[
https://issues.apache.org/jira/browse/HDFS-10419?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16238379#comment-16238379
]
Sanjay Radia commented on HDFS-10419:
-------------------------------------
HDFS-5389 describes one approach of building a NN that scales its namespace
better than the current NN.
It proposes caching only working set namespace in memory; also see [HUG -
Removing Namenode's
Limitation|https://www.slideshare.net/ydn/hadoop-meetup-hug-august-2013-removing-the-namenodes-memory-limitation].
Independent studies have also analysed LRU caching of HDFS Metadata
[Metadata Traces and Workload Models for Evaluating Big Storage
Systems|https://www.slideshare.net/ydn/hadoop-meetup-hug-august-2013-removing-the-namenodes-memory-limitation]
This approach works because in spite of having large amounts of data (say data
for the last five years) most of the data that is accessed is recent (say last
3-9 months); hence the working set can fit in memory.
> Building HDFS on top of Ozone's storage containers
> --------------------------------------------------
>
> Key: HDFS-10419
> URL: https://issues.apache.org/jira/browse/HDFS-10419
> Project: Hadoop HDFS
> Issue Type: New Feature
> Reporter: Jing Zhao
> Assignee: Jing Zhao
> Priority: Major
>
> In HDFS-7240, Ozone defines storage containers to store both the data and the
> metadata. The storage container layer provides an object storage interface
> and aims to manage data/metadata in a distributed manner. More details about
> storage containers can be found in the design doc in HDFS-7240.
> HDFS can adopt the storage containers to store and manage blocks. The general
> idea is:
> # Each block can be treated as an object and the block ID is the object's key.
> # Blocks will still be stored in DataNodes but as objects in storage
> containers.
> # The block management work can be separated out of the NameNode and will be
> handled by the storage container layer in a more distributed way. The
> NameNode will only manage the namespace (i.e., files and directories).
> # For each file, the NameNode only needs to record a list of block IDs which
> are used as keys to obtain real data from storage containers.
> # A new DFSClient implementation talks to both NameNode and the storage
> container layer to read/write.
> HDFS, especially the NameNode, can get much better scalability from this
> design. Currently the NameNode's heaviest workload comes from the block
> management, which includes maintaining the block-DataNode mapping, receiving
> full/incremental block reports, tracking block states (under/over/miss
> replicated), and joining every writing pipeline protocol to guarantee the
> data consistency. These work bring high memory footprint and make NameNode
> suffer from GC. HDFS-5477 already proposes to convert BlockManager as a
> service. If we can build HDFS on top of the storage container layer, we not
> only separate out the BlockManager from the NameNode, but also replace it
> with a new distributed management scheme.
> The storage container work is currently in progress in HDFS-7240, and the
> work proposed here is still in an experimental/exploring stage. We can do
> this experiment in a feature branch so that people with interests can be
> involved.
> A design doc will be uploaded later explaining more details.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
