[
https://issues.apache.org/jira/browse/HADOOP-11552?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14313004#comment-14313004
]
Siddharth Seth commented on HADOOP-11552:
-----------------------------------------
Thanks for the feedback [~cmccabe].
bq. Are you proposing to keep the TCP session open, but reuse the handler
thread for something else, while the RPC is progressing? It seems like this
approach would alleviate the handler thread bottleneck, but potentially run
into another bottleneck by keeping open a lot of open TCP sockets (file
descriptors). Given that we've had problems with hitting max file descriptors
in the past, couldn't this be problematic?
Yes, the intent is to keep the TPC session open and re-use the handlers
(decided by individual protocol). That's a great point about FDs remaining
open. Depending on where this mechanism is used, I believe this could be
controlled. Examples.
- AM-RM communication - one connection per AM running on a cluster. This is
typically limited to < #nodes on the cluster.
- NM-RM communication - if at all it makes sense to add it here. Would be
limited to #nodes on the cluster. (Similarly for DN-NN communication)
- Task-AM communication (MR as an example) - this definitely has the potential
to blow up, since the # of containers on a cluster can be large, and multiple
AMs could be running on the same node.
The approach doesn't force all protocols / all methods in a protocol to make
use of this, and only selective ones could be setup to use this.
It does, however, open up the option to design protocols without the #handler
limit, where applicable - and individual components would need to exercise this
wisely.
bq. Heartbeats are fundamental to how YARN works. If we move to an offer-based
system like Mesos, it seems like the way to do it would be to have the resource
manager make outgoing connections to the executors, rather than keeping TCP
sessions open a long time.
I don't think we need to make a drastic change to protocol like moving to the
offer model used in Mesos.
The request-response model used by YARN can be made faster though. Rightnow,
heartbeats are serving multiple purposes. Taking the AM-RM example, the
allocate request serves as 1) a keep-alive - this isn't required every second,
and 2) request-response for container requests - the default 1 second heartbeat
is to avoid delaying getting allocations for a previous request.
Full two-way communication would be ideal - where the RM opens connections when
required, but is a much larger change - and I doubt it'll be possible to make
this in a compatible manner, if we want to go down that route.
For individual components to use this - we could even consider adding new
methods to make use of this mechanism, and leave the old ones intact - with the
option to use either mechanism.
> Allow handoff on the server side for RPC requests
> -------------------------------------------------
>
> Key: HADOOP-11552
> URL: https://issues.apache.org/jira/browse/HADOOP-11552
> Project: Hadoop Common
> Issue Type: Improvement
> Components: ipc
> Reporter: Siddharth Seth
> Assignee: Siddharth Seth
> Attachments: HADOOP-11552.1.wip.txt
>
>
> An RPC server handler thread is tied up for each incoming RPC request. This
> isn't ideal, since this essentially implies that RPC operations should be
> short lived, and most operations which could take time end up falling back to
> a polling mechanism.
> Some use cases where this is useful.
> - YARN submitApplication - which currently submits, followed by a poll to
> check if the application is accepted while the submit operation is written
> out to storage. This can be collapsed into a single call.
> - YARN allocate - requests and allocations use the same protocol. New
> allocations are received via polling.
> The allocate protocol could be split into a request/heartbeat along with a
> 'awaitResponse'. The request/heartbeat is sent only when there's a request or
> on a much longer heartbeat interval. awaitResponse is always left active with
> the RM - and returns the moment something is available.
> MapReduce/Tez task to AM communication is another example of this pattern.
> The same pattern of splitting calls can be used for other protocols as well.
> This should serve to improve latency, as well as reduce network traffic since
> the keep-alive heartbeat can be sent less frequently.
> I believe there's some cases in HDFS as well, where the DN gets told to
> perform some operations when they heartbeat into the NN.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
