You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@spark.apache.org by Larry Xiao <xi...@sjtu.edu.cn> on 2014/09/28 05:25:39 UTC

PageRank execution imbalance, might hurt performance by 6x

Hi all!

I'm running PageRank on GraphX, and I find on some tasks on one machine 
can spend 5~6 times more time than on others, others are perfectly 
balance (around 1 second to finish).
And since time for a stage (iteration) is determined by the slowest 
task, the performance is undesirable.

I don't know if there's any internals that might make execution 
unstable? Like scheduling, garbage collection …

A stage for mapPartitions at GraphImpl.scala:409
> in mapReduceTriplets
> 408     // Map and combine.
> 409     val preAgg = view.edges.partitionsRDD.mapPartitions(_.flatMap {
> 410       case (pid, edgePartition) =>
> 411         // Choose scan method
looks like this:


        Tasks

Index 	ID 	Attempt 	Status 	Locality Level 	Executor 	Launch Time 
Duration ▴ 	GC Time 	Accumulators 	Input 	Shuffle Read 	Write Time 
Shuffle Write 	Errors
21 	787 	0 	SUCCESS 	PROCESS_LOCAL 	brick0 	2014/09/28 03:04:42 	7 s 	
	
	333.3 MB (memory) 	4.9 MB 	1 ms 	652.3 KB 	
0 	768 	0 	SUCCESS 	PROCESS_LOCAL 	brick2 	2014/09/28 03:04:42 	7 s 	
	
	531.5 MB (memory) 	8.0 MB 	2 ms 	1321.5 KB 	
9 	775 	0 	SUCCESS 	PROCESS_LOCAL 	brick0 	2014/09/28 03:04:42 	6 s 	
	
	270.4 MB (memory) 	4.1 MB 	1 ms 	659.3 KB 	
15 	781 	0 	SUCCESS 	PROCESS_LOCAL 	brick0 	2014/09/28 03:04:42 	6 s 	
	
	272.7 MB (memory) 	4.3 MB 	1 ms 	658.9 KB 	
3 	769 	0 	SUCCESS 	PROCESS_LOCAL 	brick0 	2014/09/28 03:04:42 	6 s 	
	
	285.5 MB (memory) 	4.4 MB 	1 ms 	658.5 KB 	
6 	774 	0 	SUCCESS 	PROCESS_LOCAL 	brick2 	2014/09/28 03:04:42 	6 s 	
	
	346.8 MB (memory) 	4.6 MB 	1 ms 	657.0 KB 	
12 	780 	0 	SUCCESS 	PROCESS_LOCAL 	brick2 	2014/09/28 03:04:42 	6 s 	
	
	313.2 MB (memory) 	4.4 MB 	1 ms 	645.5 KB 	
18 	786 	0 	SUCCESS 	PROCESS_LOCAL 	brick2 	2014/09/28 03:04:42 	6 s 	
	
	281.7 MB (memory) 	4.2 MB 	1 ms 	660.1 KB 	
1 	771 	0 	SUCCESS 	PROCESS_LOCAL 	brick3 	2014/09/28 03:04:42 	2 s 	
	
	339.1 MB (memory) 	5.1 MB 	1 ms 	657.4 KB 	
7 	777 	0 	SUCCESS 	PROCESS_LOCAL 	brick3 	2014/09/28 03:04:42 	2 s 	
	
	322.8 MB (memory) 	4.9 MB 	1 ms 	654.5 KB 	
13 	783 	0 	SUCCESS 	PROCESS_LOCAL 	brick3 	2014/09/28 03:04:42 	2 s 	
	
	279.8 MB (memory) 	4.6 MB 	1 ms 	655.4 KB 	
19 	789 	0 	SUCCESS 	PROCESS_LOCAL 	brick3 	2014/09/28 03:04:42 	2 s 	
	
	268.4 MB (memory) 	4.4 MB 	1 ms 	658.5 KB 	
16 	784 	0 	SUCCESS 	PROCESS_LOCAL 	brick4 	2014/09/28 03:04:42 	1 s 	
	
	339.1 MB (memory) 	5.1 MB 	1 ms 	660.1 KB 	
11 	776 	0 	SUCCESS 	PROCESS_LOCAL 	brick1 	2014/09/28 03:04:42 	1 s 	
	
	341.0 MB (memory) 	5.3 MB 	1 ms 	655.4 KB 	
2 	773 	0 	SUCCESS 	PROCESS_LOCAL 	brick5 	2014/09/28 03:04:42 	1 s 	
	
	320.9 MB (memory) 	4.9 MB 	1 ms 	655.3 KB 	
22 	790 	0 	SUCCESS 	PROCESS_LOCAL 	brick4 	2014/09/28 03:04:42 	1 s 	
	
	301.7 MB (memory) 	4.9 MB 	1 ms 	659.5 KB 	
17 	782 	0 	SUCCESS 	PROCESS_LOCAL 	brick1 	2014/09/28 03:04:42 	1 s 	
	
	317.1 MB (memory) 	5.2 MB 	1 ms 	653.7 KB 	
23 	788 	0 	SUCCESS 	PROCESS_LOCAL 	brick1 	2014/09/28 03:04:42 	1 s 	
	
	268.3 MB (memory) 	4.9 MB 	1 ms 	664.3 KB 	
8 	779 	0 	SUCCESS 	PROCESS_LOCAL 	brick5 	2014/09/28 03:04:42 	1 s 	
	
	291.3 MB (memory) 	4.6 MB 	1 ms 	660.6 KB 	
20 	791 	0 	SUCCESS 	PROCESS_LOCAL 	brick5 	2014/09/28 03:04:42 	1 s 	
	
	272.7 MB (memory) 	4.5 MB 	1 ms 	661.7 KB 	
10 	778 	0 	SUCCESS 	PROCESS_LOCAL 	brick4 	2014/09/28 03:04:42 	1 s 	
	
	276.5 MB (memory) 	4.4 MB 	1 ms 	656.4 KB 	
4 	772 	0 	SUCCESS 	PROCESS_LOCAL 	brick4 	2014/09/28 03:04:42 	1 s 	
	
	260.8 MB (memory) 	4.4 MB 	1 ms 	661.7 KB 	
14 	785 	0 	SUCCESS 	PROCESS_LOCAL 	brick5 	2014/09/28 03:04:42 	1 s 	
	
	262.7 MB (memory) 	4.3 MB 	1 ms 	651.7 KB 	
5 	770 	0 	SUCCESS 	PROCESS_LOCAL 	brick1 	2014/09/28 03:04:42 	1 s 	
	
	276.5 MB (memory) 	4.7 MB 	1 ms 	655.1 KB 	


Thanks!
Larry