertTrue(result.isResult())
downloadTimeEnd = time.time()
ds, dsSum, bytecount = result.asResult.result
if parse:
parseTimeStart = time.time()
result = simulation.compute("size(parsing.csv(ds))", timeout=240, ds=ds)
parseTimeEnd = time.time()
self.assertTrue(result.isResult())
PerformanceTestReporter.recordTest(
"python.BigBox.LargeS3.ParseTaxidata." + str(filecount),
parseTimeEnd - parseTimeStart,
None
)
else:
bytecount = bytecount.pyval
PerformanceTestReporter.recordTest(
"python.BigBox.LargeS3.TaxiSecondsPerGB." + str(filecount),
(downloadTimeEnd - downloadTimeStart) / (bytecount / 1024 / 1024.0 / 1024.0),
None
)
finally:
simulation.teardown()
def stringToInt64ParsingTest(self, threads, testName):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
#we wish we could actually test that we achieve saturation here but we can't yet.
text = """
let doALoop = fun(x) {
//pass 's' through a vector so that the compiler can't tell what it is
let s = ["2013"][0];
let res = 0
for ix in sequence(x) {
if (ix == 0)
s = s + String(ix)
res = res + Int64(s) + ix
}
res
};
Vector.range(__thread_count__) ~~ {doALoop(20000000 + _)}
""".replace("__thread_count__", str(threads))
_, simulation = \
self.computeUsingSeveralWorkers(
"1+1",
s3,
1,
timeout = 240,
memoryLimitMb = 55 * 1024,
threadCount = 30,
returnSimulation = True,
useInMemoryCache = False
)
t0 = time.time()
_, simulation = \
self.computeUsingSeveralWorkers(
"1+1",
s3,
1,
timeout = 240,
memoryLimitMb = 55 * 1024,
threadCount = 30,
returnSimulation = True,
useInMemoryCache = False
)
try:
t0 = time.time()
result = simulation.compute(text, timeout=240)
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(testName,
totalTimeToReturnResult, None)
finally:
simulation.teardown()
def testMemoryUpdate(self):
measurementTime = self.measurementTime
for cores in self.coreList:
for allocSize in self.allocSizeList:
PerformanceTestReporter.recordTest(
"python.BigBox.MemoryUpdate.SecondsPerGB.%sCore_%sMB" %
(cores, allocSize),
self.measureMemoryUpdatePerformance(
measurementTime, 1024 * 1024 * allocSize, cores), None)
def testMemoryUpdate(self):
measurementTime = self.measurementTime
for cores in self.coreList:
for allocSize in self.allocSizeList:
PerformanceTestReporter.recordTest(
"python.BigBox.MemoryUpdate.SecondsPerGB.%sCore_%sMB" % (cores, allocSize),
self.measureMemoryUpdatePerformance(measurementTime, 1024 * 1024 * allocSize, cores),
None
)
def testMmapAllocation(self):
measurementTime = self.measurementTime
for cores in self.coreList:
for allocSize in self.allocSizeList:
PerformanceTestReporter.recordTest(
"python.BigBox.MmapAlloc.SecondsPerGB.%sCore_%sMB" % (cores, allocSize),
self.measureMmapPerformance(measurementTime, 1024 * 1024 * allocSize, cores, False),
None
)
def gbmRegressionFittingTest(self, nRows, nColumns, depth, nThreads,
maxBoosts):
testName = self.getTestName(nRows, nColumns, depth, maxBoosts,
nThreads)
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
result, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
self.dataGenerationScript(nRows, nColumns),
s3,
1,
timeout=360,
memoryLimitMb=30 * 1024,
threadCount=nThreads,
returnSimulation=True,
useInMemoryCache=False)
try:
self.assertTrue(result.isResult())
dfPredictors, dfResponse = result.asResult.result
fitter = simulation.compute(
self.regressionScript(depth, 1),
timeout=360,
dfResponse=dfResponse,
dfPredictors=dfPredictors).asResult.result
t0 = time.time()
for nBoosts in range(1, maxBoosts):
testName = self.getTestName(nRows, nColumns, depth, nBoosts,
nThreads)
predictions = simulation.compute(
"fitter.predictionsAndPseudoresiduals()",
timeout=360,
fitter=fitter).asResult.result
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(testName + "_predict",
totalTimeToReturnResult,
None)
fitter = simulation.compute(
"fitter.nextGivenPredictions(predictions)",
timeout=360,
fitter=fitter,
predictions=predictions).asResult.result
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(testName,
totalTimeToReturnResult,
None)
finally:
simulation.teardown()
def stringToInt64ParsingTest(self, threads, testName):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
#we wish we could actually test that we achieve saturation here but we can't yet.
