[HDFS-17085] Erasure coding: readTo is computed large than actually needed during pread

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Not A Problem
• Affects Version/s: 3.4.0
• Component/s: erasure-coding
• Target Version/s: 3.4.0
• Assignee: farmmamba
• Reporter: farmmamba
• Created: 13/Jul/23 09:23
• Updated: 13/Jul/23 13:15
• Resolved: 13/Jul/23 13:15

Description

In HDFS-16520，it improved EC pread by introducing a readTo field.

But, the way it was calculated seems still have some room for improvement.

Now, it was calculated by below code:

for (AlignedStripe stripe : stripes) {
  readTo = Math.max(readTo, stripe.getOffsetInBlock() + stripe.getSpanInBlock());
} 

But in the followed code, for every AlignedStripe object, it uses max readTo to construct StripeReader. I think there still exists waste of resource.

for (AlignedStripe stripe : stripes) {
  // Parse group to get chosen DN location
  StripeReader preader = new PositionStripeReader(stripe, ecPolicy, blks,
      preaderInfos, corruptedBlocks, decoder, this);
  preader.setReadTo(readTo);
  try {
    preader.readStripe();
  } finally {
    preader.close();
  }
}