1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
|
package spark.scheduler
import java.util.Properties
import spark.scheduler.cluster.TaskInfo
import spark.util.Distribution
import spark.{Logging, SparkContext, TaskEndReason, Utils}
import spark.executor.TaskMetrics
sealed trait SparkListenerEvents
case class SparkListenerStageSubmitted(stage: Stage, taskSize: Int) extends SparkListenerEvents
case class StageCompleted(val stageInfo: StageInfo) extends SparkListenerEvents
case class SparkListenerTaskEnd(task: Task[_], reason: TaskEndReason, taskInfo: TaskInfo,
taskMetrics: TaskMetrics) extends SparkListenerEvents
case class SparkListenerJobStart(job: ActiveJob, properties: Properties = null)
extends SparkListenerEvents
case class SparkListenerJobEnd(job: ActiveJob, jobResult: JobResult)
extends SparkListenerEvents
trait SparkListener {
/**
* Called when a stage is completed, with information on the completed stage
*/
def onStageCompleted(stageCompleted: StageCompleted) { }
/**
* Called when a stage is submitted
*/
def onStageSubmitted(stageSubmitted: SparkListenerStageSubmitted) { }
/**
* Called when a task ends
*/
def onTaskEnd(taskEnd: SparkListenerTaskEnd) { }
/**
* Called when a job starts
*/
def onJobStart(jobStart: SparkListenerJobStart) { }
/**
* Called when a job ends
*/
def onJobEnd(jobEnd: SparkListenerJobEnd) { }
}
/**
* Simple SparkListener that logs a few summary statistics when each stage completes
*/
class StatsReportListener extends SparkListener with Logging {
override def onStageCompleted(stageCompleted: StageCompleted) {
import spark.scheduler.StatsReportListener._
implicit val sc = stageCompleted
this.logInfo("Finished stage: " + stageCompleted.stageInfo)
showMillisDistribution("task runtime:", (info, _) => Some(info.duration))
//shuffle write
showBytesDistribution("shuffle bytes written:",(_,metric) => metric.shuffleWriteMetrics.map{_.shuffleBytesWritten})
//fetch & io
showMillisDistribution("fetch wait time:",(_, metric) => metric.shuffleReadMetrics.map{_.fetchWaitTime})
showBytesDistribution("remote bytes read:", (_, metric) => metric.shuffleReadMetrics.map{_.remoteBytesRead})
showBytesDistribution("task result size:", (_, metric) => Some(metric.resultSize))
//runtime breakdown
val runtimePcts = stageCompleted.stageInfo.taskInfos.map{
case (info, metrics) => RuntimePercentage(info.duration, metrics)
}
showDistribution("executor (non-fetch) time pct: ", Distribution(runtimePcts.map{_.executorPct * 100}), "%2.0f %%")
showDistribution("fetch wait time pct: ", Distribution(runtimePcts.flatMap{_.fetchPct.map{_ * 100}}), "%2.0f %%")
showDistribution("other time pct: ", Distribution(runtimePcts.map{_.other * 100}), "%2.0f %%")
}
}
object StatsReportListener extends Logging {
//for profiling, the extremes are more interesting
val percentiles = Array[Int](0,5,10,25,50,75,90,95,100)
val probabilities = percentiles.map{_ / 100.0}
val percentilesHeader = "\t" + percentiles.mkString("%\t") + "%"
def extractDoubleDistribution(stage:StageCompleted, getMetric: (TaskInfo,TaskMetrics) => Option[Double]): Option[Distribution] = {
Distribution(stage.stageInfo.taskInfos.flatMap{
case ((info,metric)) => getMetric(info, metric)})
}
//is there some way to setup the types that I can get rid of this completely?
def extractLongDistribution(stage:StageCompleted, getMetric: (TaskInfo,TaskMetrics) => Option[Long]): Option[Distribution] = {
extractDoubleDistribution(stage, (info, metric) => getMetric(info,metric).map{_.toDouble})
}
def showDistribution(heading: String, d: Distribution, formatNumber: Double => String) {
val stats = d.statCounter
logInfo(heading + stats)
val quantiles = d.getQuantiles(probabilities).map{formatNumber}
logInfo(percentilesHeader)
logInfo("\t" + quantiles.mkString("\t"))
}
def showDistribution(heading: String, dOpt: Option[Distribution], formatNumber: Double => String) {
dOpt.foreach { d => showDistribution(heading, d, formatNumber)}
}
def showDistribution(heading: String, dOpt: Option[Distribution], format:String) {
def f(d:Double) = format.format(d)
showDistribution(heading, dOpt, f _)
}
def showDistribution(heading:String, format: String, getMetric: (TaskInfo,TaskMetrics) => Option[Double])
(implicit stage: StageCompleted) {
showDistribution(heading, extractDoubleDistribution(stage, getMetric), format)
}
def showBytesDistribution(heading:String, getMetric: (TaskInfo,TaskMetrics) => Option[Long])
(implicit stage: StageCompleted) {
showBytesDistribution(heading, extractLongDistribution(stage, getMetric))
}
def showBytesDistribution(heading: String, dOpt: Option[Distribution]) {
dOpt.foreach{dist => showBytesDistribution(heading, dist)}
}
def showBytesDistribution(heading: String, dist: Distribution) {
showDistribution(heading, dist, (d => Utils.memoryBytesToString(d.toLong)): Double => String)
}
def showMillisDistribution(heading: String, dOpt: Option[Distribution]) {
showDistribution(heading, dOpt, (d => StatsReportListener.millisToString(d.toLong)): Double => String)
}
def showMillisDistribution(heading: String, getMetric: (TaskInfo, TaskMetrics) => Option[Long])
(implicit stage: StageCompleted) {
showMillisDistribution(heading, extractLongDistribution(stage, getMetric))
}
val seconds = 1000L
val minutes = seconds * 60
val hours = minutes * 60
/**
* reformat a time interval in milliseconds to a prettier format for output
*/
def millisToString(ms: Long) = {
val (size, units) =
if (ms > hours) {
(ms.toDouble / hours, "hours")
} else if (ms > minutes) {
(ms.toDouble / minutes, "min")
} else if (ms > seconds) {
(ms.toDouble / seconds, "s")
} else {
(ms.toDouble, "ms")
}
"%.1f %s".format(size, units)
}
}
case class RuntimePercentage(executorPct: Double, fetchPct: Option[Double], other: Double)
object RuntimePercentage {
def apply(totalTime: Long, metrics: TaskMetrics): RuntimePercentage = {
val denom = totalTime.toDouble
val fetchTime = metrics.shuffleReadMetrics.map{_.fetchWaitTime}
val fetch = fetchTime.map{_ / denom}
val exec = (metrics.executorRunTime - fetchTime.getOrElse(0l)) / denom
val other = 1.0 - (exec + fetch.getOrElse(0d))
RuntimePercentage(exec, fetch, other)
}
}
|