text = """
let doALoop = fun(x) {
//pass 's' through a vector so that the compiler can't tell what it is
let s = ["2013"][0];
let res = 0
for ix in sequence(x) {
if (ix == 0)
s = s + String(ix)
res = res + Int64(s) + ix
}
res
};
Vector.range(__thread_count__) ~~ {doALoop(20000000 + _)}
""".replace("__thread_count__", str(threads))
_, simulation = \
self.computeUsingSeveralWorkers(
"1+1",
s3,
1,
timeout = 240,
memoryLimitMb = 55 * 1024,
threadCount = 30,
returnSimulation = True,
useInMemoryCache = False
)
t0 = time.time()
_, simulation = \
self.computeUsingSeveralWorkers(
"1+1",
s3,
1,
timeout = 240,
memoryLimitMb = 55 * 1024,
threadCount = 30,
returnSimulation = True,
useInMemoryCache = False
)
try:
t0 = time.time()
result = simulation.compute(text, timeout=240)
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(testName, totalTimeToReturnResult, None)
finally:
simulation.teardown()
def test_knows_is_reporting(self):
with SetEnv(
PerformanceTestReporter.TEST_DATA_LOCATION_ENVIRONMENT_VARIABLE,
None):
self.assertFalse(PerformanceTestReporter.isCurrentlyTesting())
with SetEnv(
PerformanceTestReporter.TEST_DATA_LOCATION_ENVIRONMENT_VARIABLE,
"./testResults.json"):
self.assertTrue(PerformanceTestReporter.isCurrentlyTesting())
def testMmapAllocation(self):
measurementTime = self.measurementTime
for cores in self.coreList:
for allocSize in self.allocSizeList:
PerformanceTestReporter.recordTest(
"python.BigBox.MmapAlloc.SecondsPerGB.%sCore_%sMB" %
(cores, allocSize),
self.measureMmapPerformance(measurementTime,
1024 * 1024 * allocSize, cores,
False), None)
def test_cant_report_nonsensical_timing(self):
tempDir = tempfile.mkdtemp()
tempFile = os.path.join(tempDir, "data.json")
with SetEnv(
PerformanceTestReporter.TEST_DATA_LOCATION_ENVIRONMENT_VARIABLE,
tempFile
):
with self.assertRaises(Exception):
PerformanceTestReporter.recordTest("test1","not a float",None)
def test_effectiveParallelism(self):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
#do a burn-in run
self.computeUsingSeveralWorkers("""
let v = Vector.range(5000000, { (1,_) } );
let f = fun(ix) {
let res = 0
for x in sequence( (ix - 2000) >>> 0, ix )
res = res + size(v[x])
res
}
Vector.range(size(v), f).sum()
""",
s3,
2,
wantsStats=True,
timeout=240,
memoryLimitMb=500)[1]
t0 = time.time()
stats = self.computeUsingSeveralWorkers("""
let v = Vector.range(5000000, { (1,_) } );
let f = fun(ix) {
let res = 0
for x in sequence( (ix - 2000) >>> 0, ix )
res = res + size(v[x])
res
}
Vector.range(size(v), f).sum()
""",
s3,
2,
wantsStats=True,
timeout=240,
memoryLimitMb=500)[1]
timeElapsed = time.time() - t0
totalTime = stats.timeSpentInInterpreter + stats.timeSpentInCompiler
effParallelism = totalTime / timeElapsed
PerformanceTestReporter.recordTest(
"python.cumulus.EffectiveParallelism.elapsed", timeElapsed, None)
PerformanceTestReporter.recordTest(
"python.cumulus.EffectiveParallelism.effectiveCores",
effParallelism, {},
units='count')
def gbmRegressionFittingTest(self, nRows, nColumns, depth, nThreads, maxBoosts):
testName = self.getTestName(nRows, nColumns, depth, maxBoosts, nThreads)
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
result, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
self.dataGenerationScript(nRows, nColumns),
s3,
1,
timeout = 360,
memoryLimitMb = 30 * 1024,
threadCount = nThreads,
returnSimulation = True,
useInMemoryCache = False
)
try:
self.assertTrue(result.isResult())
dfPredictors, dfResponse = result.asResult.result
fitter = simulation.compute(
self.regressionScript(depth, 1),
timeout = 360,
dfResponse = dfResponse,
dfPredictors = dfPredictors
).asResult.result
t0 = time.time()
for nBoosts in range(1, maxBoosts):
testName = self.getTestName(nRows, nColumns, depth, nBoosts, nThreads)
predictions = simulation.compute(
"fitter.predictionsAndPseudoresiduals()",
timeout = 360,
fitter = fitter
).asResult.result
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(
testName + "_predict", totalTimeToReturnResult, None)
