Error: GC overhead limit exceeded. Child Error org.apache.hadoop.mapred.TaskRunner.run

Hello everyone,

I'm trying to run pregel style implementation of strongly connected components algorithm on top of giraph. A user defined writable (LongPairWritable) used as the type of the messages between the vertices. The algorithm works well in single node. However, whenever I tested using large graph (around 1 million vertices) in cluster, a GC overhead limit exceeded caused by out of memory exception. The error comes after attempt=0 super-step=1 with 2 minutes pending. I'm wondering how to solve the problem and what causes it?

Is it possible that LongPairWritable cause the problem?
I tried a simple test:
Succeeded job:
Comput(Iterator<LongPairWritable> msgIterator){
If(getSuperstep() >=20){
voteToHalt();
} else {
//Do nothing
}
}

Failed job:
Comput(Iterator<LongPairWritable> msgIterator){
If(getSuperstep() >=20){
voteToHalt();
} else {
sendToAll(new LongPairWritable(new LongWritable(0),new LongWritable(0));
}
}

Thanks in advance.

Amani