[
https://issues.apache.org/jira/browse/HDFS-5746?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13866564#comment-13866564
]
Colin Patrick McCabe commented on HDFS-5746:
--------------------------------------------
See here for some notes about the strategy for 5182:
https://issues.apache.org/jira/browse/HDFS-5182?focusedCommentId=13866545&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-13866545
* Add SharedFileDescriptorFactory. This is a class which can produce anonymous
shared memory segments suitable for passing from the DataNode to the DFSClient
via file descriptor passing. It would have been nice to do this without JNI,
but unfortunately we don't have {{open(O_EXCL)}} support in JDK6, which we're
still supporting. There is {{NativeIO#open}}, but it doesn't allow me to
cleanly separate {{EEXIST}} errors from other errors (the error gets turned
into a textual exception which I don't want to parse). Also there may be some
symlink issues with the JDK6 java APIs for listing files in a directory, etc.
Overall, the native implementation was just easier. This is something we
should probably revisit with JDK7, of course.
* Add {{NativeIO#mmap}} and {{NativeIO#munmap}}. Although it would be nicer to
use {{FileChannel#map}}, there is no public interface to get access to the
virtual memory address of a {{MappedByteBuffer}}, and I needed that. Luckily,
the amount of code needed to just call mmap is really small.
* I didn't want to duplicate the code used to stuff a reference count + closed
bit into {{DomainSocket#refCount}}, so I factored it out into
{{CloseableReferenceCount}}. This class is now used in both DomainSocket and
{{ShortCircuitSharedMemorySegment}}.
* {{DomainSocketWatcher}} is a thread which calls poll() in a loop. This will
be used to detect when a DFSClient has closed and its shared memory segments
can be closed, by detecting when their associated DomainSockets are closed. I
used poll() here rather than select() since select() has some limitations with
high-numbered file descriptors on some platforms. Also, poll's interface is a
bit simpler. It would have been nice to use Java NIO for this, but
{{DomainSocket}} is not integrated with NIO. poll() doesn't scale as well as
epoll() and other platform-specific functions, but we don't need it to, since
this just for handling clients closing, which should be a relatively infrequent
event. We're not using this for handling every packet sent through a webserver
or something.
* {{ShortCircuitSharedMemorySegment}} is entirely in Java, using
{{sun.misc.Unsafe}} for the anchor / unanchor / etc. operations. This is
preferrable to using JNI for this, since {{Unsafe#compareAndSwap}} will be
inlined by the JVM. (Thanks to [~tlipcon] for pointing out the existence of
these functions).
> add ShortCircuitSharedMemorySegment
> -----------------------------------
>
> Key: HDFS-5746
> URL: https://issues.apache.org/jira/browse/HDFS-5746
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Components: datanode, hdfs-client
> Reporter: Colin Patrick McCabe
> Assignee: Colin Patrick McCabe
> Fix For: 3.0.0
>
>
> Add ShortCircuitSharedMemorySegment, which will be used to communicate
> information between the datanode and the client about whether a replica is
> mlocked.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