fitter = simulation.compute(
"fitter.nextGivenPredictions(predictions)",
timeout = 360,
fitter = fitter,
predictions = predictions
).asResult.result
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(
testName, totalTimeToReturnResult, None)
finally:
simulation.teardown()
def test_CalculationRicochet(self):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
text = """
let f = fun(ct, seed = 1) {
let x = 0
let res = []
let it = iterator(math.random.UniformReal(0, size(v), seed))
for ix in sequence(ct) {
let x = Int64(pull it)
res = res :: (x / Float64(size(v)), v[x])
}
return res
}
v[2]
f(__count__,__seed__)
"""
vResult, sim = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
"Vector.range(125000000, math.log)",
s3,
4,
timeout=120,
memoryLimitMb=400,
threadCount=1,
useInMemoryCache=True,
returnSimulation=True)
try:
v = vResult.asResult.result
t0 = time.time()
sim.compute(text.replace("__seed__",
"1").replace("__count__", "1000"),
timeout=120,
v=v)
PerformanceTestReporter.recordTest(
"python.InMemoryCumulus.Ricochet1000.Pass1",
time.time() - t0, None)
t0 = time.time()
sim.compute(text.replace("__seed__",
"2").replace("__count__", "1000"),
timeout=120,
v=v)
PerformanceTestReporter.recordTest(
"python.InMemoryCumulus.Ricochet1000.Pass2",
time.time() - t0, None)
finally:
sim.teardown()
def largeDatasetJoinTest(self, mbOfData, columns, threads, machineCount, ratio = .5):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
t0 = time.time()
result, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
self.dataGenerationScript(mbOfData, columns),
s3,
machineCount,
timeout = 360,
memoryLimitMb = mbOfData / ratio / machineCount,
#channelThroughputMBPerSecond = 100.0,
threadCount = threads,
returnSimulation = True,
useInMemoryCache = False,
disableEventHandler = True
)
try:
self.assertTrue(result.isResult())
data = result.asResult.result
joinScript = """
let leftDF = dataframe.DataFrame(data[,size(data)/2])
let rightDF = dataframe.DataFrame(data[size(data)/2,])
size(leftDF.join(rightDF, on: "C0", how: `outer, chunkSize: 1000000, areSorted:true))
"""
t0 = time.time()
result = simulation.compute(
joinScript,
timeout=1080,
data=data
)
totalTimeToReturnResult = time.time() - t0
logging.info("Total time to join: %s", totalTimeToReturnResult)
self.assertTrue(result.isResult(), result)
PerformanceTestReporter.recordTest(
"algorithms.Join.inMemory_%sMB_%scols_%sthreads_%smachines" %
(mbOfData, columns,threads,machineCount),
totalTimeToReturnResult,
None
)
finally:
dfResponse = None
dfPredictors = None
result = None
simulation.teardown()
def test_disk_scans(self):
s3 = ActualS3Interface.ActualS3InterfaceFactory()
objectStore = S3ObjectStore.S3ObjectStore(
s3,
Setup.config().userDataS3Bucket,
prefix="test_object_cache/"
)
_, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
"1+1",
s3,
1,
memoryLimitMb=1 * 1024,
threadCount=30,
returnSimulation=True,
ioTaskThreadOverride=8,
objectStore=objectStore,
useInMemoryCache=False #use an actual disk cache for this
)
try:
gigabytes = 8
t0 = time.time()
resultVectors = []
for ix in range(gigabytes):
result = simulation.compute("Vector.range(125000000 + %s)" % ix, timeout=120)
resultVectors.append(result.asResult.result)
t1 = time.time()
intResults = []
for vec in resultVectors:
result = simulation.compute("v.sum()", timeout = 120, v=vec)
intResults.append(result.asResult.result.pyval)
self.assertTrue(len(intResults) == gigabytes)
PerformanceTestReporter.recordTest(
"python.BigBox.Disk.Write.10GB",
t1 - t0,
None
)
PerformanceTestReporter.recordTest(
"python.BigBox.Disk.WriteAndScan.%sGB" % gigabytes,
time.time() - t0,
None
)
finally:
simulation.teardown()
def gbmRegressionFittingTest(self,
nRows,
nColumns,
depth,
nThreads,
nBoosts,
copies,
report=True):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
result, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
self.dataGenerationScript(nRows, nColumns),
s3,
1,
timeout=360,
memoryLimitMb=30 * 1024,
threadCount=nThreads,
returnSimulation=True,
