问题描述
我正在做一些流处理框架之间的基准比较,
我在这方面选择了 WordCount 这样的“Hello world”任务(有一些曲折),到目前为止测试了 Flink 和 Hazelcast Jet,结果是 Flink 需要 80+s 才能完成,而 Jet 只需要 30+s
我知道 Flink 很受欢迎,我在这里做错了什么?真的很好奇这个
我的示例代码在这里
https://github.com/ChinW/stream-processing-compare
以下是详细信息(规范、管道、日志)
经过测试的 WordCount 管道
Source (read from file,5MB)
-> Process: Split line into words (Here here is a bomb,every word emit 1000 times)
-> Group/Count
-> Sink (do nothing)
我的本地结果
- MacBook Pro(13 英寸,2020 年,四个雷雳 3 端口)
- 2 GHz 四核 Intel Core i5(8 个逻辑处理器)
- 16 GB 3733 MHz LPDDR4X
- JDK 11
管道:
digraph DAG {
"items" [localParallelism=1];
"fused(flat-map,filter)" [localParallelism=8];
"group-and-aggregate-prepare" [localParallelism=8];
"group-and-aggregate" [localParallelism=8];
"do-nothing-sink" [localParallelism=1];
"items" -> "fused(flat-map,filter)" [queueSize=1024];
"fused(flat-map,filter)" -> "group-and-aggregate-prepare" [label="partitioned",queueSize=1024];
subgraph cluster_0 {
"group-and-aggregate-prepare" -> "group-and-aggregate" [label="distributed-partitioned",queueSize=1024];
}
"group-and-aggregate" -> "do-nothing-sink" [queueSize=1024];
}
日志:
Start time: 2021-04-18T13:52:52.106
Duration: 00:00:36.459
Jet: finish in 36.45935081 seconds.
Start time: 2021-04-19T16:51:53.806
Duration: 00:00:30.143
Jet: finish in 30.625740453 seconds.
Start time: 2021-04-19T16:52:48.906
Duration: 00:00:37.207
Jet: finish in 37.862554137 seconds.
Scala 2.11 的 Flink 1.12.2
flink-config.yaml
配置:
jobmanager.rpc.address: localhost
jobmanager.rpc.port: 6123
jobmanager.memory.process.size: 2096m
taskmanager.memory.process.size: 12288m
taskmanager.numberOfTaskSlots: 8
parallelism.default: 8
管道:
{
"nodes" : [ {
"id" : 1,"type" : "Source: Custom Source","pact" : "Data Source","contents" : "Source: Custom Source","parallelism" : 1
},{
"id" : 2,"type" : "Flat Map","pact" : "Operator","contents" : "Flat Map","parallelism" : 8,"predecessors" : [ {
"id" : 1,"ship_strategy" : "REBALANCE","side" : "second"
} ]
},{
"id" : 4,"type" : "Keyed Aggregation","contents" : "Keyed Aggregation","predecessors" : [ {
"id" : 2,"ship_strategy" : "HASH",{
"id" : 5,"type" : "Sink: Unnamed","pact" : "Data Sink","contents" : "Sink: Unnamed","predecessors" : [ {
"id" : 4,"ship_strategy" : "FORWARD","side" : "second"
} ]
} ]
}
日志:
❯ flink run -c chiw.spc.flink.FlinkWordCountKt stream-processing-compare-1.0-SNAPSHOT.jar
Job has been submitted with JobID 163ce849a663e45f3c3028a98f260e7c
Program execution finished
Job with JobID 163ce849a663e45f3c3028a98f260e7c has finished.
Job Runtime: 88614 ms
❯ flink run -c chiw.spc.flink.FlinkWordCountKt stream-processing-compare-1.0-SNAPSHOT.jar
Job has been submitted with JobID fcf12488204969299e4e5d7f23f4ea6e
Program execution finished
Job with JobID fcf12488204969299e4e5d7f23f4ea6e has finished.
Job Runtime: 90165 ms
❯ flink run -c chiw.spc.flink.FlinkWordCountKt stream-processing-compare-1.0-SNAPSHOT.jar
Job has been submitted with JobID 37e349e4fad90cd7405546d30239afa4
Program execution finished
Job with JobID 37e349e4fad90cd7405546d30239afa4 has finished.
Job Runtime: 78908 ms
非常感谢您的帮助!
解决方法
我认为您没有做错任何事情,我们的测试表明 Jet 比 Spark 和 Flink 快得多,而字数统计是我们用来衡量这一点的示例之一。
,鉴于您的炸弹会产生大量的小物品(而不是少量的大物品),我最好的猜测是为什么 Jet 在这里可能具有优势是它的单一生产者- 单消费者 (SPSC) 队列与类似协程的并发性相结合。
您有 8 个平面映射阶段和 8 个聚合阶段。 Jet 将在总共 8 个线程上执行此操作(假设您有 8 个 availableProcessors
),因此几乎不会在操作系统级别进行线程调度。数据将以大块的形式在线程之间移动:flatMap
将一次排队 1024 个,然后每个聚合器将拉出所有指定给它的项目。 SPSC 队列上的通信不受其他线程的任何干扰:每个聚合处理器有 8 个输入队列,一个专用于每个平面映射器。
在 Flink 中,每个阶段将启动另外 8 个线程,我还注意到接收器的并行度为 8,因此这是 24 个线程,另一个用于源。操作系统必须将它们安排在 8 个物理内核上。通信将通过多生产者单消费者 (MPSC) 队列进行,这意味着所有平面映射器线程必须协调,以便一次只有一个线程将一个项目排队到任何给定的聚合器,并且争用导致热 CAS 循环所有线程。
要确认这种怀疑,请尝试收集一些分析数据。如果上面的故事是正确的,你应该看到 Flink 花费了大量的 CPU 时间来排队数据。