Computing Block's range will miss some cell during blocking.
------------------------------------------------------------
Key: HAMA-130
URL: https://issues.apache.org/jira/browse/HAMA-130
Project: Hama
Issue Type: Bug
Components: implementation
Affects Versions: 0.1.0
Reporter: Samuel Guo
Assignee: Samuel Guo
Priority: Critical
Fix For: 0.1.0
As the code below, in *BlockingMapper*, we compute the block's col range in a
loop.
if the block's size can integer-divide the matrix' column size, it is right.
but if the block's size can not integer-divide the matrix' column size, some
cells will be missed.
for (int i = 0; i < mBlockNum; i++) {
startColumn = i * mBlockColSize;
endColumn = startColumn + mBlockColSize - 1;
output.collect(new BlockID(blkRow, i), new VectorWritable(key.get(),
dv.subVector(startColumn, endColumn)));
}
for examples:
if the block num is 3, the matrix is 100 * 100. mBlockSize = 100 / 3 = 33.
then <0 ~ 32>, <33 ~ 65>, <66 ~ 98> will be counted. At the same time we will
lose column 99.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