useInMemoryCache=False)
try:
self.assertTrue(result.isResult())
dfPredictors, dfResponse = result.asResult.result
builder = simulation.compute(
self.regressionScript(depth, nBoosts),
timeout=360,
dfResponse=dfResponse,
dfPredictors=dfPredictors).asResult.result
t0 = time.time()
testName = self.getTestName(nRows, nColumns, depth, nBoosts,
nThreads, copies)
result = simulation.compute(
"Vector.range(%s).apply(fun(x) { builder.fit(dfPredictors[,-x-1], dfResponse[,-x-1]) })"
% copies,
timeout=360,
builder=builder,
dfPredictors=dfPredictors,
dfResponse=dfResponse,
).asResult.result
totalTimeToReturnResult = time.time() - t0
if report:
PerformanceTestReporter.recordTest(testName,
totalTimeToReturnResult,
None)
finally:
simulation.teardown()
def test_effectiveParallelism(self):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
#do a burn-in run
self.computeUsingSeveralWorkers("""
let v = Vector.range(5000000, { (1,_) } );
let f = fun(ix) {
let res = 0
for x in sequence( (ix - 2000) >>> 0, ix )
res = res + size(v[x])
res
}
Vector.range(size(v), f).sum()
""", s3, 2, wantsStats = True, timeout=240, memoryLimitMb=500
)[1]
t0 = time.time()
stats = self.computeUsingSeveralWorkers("""
let v = Vector.range(5000000, { (1,_) } );
let f = fun(ix) {
let res = 0
for x in sequence( (ix - 2000) >>> 0, ix )
res = res + size(v[x])
res
}
Vector.range(size(v), f).sum()
""", s3, 2, wantsStats = True, timeout=240, memoryLimitMb=500
)[1]
timeElapsed = time.time() - t0
totalTime = stats.timeSpentInInterpreter + stats.timeSpentInCompiler
effParallelism = totalTime / timeElapsed
PerformanceTestReporter.recordTest(
"python.cumulus.EffectiveParallelism.elapsed",
timeElapsed,
None
)
PerformanceTestReporter.recordTest(
"python.cumulus.EffectiveParallelism.effectiveCores",
effParallelism,
{},
units='count'
)
def test_some_throughput(self):
def toTest(n):
FORA.eval("""let v = [0, 0.0]; let res = 0;
for ix in sequence(%s * 100000000) {
res = res + v[0] + v[1];
}
res""" % n)
PerformanceTestReporter.testThroughput(
"fora_lang.LangTestPerf.vector.heterogeneousVectorAccessThroughput_100mm",
toTest,
maxNToSearch=10,
timeoutInSec=5.0)
def largeDatasetJoinTest(self,
mbOfData,
columns,
threads,
machineCount,
ratio=.5):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
t0 = time.time()
result, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
self.dataGenerationScript(mbOfData, columns),
s3,
machineCount,
timeout=360,
memoryLimitMb=mbOfData / ratio / machineCount,
#channelThroughputMBPerSecond = 100.0,
threadCount=threads,
returnSimulation=True,
useInMemoryCache=False,
disableEventHandler=True)
try:
self.assertTrue(result.isResult())
data = result.asResult.result
joinScript = """
let leftDF = dataframe.DataFrame(data[,size(data)/2])
let rightDF = dataframe.DataFrame(data[size(data)/2,])
size(leftDF.join(rightDF, on: "C0", how: `outer, chunkSize: 1000000, areSorted:true))
"""
t0 = time.time()
result = simulation.compute(joinScript, timeout=1080, data=data)
totalTimeToReturnResult = time.time() - t0
logging.info("Total time to join: %s", totalTimeToReturnResult)
self.assertTrue(result.isResult(), result)
PerformanceTestReporter.recordTest(
"algorithms.Join.inMemory_%sMB_%scols_%sthreads_%smachines" %
(mbOfData, columns, threads, machineCount),
totalTimeToReturnResult, None)
finally:
dfResponse = None
dfPredictors = None
result = None
simulation.teardown()
def test_some_throughput(self):
def toTest(n):
FORA.eval("""let v = [0, 0.0]; let res = 0;
for ix in sequence(%s * 100000000) {
res = res + v[0] + v[1];
}
res""" % n)
PerformanceTestReporter.testThroughput(
"fora_lang.LangTestPerf.vector.heterogeneousVectorAccessThroughput_100mm",
toTest,
maxNToSearch=10,
timeoutInSec=5.0
)
def test_throughputThrowsIfNonePassed(self):
tempDir = tempfile.mkdtemp()
tempFile = os.path.join(tempDir, "data.json")
with SetEnv(
PerformanceTestReporter.TEST_DATA_LOCATION_ENVIRONMENT_VARIABLE,
tempFile
):
def testFunOfN(n):
raise PerformanceTestReporter.TimedOutException("timed out!!")
