Hi, Owen:
Thanks for your detail answer. I like to use this as a chance to confirm what I
understand.
As you said, normally it is one executor per application per work. So in my
standalone one box environment, if I want to use all 48G memory allocated to
worker for my application, I should ask 48G memory for the executor in the
spark shell, right? Because 48G is too big for a JVM heap in normal case, I can
and should consider to start multi workers in one box, to lower the executor
memory, but still use all 48G memory.
In the spark document, about the "-- cores" parameter, the default is all
available cores, so it means using all available cores in all workers, even in
the cluster environment? If so, in default case, if one client submit a huge
job, it will use all the available cores from the cluster for all the tasks it
generates?
One thing is still not clear is in the given example I have, if 10 tasks (1 per
partition) will execute, but there is one executor per application, in this
case, I have the following 2 questions, assuming that the worker memory is set
to 48G, and executor memory is set to 4G, and I use one spark-shell to connect
to the master to submit my application:
1) How many executor will be created on this box (Or even in the cluster it it
is running in the cluster)? I don't see any spark configuration related to set
number of executor in spark shell. If it is more than one, how this number is
calculated?2) Do you mean that one partition (or one task for it) will be run
by one executor? Is that one executor will run the task sequentially, but job
concurrency comes from that multi executors could run synchronous, right?
Thanks
> From: so...@cloudera.com
> Date: Sun, 8 Feb 2015 20:53:15 +0000
> Subject: Re: Spark concurrency question
> To: java8...@hotmail.com
> CC: user@spark.apache.org
>
> I think I have this right:
>
> You will run one executor per application per worker. Generally there
> is one worker per machine, and it manages all of the machine's
> resources. So if you want one app to use this whole machine you need
> to ask for 48G and 24 cores. That's better than splitting up the
> resources such that no executor can use more than 4G.
>
> (However with big heaps >32G it can make sense to limit the size of an
> executor, so for example, you could configure to run 3 workers per
> machine each controlling 8 cores and 16G, and ask for smaller
> executors. Still I don't think it would make sense to run 12 workers
> per machine here.)
>
> 10 tasks (1 per partition) will execute. They generally get assigned
> to favor data locality, but here everything's local. If you had 3
> executors of 8 cores, I'm not sure if it's guaranteed to balance but
> it should be using at least 2 executors, since there are 10 tasks and
> 8*3=24 slots.
>
> In your initial scenario, I think it may be waiting because the single
> worker has all of its cores devoted to your first app's single
> executor. You can ask for fewer cores in each spark-shell.
>
> Not sure what you mean about threads. Yes of course threads are used
> within one JVM / executor. It's not an executor per partition; it's a
> task per partition and 1 executor per application per worker (and
> usually 1 worker per machine but not always). One task executes
> serially in one thread and as many tasks as slots can run
> concurrently, and that's 1 slot per core that the executor is using. I
> suppose in theory you could write a function that starts its own
> threads too, but that's not generally a good idea or necessary.
>
> Did you read the docs on the site?
> http://spark.apache.org/docs/latest/cluster-overview.html
> http://spark.apache.org/docs/latest/spark-standalone.html
>
> On Sun, Feb 8, 2015 at 7:18 PM, java8964 <java8...@hotmail.com> wrote:
> > Hi, I have some questions about how the spark run the job concurrently.
> >
> > For example, if I setup the Spark on one standalone test box, which has 24
> > core and 64G memory. I setup the Worker memory to 48G, and Executor memory
> > to 4G, and using spark-shell to run some jobs. Here is something confusing
> > me:
> >
> > 1) Does the above setting mean that I can have up to 12 Executor running in
> > this box at same time?
> > 2) Let's assume that I want to do a line count of one 1280M HDFS file, which
> > has 10 blocks as 128M per block. In this case, when the Spark program starts
> > to run, will it kick off one executor using 10 threads to read these 10
> > blocks hdfs data, or 10 executors to read one block each? Or in other way? I
> > read the Apache spark document, so I know that this 1280M HDFS file will be
> > split as 10 partitions. But how the executor run them, I am not clear.
> > 3) In my test case, I started one Spark-shell to run a very expensive job. I
> > saw in the Spark web UI, there are 8 stages generated, with 200 to 400 tasks
> > in each stage, and the tasks started to run. At this time, I started another
> > spark shell to connect to master, and try to run a small spark program. From
> > the spark-shell, it shows my new small program is in a wait status for
> > resource. Why? And what kind of resources it is waiting for? If it is
> > waiting for memory, does this means that there are 12 concurrent tasks
> > running in the first program, took 12 * 4G = 48G memory given to the worker,
> > so no more resource available? If so, in this case, then one running task is
> > one executor?
> > 4) In MapReduce, the count of map and reducer tasks are the resource used by
> > the cluster. My understanding is Spark is using multithread, instead of
> > individual JVM processor. In this case, is the Executor using 4G heap to
> > generate multithreads? My real question is that if each executor
> > corresponding to each RDD partition, or executor could span thread for a RDD
> > partition? On the other hand, how the worker decides how many executors to
> > be created?
> >
> > If there is any online document answering the above questions, please let me
> > know. I searched in the Apache Spark site, but couldn't find it.
> >
> > Thanks
> >
> > Yong
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: user-unsubscr...@spark.apache.org
> For additional commands, e-mail: user-h...@spark.apache.org
>