[
https://issues.apache.org/jira/browse/HDFS-6607?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14048577#comment-14048577
]
Abdullah Alamoudi commented on HDFS-6607:
-----------------------------------------
I know I didn't give enough details. What I've done is the following: I built
B-Tree indexes on records stored in HDFS files where I am using the records
positions offsets to point to the records so I can perform seek and read in
order to get the records of interest. when I search my index and before I use
the list of offsets to access the records I sort them so I can access the
records sequentially when possible. After that, I wanted to test the
performance for the indexes and found out that they perform better when the
number of accessed records is 60M than when they are 3M which doesn't make
sense.
So I went and started looking at the seek implementation and came up with this
conjecture that when I have more records and I am accessing them sequentially,
the chances for the next record to be in the buffer is higher and so I am doing
less of (closing current reader, creating new sockets and looking for the block
location, etc) than when I am accessing less number of records.
As for the heavy stuff being in the read and not the seek, you are right but
what I meant is that they are related as they occur because I perform a seek
followed by a read :-). What I think should be done is that if the next desired
position is in the same currently being read block, a cheaper way to do this
using the same block reader should be implemented.
> Improve DFSInputStream forward seek performance
> -----------------------------------------------
>
> Key: HDFS-6607
> URL: https://issues.apache.org/jira/browse/HDFS-6607
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: hdfs-client, performance
> Affects Versions: 2.4.1
> Reporter: Abdullah Alamoudi
>
> When having a DFSInputStream open and seeking to a position that resides in
> the same block, if the target position is in the TCP buffer already, the seek
> is performed efficiently simply by eating up the intervening data. See line
> 1368 in the file:
> hadoop-common/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/DFSInputStream.java.
> However, if the position is in the same block but after the TCP buffer, the
> inputstream performs a set of actions including closing the current block
> reader, locating the block again, selecting a data node and creating a new
> block reader. During this, many objects are created and all of this is very
> inefficient for users with random access needs (e.g index access).
> I have conducted some experiments which showed that reading 3,000,000 records
> using seeks and reads is slower than reading 60,000,000 records using seeks
> and reads as well which shows the need to improve the seek implementation.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