with self.assertRaises(AssertionError):
PerformanceTestReporter.testThroughput(
"test1", testFunOfN = testFunOfN)
def vector_dot_product(self, dimension):
with self.ufora.remotely:
a = np.arange(dimension)
b = np.arange(dimension)
def f(n):
with self.ufora.remotely:
for _ in xrange(n):
np.dot(a, b)
PerformanceTestReporter.testThroughput(
"pyfora.numpy.vector_dot_product_%d" % dimension,
f,
maxNToSearch=20,
timeoutInSec=20.0)
def array_binary_operation(self, dimension, op, test_name):
with self.ufora.remotely:
a = np.arange(dimension)
b = np.arange(dimension)
def f(n):
with self.ufora.remotely:
for _ in xrange(n):
op(a, b)
PerformanceTestReporter.testThroughput("pyfora.numpy.%s_%d" %
(test_name, dimension),
f,
maxNToSearch=20,
timeoutInSec=20.0)
def largeDatasetBigLMTest(self, mbOfData, columns, threads, testName):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
t0 = time.time()
result, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
self.dataGenerationScript(mbOfData, columns),
s3,
1,
timeout=360,
memoryLimitMb=50 * 1024,
threadCount=threads,
returnSimulation=True,
useInMemoryCache=False)
if testName is not None:
PerformanceTestReporter.recordTest(testName + "_create",
time.time() - t0, None)
try:
self.assertTrue(result.isResult())
dfResponse, dfPredictors = result.asResult.result
regressionScript = """
let model = math.regression.LinearRegression(dfPredictors, dfResponse, fitIntercept: false);
let coefficients = model.coefficients();
coefficients[0]
"""
t0 = time.time()
result = simulation.compute(regressionScript,
timeout=1080,
dfResponse=dfResponse,
dfPredictors=dfPredictors)
totalTimeToReturnResult = time.time() - t0
self.assertTrue(result.isResult())
if testName is not None:
PerformanceTestReporter.recordTest(testName,
totalTimeToReturnResult,
None)
finally:
dfResponse = None
dfPredictors = None
result = None
simulation.teardown()
def largeDatasetBigLMTest(self, mbOfData, columns, threads, testName):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
t0 = time.time()
result, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
self.dataGenerationScript(mbOfData, columns),
s3,
1,
timeout = 360,
memoryLimitMb = 50 * 1024,
threadCount = threads,
returnSimulation = True,
useInMemoryCache = False
)
if testName is not None:
PerformanceTestReporter.recordTest(testName + "_create", time.time() - t0, None)
try:
self.assertTrue(result.isResult())
dfResponse, dfPredictors = result.asResult.result
regressionScript = """
let model = math.regression.LinearRegression(dfPredictors, dfResponse, fitIntercept: false);
let coefficients = model.coefficients();
coefficients[0]
"""
t0 = time.time()
result = simulation.compute(
regressionScript,
timeout=1080,
dfResponse=dfResponse,
dfPredictors=dfPredictors
)
totalTimeToReturnResult = time.time() - t0
self.assertTrue(result.isResult())
if testName is not None:
PerformanceTestReporter.recordTest(testName, totalTimeToReturnResult, None)
finally:
dfResponse = None
dfPredictors = None
result = None
simulation.teardown()
def array_binary_operation(self, dimension, op, test_name):
with self.ufora.remotely:
a = np.arange(dimension)
b = np.arange(dimension)
def f(n):
with self.ufora.remotely:
for _ in xrange(n):
op(a, b)
PerformanceTestReporter.testThroughput(
"pyfora.numpy.%s_%d" % (test_name, dimension),
f,
maxNToSearch=20,
timeoutInSec=20.0
)
def vector_dot_product(self, dimension):
with self.ufora.remotely:
a = np.arange(dimension)
b = np.arange(dimension)
def f(n):
with self.ufora.remotely:
for _ in xrange(n):
np.dot(a, b)
PerformanceTestReporter.testThroughput(
"pyfora.numpy.vector_dot_product_%d" % dimension,
f,
maxNToSearch=20,
timeoutInSec=20.0
)
def roundtripConvert(self, pyObject, testName):
try:
_, timings = self._roundtripConvert(pyObject)
for k in sorted(timings):
print k, timings[k]
PerformanceTestReporter.recordTest(
testName=testName+"."+k,
elapsedTime=timings[k],
metadata=None
)
except:
import traceback
traceback.print_exc()
self.assertTrue(False)
def diskThroughputTest(self, gb):
if os.getenv("CUMULUS_DATA_DIR") is None:
dataDir = tempfile.mkdtemp()
else:
dataDir = os.getenv("CUMULUS_DATA_DIR")
dataDir = os.path.join(dataDir, str(uuid.uuid4()))
diskCache = CumulusNative.DiskOfflineCache(
callbackScheduler,
dataDir,
100 * 1024 * 1024 * 1024,
