[
https://issues.apache.org/jira/browse/HDFS-3148?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13243592#comment-13243592
]
Todd Lipcon commented on HDFS-3148:
-----------------------------------
- I think it makes more sense to make {{getLocalInterfaceAddrs}} static, and
take {{localInterfaces}} as a parameter.
----
{code}
+ public static final String DFS_CLIENT_LOCAL_INTERFACES =
"dfs.client.local.interfaces";
{code}
Move this higher in the file, near the other DFS_CLIENT configs
----
{code}
+ final int idx = r.nextInt(localInterfaceAddrs.length);
+ final SocketAddress addr = localInterfaceAddrs[idx];
+ if (LOG.isDebugEnabled()) {
+ LOG.debug("Using local interface " + localInterfaces[idx] + " " + addr);
{code}
This doesn't seem right, since {{localInterfaces}} and {{localInterfaceAddrs}}
may have different lengths -- a given configured local interface could have
multiple addrs in the {{localInterfaceAddrs}} list.
This brings up another question: if a NIC has multiple IPs, should it be
weighted in the load balancing based on the number of IPs assigned? That
doesn't seem right.
Maybe the right solution to both of these issues is to actually require that
the list of addresses decided upon has at most one IP corresponding to each
device?
Another possibility is that you could change the member variable to a
MultiMap<String, SocketAddress> -- first randomly choose a key from the map,
and then randomly choose among that key's values. My hunch is this would give
the right behavior most of the time.
----
{code}
+ <description>A comma separate list of network interface names to use
+ for data transfer between the client and datanodes. When creating
{code}
typo: comma separate*d* list
> The client should be able to use multiple local interfaces for data transfer
> ----------------------------------------------------------------------------
>
> Key: HDFS-3148
> URL: https://issues.apache.org/jira/browse/HDFS-3148
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Components: hdfs client
> Reporter: Eli Collins
> Assignee: Eli Collins
> Attachments: hdfs-3148-b1.txt, hdfs-3148.txt
>
>
> HDFS-3147 covers using multiple interfaces on the server (Datanode) side.
> Clients should also be able to utilize multiple *local* interfaces for
> outbound connections instead of always using the interface for the local
> hostname. This can be accomplished with a new configuration parameter
> ({{dfs.client.local.interfaces}}) that accepts a list of interfaces the
> client should use. Acceptable configuration values are the same as the
> {{dfs.datanode.available.interfaces}} parameter. The client binds its socket
> to a specific interface, which enables outbound traffic to use that
> interface. Binding the client socket to a specific address is not sufficient
> to ensure egress traffic uses that interface. Eg if multiple interfaces are
> on the same subnet the host requires IP rules that use the source address
> (which bind sets) to select the destination interface. The SO_BINDTODEVICE
> socket option could be used to select a specific interface for the connection
> instead, however it requires JNI (is not in Java's SocketOptions) and root
> access, which we don't want to require clients have.
> Like HDFS-3147, the client can use multiple local interfaces for data
> transfer. Since the client already cache their connections to DNs choosing a
> local interface at random seems like a good policy. Users can also pin a
> specific client to a specific interface by specifying just that interface in
> dfs.client.local.interfaces.
> This change was discussed in HADOOP-6210 a while back, and is actually
> useful/independent of the other HDFS-3140 changes.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
