pvary commented on a change in pull request #690: HIVE-21910: Multiple target
location generation in HostAffinitySplitLocationProvider
URL: https://github.com/apache/hive/pull/690#discussion_r298573326
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/exec/tez/HostAffinitySplitLocationProvider.java
##########
@@ -72,11 +78,17 @@ public HostAffinitySplitLocationProvider(List<String>
knownLocations) {
FileSplit fsplit = (FileSplit) split;
String splitDesc = "Split at " + fsplit.getPath() + " with offset= " +
fsplit.getStart()
+ ", length=" + fsplit.getLength();
- List<String> preferredLocations = preferLocations(fsplit);
- String location =
- preferredLocations.get(determineLocation(preferredLocations,
fsplit.getPath().toString(),
- fsplit.getStart(), splitDesc));
- return (location != null) ? new String[] { location } : null;
+ List<String> preferredLocations = new ArrayList<>(preferLocations(fsplit));
+ List<String> finalLocations = new ArrayList<>(numberOfLocations);
+ // Generate new preferred locations until we need more, or we do not have
any preferred
+ // location left
+ while (finalLocations.size() < numberOfLocations &&
preferredLocations.size() > 0) {
+ String nextLocation =
preferredLocations.get(determineLocation(preferredLocations,
+ fsplit.getPath().toString(), fsplit.getStart(), splitDesc));
+ finalLocations.add(nextLocation);
+ preferredLocations.remove(nextLocation);
Review comment:
I did some measurements for the split generation with this code:
`
@Test (timeout = 5000000)
public void testOrcSplitsBasic() throws IOException {
HostAffinitySplitLocationProvider locationProvider = new
HostAffinitySplitLocationProvider(executorLocations, true, 1);
InputSplit os1 = createMockFileSplit(true, "path1", 0, 1000, new String[]
{locations.get(0), locations.get(1), locations.get(2), locations.get(3)});
long start = System.nanoTime();
for(int i=0;i<100000;i++) {
locationProvider.getLocations(os1);
}
LOG.error("TIME: " + (System.nanoTime()-start)/1000000);
}
`
I got the following results:
Original code (~6100ms for 100k requests):
- 5859
- 6511
- 6813
- 5721
- 5663
New code with 1 location (~5823ms for 100k requests):
- 5877
- 5621
- 5613
- 5883
- 6120
New code with 2 locations (~6579ms for 100k request):
- 6433
- 6825
- 6574
- 6444
- 6621
I do not see why the new code should be faster, so this means probably high
variation for the data. Generating 2 locations instead of 1 seems like a 10%
overhead. Since this is 0.006ms per request this seems reasonable for me.
What is your opinion?
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