100000
)
fiftyMegabytes = ForaNative.encodeStringInSerializedObject(" " * 1024 * 1024 * 50)
logging.info("Writing to %s", dataDir)
try:
t0 = time.time()
for ix in range(gb * 20):
diskCache.store(
ForaNative.PageId(HashNative.Hash.sha1(str(ix)), 50 * 1024 * 1024, 50 * 1024 * 1024),
fiftyMegabytes
)
PerformanceTestReporter.recordTest(
"python.BigBox.Disk.Write%sGB" % gb,
time.time() - t0,
None
)
t0 = time.time()
for ix in range(gb * 20):
diskCache.loadIfExists(
ForaNative.PageId(HashNative.Hash.sha1(str(ix)), 50 * 1024 * 1024, 50 * 1024 * 1024)
)
PerformanceTestReporter.recordTest(
"python.BigBox.Disk.Read%sGB" % gb,
time.time() - t0,
None
)
finally:
shutil.rmtree(dataDir)
def regressionTreePredictionTest(self,
mbOfData,
columns,
testName,
treeDepth,
threads,
minSamplesSplit=50):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
result, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
self.dataGenerationScript(mbOfData, columns),
s3,
1,
timeout=360,
memoryLimitMb=45 * 1024,
threadCount=threads,
returnSimulation=True,
useInMemoryCache=False)
try:
self.assertTrue(result.isResult())
dfResponse, dfPredictors = result.asResult.result
fitTree = simulation.compute(self.regressionScript(
treeDepth, minSamplesSplit - 1),
timeout=120,
dfResponse=dfResponse,
dfPredictors=dfPredictors)
def predictionScript(dirtyFlag=1):
return ";(%s; fitRegressionTree.predict(dfPredictors));" % dirtyFlag
t0 = time.time()
result = simulation.compute(
predictionScript(),
timeout=120,
dfPredictors=dfPredictors,
fitRegressionTree=fitTree.asResult.result)
totalTimeToReturnResult = time.time() - t0
self.assertTrue(result.isResult())
PerformanceTestReporter.recordTest(testName,
totalTimeToReturnResult, None)
finally:
simulation.teardown()
def test_disk_scans(self):
s3 = ActualS3Interface.ActualS3InterfaceFactory()
objectStore = S3ObjectStore.S3ObjectStore(
s3, Setup.config().userDataS3Bucket, prefix="test_object_cache/")
_, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
"1+1",
s3,
1,
memoryLimitMb=1 * 1024,
threadCount=30,
returnSimulation=True,
ioTaskThreadOverride=8,
objectStore=objectStore,
useInMemoryCache=False #use an actual disk cache for this
)
try:
gigabytes = 8
t0 = time.time()
resultVectors = []
for ix in range(gigabytes):
result = simulation.compute("Vector.range(125000000 + %s)" %
ix,
timeout=120)
resultVectors.append(result.asResult.result)
t1 = time.time()
intResults = []
for vec in resultVectors:
result = simulation.compute("v.sum()", timeout=120, v=vec)
intResults.append(result.asResult.result.pyval)
self.assertTrue(len(intResults) == gigabytes)
PerformanceTestReporter.recordTest("python.BigBox.Disk.Write.10GB",
t1 - t0, None)
PerformanceTestReporter.recordTest(
"python.BigBox.Disk.WriteAndScan.%sGB" % gigabytes,
time.time() - t0, None)
finally:
simulation.teardown()
def test_reporting_to_file(self):
tempDir = tempfile.mkdtemp()
tempFile = os.path.join(tempDir, "data.json")
with SetEnv(
PerformanceTestReporter.TEST_DATA_LOCATION_ENVIRONMENT_VARIABLE,
tempFile
):
PerformanceTestReporter.recordTest("test1.result", 10.0, {"some":"metadata"})
PerformanceTestReporter.recordTest("test1.result", None, {"some":"metadata"})
testData = PerformanceTestReporter.loadTestsFromFile(tempFile)
self.assertEqual(testData,
[{"name":"test1.result", "time":10.0, "metadata": {"some":"metadata"}},
{"name":"test1.result", "time":None, "metadata": {"some":"metadata"}}
])
def test_CalculationRicochet(self):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
text = """
let f = fun(ct, seed = 1) {
let x = 0
let res = []
let it = iterator(math.random.UniformReal(0, size(v), seed))
for ix in sequence(ct) {
let x = Int64(pull it)
res = res :: (x / Float64(size(v)), v[x])
}
return res
}
v[2]
f(__count__,__seed__)
"""
vResult, sim = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
"Vector.range(125000000, math.log)",
s3,
4,
timeout = 120,
memoryLimitMb=400,
threadCount = 1,
useInMemoryCache = True,
returnSimulation = True
)
try:
v = vResult.asResult.result
t0 = time.time()
sim.compute(text.replace("__seed__", "1").replace("__count__", "1000"), timeout = 120, v = v)
PerformanceTestReporter.recordTest("python.InMemoryCumulus.Ricochet1000.Pass1", time.time() - t0,None)
t0 = time.time()
