[
https://issues.apache.org/jira/browse/HDFS-12051?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16353441#comment-16353441
]
Misha Dmitriev commented on HDFS-12051:
---------------------------------------
I tested my change in a relatively small cluster that simulates some customer's
workload at a smaller scale. It contains about 15M HDFS files, NN heap
oscillates between about 2.1 and 2.9GiB. Before the change, FSImage load time
is 48 sec, and there was the following overhead due to duplicate byte[] arrays:
|| *Overhead* || *# objects* || *Unique objects* || *Class name* ||
| 336,794K (14.8%)| 11,309,068| 3,132,447|byte[]|
I first applied the current patch, where the NameCache size is set as 1/512th
of the total heap size. It resulted in a cache with 1M entries, taking 4MiB of
memory. It turns out that when the cache size is considerably smaller than the
number of unique arrays (3M in the above case), it leaves behind some duplicate
objects. We saved ~100MiB of memory:
|| *Overhead* || *# objects* || *Unique objects* || *Class name* ||
| 232,435K (10.7%)| 8,599,197| 3,132,367|byte[]|
Next, I changed my patch so that by default NameCache is set as 1/400 of the
total heap size. This resulted in a cache with 2M entries, and it eliminated
almost all duplicate objects and saved, compared to the original NameNode, more
than 260MiB of memory. The FSImage load time was 50 sec.
|| *Overhead* || *# objects* || *Unique objects* || *Class name* ||
| 76,209K (3.8%)| 4,866,933| 3,132,450|byte[]|
> Reimplement NameCache in NameNode: Intern duplicate byte[] arrays (mainly
> those denoting file/directory names) to save memory
> -----------------------------------------------------------------------------------------------------------------------------
>
> Key: HDFS-12051
> URL: https://issues.apache.org/jira/browse/HDFS-12051
> Project: Hadoop HDFS
> Issue Type: Improvement
> Reporter: Misha Dmitriev
> Assignee: Misha Dmitriev
> Priority: Major
> Attachments: HDFS-12051-NameCache-Rewrite.pdf, HDFS-12051.01.patch,
> HDFS-12051.02.patch, HDFS-12051.03.patch, HDFS-12051.04.patch,
> HDFS-12051.05.patch, HDFS-12051.06.patch, HDFS-12051.07.patch,
> HDFS-12051.08.patch, HDFS-12051.09.patch
>
>
> When snapshot diff operation is performed in a NameNode that manages several
> million HDFS files/directories, NN needs a lot of memory. Analyzing one heap
> dump with jxray (www.jxray.com), we observed that duplicate byte[] arrays
> result in 6.5% memory overhead, and most of these arrays are referenced by
> {{org.apache.hadoop.hdfs.server.namenode.INodeFileAttributes$SnapshotCopy.name}}
> and {{org.apache.hadoop.hdfs.server.namenode.INodeFile.name}}:
> {code:java}
> 19. DUPLICATE PRIMITIVE ARRAYS
> Types of duplicate objects:
> Ovhd Num objs Num unique objs Class name
> 3,220,272K (6.5%) 104749528 25760871 byte[]
> ....
> 1,841,485K (3.7%), 53194037 dup arrays (13158094 unique)
> 3510556 of byte[17](112, 97, 114, 116, 45, 109, 45, 48, 48, 48, ...), 2228255
> of byte[8](48, 48, 48, 48, 48, 48, 95, 48), 357439 of byte[17](112, 97, 114,
> 116, 45, 109, 45, 48, 48, 48, ...), 237395 of byte[8](48, 48, 48, 48, 48, 49,
> 95, 48), 227853 of byte[17](112, 97, 114, 116, 45, 109, 45, 48, 48, 48, ...),
> 179193 of byte[17](112, 97, 114, 116, 45, 109, 45, 48, 48, 48, ...), 169487
> of byte[8](48, 48, 48, 48, 48, 50, 95, 48), 145055 of byte[17](112, 97, 114,
> 116, 45, 109, 45, 48, 48, 48, ...), 128134 of byte[8](48, 48, 48, 48, 48, 51,
> 95, 48), 108265 of byte[17](112, 97, 114, 116, 45, 109, 45, 48, 48, 48, ...)
> ... and 45902395 more arrays, of which 13158084 are unique
> <--
> org.apache.hadoop.hdfs.server.namenode.INodeFileAttributes$SnapshotCopy.name
> <-- org.apache.hadoop.hdfs.server.namenode.snapshot.FileDiff.snapshotINode
> <-- {j.u.ArrayList} <--
> org.apache.hadoop.hdfs.server.namenode.snapshot.FileDiffList.diffs <--
> org.apache.hadoop.hdfs.server.namenode.snapshot.FileWithSnapshotFeature.diffs
> <-- org.apache.hadoop.hdfs.server.namenode.INode$Feature[] <--
> org.apache.hadoop.hdfs.server.namenode.INodeFile.features <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlockInfo.bc <-- ... (1
> elements) ... <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlocksMap$1.entries <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlocksMap.blocks <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlockManager.blocksMap <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlockManager$BlockReportProcessingThread.this$0
> <-- j.l.Thread[] <-- j.l.ThreadGroup.threads <-- j.l.Thread.group <-- Java
> Static: org.apache.hadoop.fs.FileSystem$Statistics.STATS_DATA_CLEANER
> 409,830K (0.8%), 13482787 dup arrays (13260241 unique)
> 430 of byte[32](116, 97, 115, 107, 95, 49, 52, 57, 55, 48, ...), 353 of
> byte[32](116, 97, 115, 107, 95, 49, 52, 57, 55, 48, ...), 352 of
> byte[32](116, 97, 115, 107, 95, 49, 52, 57, 55, 48, ...), 350 of
> byte[32](116, 97, 115, 107, 95, 49, 52, 57, 55, 48, ...), 342 of
> byte[32](116, 97, 115, 107, 95, 49, 52, 57, 55, 48, ...), 341 of
> byte[32](116, 97, 115, 107, 95, 49, 52, 57, 55, 48, ...), 341 of
> byte[32](116, 97, 115, 107, 95, 49, 52, 57, 55, 48, ...), 340 of
> byte[32](116, 97, 115, 107, 95, 49, 52, 57, 55, 48, ...), 337 of
> byte[32](116, 97, 115, 107, 95, 49, 52, 57, 55, 48, ...), 334 of
> byte[32](116, 97, 115, 107, 95, 49, 52, 57, 55, 48, ...)
> ... and 13479257 more arrays, of which 13260231 are unique
> <-- org.apache.hadoop.hdfs.server.namenode.INodeFile.name <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlockInfo.bc <--
> org.apache.hadoop.util.LightWeightGSet$LinkedElement[] <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlocksMap$1.entries <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlocksMap.blocks <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlockManager.blocksMap <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlockManager$BlockReportProcessingThread.this$0
> <-- j.l.Thread[] <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlocksMap$1.entries <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlocksMap.blocks <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlockManager.blocksMap <--
> org.apache.hadoop.hdfs.server.blockmanagement.BlockManager$BlockReportProcessingThread.this$0
> <-- j.l.Thread[] <-- j.l.ThreadGroup.threads <-- j.l.Thread.group <-- Java
> Static: org.apache.hadoop.fs.FileSystem$Statistics.STATS_DATA_CLEANER
> ....
> {code}
> There are several other places in NameNode code which also produce duplicate
> {{byte[]}} arrays.
> To eliminate this duplication and reclaim memory, we will need to write a
> small class similar to StringInterner, but designed specifically for byte[]
> arrays.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