sim.compute(text.replace("__seed__", "2").replace("__count__", "1000"), timeout = 120, v = v)
PerformanceTestReporter.recordTest("python.InMemoryCumulus.Ricochet1000.Pass2", time.time() - t0,None)
finally:
sim.teardown()
def gbmRegressionFittingTest(self, nRows, nColumns, depth, nThreads, nBoosts, copies, report=True):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
result, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
self.dataGenerationScript(nRows, nColumns),
s3,
1,
timeout = 360,
memoryLimitMb = 30 * 1024,
threadCount = nThreads,
returnSimulation = True,
useInMemoryCache = False
)
try:
self.assertTrue(result.isResult())
dfPredictors, dfResponse = result.asResult.result
builder = simulation.compute(
self.regressionScript(depth, nBoosts),
timeout = 360,
dfResponse = dfResponse,
dfPredictors = dfPredictors
).asResult.result
t0 = time.time()
testName = self.getTestName(nRows, nColumns, depth, nBoosts, nThreads, copies)
result = simulation.compute(
"Vector.range(%s).apply(fun(x) { builder.fit(dfPredictors[,-x-1], dfResponse[,-x-1]) })"
% copies,
timeout = 360,
builder=builder,
dfPredictors=dfPredictors,
dfResponse=dfResponse,
).asResult.result
totalTimeToReturnResult = time.time() - t0
if report:
PerformanceTestReporter.recordTest(testName, totalTimeToReturnResult, None)
finally:
simulation.teardown()
def matrix_dot_product(self, dimension):
with self.ufora.remotely:
a = np.arange(dimension**2).reshape(
(int(dimension), int(dimension)))
b = np.arange(dimension**2).reshape(
(int(dimension), int(dimension)))
def f(n):
with self.ufora.remotely:
for _ in xrange(n):
np.dot(a, b)
PerformanceTestReporter.testThroughput(
"pyfora.numpy.matrix_dot_product_%dx%d" % (dimension, dimension),
f,
maxNToSearch=20,
timeoutInSec=20.0)
def matrix_dot_product(self, dimension):
with self.ufora.remotely:
a = np.arange(dimension**2).reshape(
(int(dimension), int(dimension)))
b = np.arange(dimension**2).reshape(
(int(dimension), int(dimension)))
def f(n):
with self.ufora.remotely:
for _ in xrange(n):
np.dot(a, b)
PerformanceTestReporter.testThroughput(
"pyfora.numpy.matrix_dot_product_%dx%d" % (dimension, dimension),
f,
maxNToSearch=20,
timeoutInSec=20.0
)
def regressionTreePredictionTest(self, mbOfData, columns, testName,
treeDepth, threads, minSamplesSplit=50):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
result, simulation = InMemoryCumulusSimulation.computeUsingSeveralWorkers(
self.dataGenerationScript(mbOfData, columns),
s3,
1,
timeout = 360,
memoryLimitMb = 45 * 1024,
threadCount = threads,
returnSimulation = True,
useInMemoryCache = False
)
try:
self.assertTrue(result.isResult())
dfResponse, dfPredictors = result.asResult.result
fitTree = simulation.compute(
self.regressionScript(treeDepth, minSamplesSplit - 1),
timeout=120,
dfResponse=dfResponse,
dfPredictors=dfPredictors
)
def predictionScript(dirtyFlag=1):
return ";(%s; fitRegressionTree.predict(dfPredictors));" % dirtyFlag
t0 = time.time()
result = simulation.compute(
predictionScript(),
timeout=120,
dfPredictors=dfPredictors,
fitRegressionTree=fitTree.asResult.result
)
totalTimeToReturnResult = time.time() - t0
self.assertTrue(result.isResult())
PerformanceTestReporter.recordTest(testName, totalTimeToReturnResult, None)
finally:
simulation.teardown()
def test_throughputDoesNotFailOnTimeoutIfSomePassed(self):
tempDir = tempfile.mkdtemp()
tempFile = os.path.join(tempDir, "data.json")
with SetEnv(
PerformanceTestReporter.TEST_DATA_LOCATION_ENVIRONMENT_VARIABLE,
tempFile
):
def testFunOfN(n):
if n < 10:
pass
else:
raise PerformanceTestReporter.TimedOutException("timed out!!")
PerformanceTestReporter.testThroughput(
"test1", testFunOfN = testFunOfN)
testData = PerformanceTestReporter.loadTestsFromFile(tempFile)
self.assertEqual(len(testData), 1)
def validatePerfForExpression(self, testName, foraExpr, callResult=False):
subprocessArgs = \
[sys.executable, ValidatePerf.__file__, foraExpr]
if callResult:
subprocessArgs.append("callResult")
self.validateTimingsForSubprocessCall(
testName, subprocessArgs,
PerformanceTestReporter.getCurrentStackframeFileAndLine(
framesAbove=2))
def stringToDatetimeParsingTest(self, threads, testName):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
#we wish we could actually test that we achieve saturation here but we can't yet.
text = """
let s = ["2013-01-01 15:18:10"][0];
let doALoop = fun(x) {
let res = 0
for ix in sequence(x) {
res = res + DateTime(s).year
}
res
};
Vector.range(__thread_count__) ~~ {doALoop(1000000 + _)}
""".replace("__thread_count__", str(threads))
t0 = time.time()
_, simulation = \
self.computeUsingSeveralWorkers(
"1+1",
s3,
1,
timeout = 240,
memoryLimitMb = 55 * 1024 / workers,
threadCount = 30,
returnSimulation = True,
useInMemoryCache = False
)
try:
t0 = time.time()
result = simulation.compute(text, timeout=240)
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(testName,
totalTimeToReturnResult, None)
finally:
simulation.teardown()
def validatePerfForExpression(self, testName, foraExpr, callResult = False):
subprocessArgs = \
[sys.executable, ValidatePerf.__file__, foraExpr];
if callResult:
subprocessArgs.append("callResult")
self.validateTimingsForSubprocessCall(
testName,
subprocessArgs,
PerformanceTestReporter.getCurrentStackframeFileAndLine(framesAbove=2)
)
def stringToDatetimeParsingTest(self, threads, testName):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
#we wish we could actually test that we achieve saturation here but we can't yet.
text = """
let s = ["2013-01-01 15:18:10"][0];
let doALoop = fun(x) {
let res = 0
for ix in sequence(x) {
res = res + DateTime(s).year
}
res
};
Vector.range(__thread_count__) ~~ {doALoop(1000000 + _)}
""".replace("__thread_count__", str(threads))
t0 = time.time()
_, simulation = \
self.computeUsingSeveralWorkers(
"1+1",
s3,
1,
timeout = 240,
memoryLimitMb = 55 * 1024 / workers,
threadCount = 30,
returnSimulation = True,
useInMemoryCache = False
)
try:
t0 = time.time()
result = simulation.compute(text, timeout=240)
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(testName, totalTimeToReturnResult, None)
finally:
simulation.teardown()
def validateTimingsForSubprocessCall(
self,
testName,
subprocessArgs,
meta,
timeout = 600.0
):
resultCode, out, err = SubprocessRunner.callAndReturnResultAndOutput(
subprocessArgs,
timeout = timeout
)
if resultCode != 0:
meta.update({"failure": "subprocess call returned error"})
if PerformanceTestReporter.isCurrentlyTesting():
PerformanceTestReporter.recordTest(
testName,
None,
meta
)
assert resultCode == 0, err
logging.info("Actual time was %s for %s", out[0], subprocessArgs)
measuredTiming = float(out[0]) / self.baseTiming
if PerformanceTestReporter.isCurrentlyTesting():
PerformanceTestReporter.recordTest(
"fora_lang." + testName,
float(out[0]),
meta
)
def diskThroughputTest(self, gb):
if os.getenv("CUMULUS_DATA_DIR") is None:
dataDir = tempfile.mkdtemp()
else:
dataDir = os.getenv("CUMULUS_DATA_DIR")
dataDir = os.path.join(dataDir, str(uuid.uuid4()))
diskCache = CumulusNative.DiskOfflineCache(callbackScheduler, dataDir,
100 * 1024 * 1024 * 1024,
100000)
fiftyMegabytes = ForaNative.encodeStringInSerializedObject(" " * 1024 *
1024 * 50)
logging.info("Writing to %s", dataDir)
try:
t0 = time.time()
for ix in range(gb * 20):
diskCache.store(
ForaNative.PageId(HashNative.Hash.sha1(str(ix)),
50 * 1024 * 1024, 50 * 1024 * 1024),
fiftyMegabytes)
PerformanceTestReporter.recordTest(
"python.BigBox.Disk.Write%sGB" % gb,
time.time() - t0, None)
t0 = time.time()
for ix in range(gb * 20):
diskCache.loadIfExists(
ForaNative.PageId(HashNative.Hash.sha1(str(ix)),
50 * 1024 * 1024, 50 * 1024 * 1024))
PerformanceTestReporter.recordTest(
"python.BigBox.Disk.Read%sGB" % gb,
time.time() - t0, None)
finally:
shutil.rmtree(dataDir)
def loopScalabilityTestTest(self, threads, testName):
s3 = InMemoryS3Interface.InMemoryS3InterfaceFactory()
text = """
let doALoop = fun(x) {
let res = 0
for ix in sequence(x) {
res = res + ix + 1
}
res
};
Vector.range(__thread_count__) ~~ {doALoop(1000000000 + _)}
""".replace("__thread_count__", str(threads))
t0 = time.time()
_, simulation = \
self.computeUsingSeveralWorkers(
"1+1",
s3,
1,
timeout = 240,
memoryLimitMb = 55 * 1024,
threadCount = 30,
returnSimulation = True,
useInMemoryCache = False
)
try:
t0 = time.time()
result = simulation.compute(text, timeout=240)
totalTimeToReturnResult = time.time() - t0
PerformanceTestReporter.recordTest(testName, totalTimeToReturnResult, None)
finally:
simulation.teardown()
