spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Shangyu Luo <lsy...@gmail.com>
Subject Re: cluster hangs for no apparent reason
Date Thu, 07 Nov 2013 23:34:03 GMT
I am not sure. But in their RDD paper they have mentioned the usage of
broadcast variable. Sometimes you may need local variable in many
map-reduce jobs and you do not want to copy them to all worker nodes
multiple times. Then the broadcast variable is a good choice


2013/11/7 Walrus theCat <walrusthecat@gmail.com>

> Shangyu,
>
> Thanks for the tip re: the flag!  Maybe the broadcast variable is only for
> "complex" data structures?
>
>
> On Sun, Nov 3, 2013 at 7:58 PM, Shangyu Luo <lsyurd@gmail.com> wrote:
>
>> I met the problem of 'Too many open files' before. One solution is
>> adding 'ulimit -n 100000' in the spark-env.sh file.
>> Basically, I think the local variable may not be a problem as I have
>> written programs with local variables as parameters for functions and the
>> programs work.
>>
>>
>> 2013/11/3 Walrus theCat <walrusthecat@gmail.com>
>>
>>> Hi Shangyu,
>>>
>>> I appreciate your ongoing correspondence.  To clarify, my solution
>>> didn't work, and I didn't expect it to. I was digging through the logs, and
>>> I found a series of exceptions (in only one of the workers):
>>>
>>> 13/11/03 17:51:05 INFO client.DefaultHttpClient: Retrying connect
>>> 13/11/03 17:51:05 INFO http.AmazonHttpClient: Unable to execute HTTP request:
Too many open files
>>> java.net.SocketException: Too many open files
>>> ...
>>> at com.amazonaws.services.s3.AmazonS3Client.getObject(AmazonS3Client.java:808)
>>> ...
>>>
>>> I don't know why, because I do close those streams, but I'll look into it.
>>>
>>>
>>>
>>> As an aside, I make references to a spark.util.Vector from a parallelized context
(in an RDD.map operation), as per the Logistic Regression example that Spark came with, and
it seems to work out (the following from the examples, you'll see that 'w' is not a broadcast
variable, and 'points' is an RDD):
>>>
>>>
>>>
>>>     var w = Vector(D, _ => 2 * rand.nextDouble - 1)
>>>     println("Initial w: " + w)
>>>
>>>     for (i <- 1 to ITERATIONS) {
>>>       println("On iteration " + i)
>>>       val gradient = points.map { p =>
>>>
>>>
>>>
>>>         (1 / (1 + exp(-p.y * (w dot p.x))) - 1) * p.y * p.x
>>>       }.reduce(_ + _)
>>>       w -= gradient
>>>     }
>>>
>>>
>>>
>>>
>>> On Sun, Nov 3, 2013 at 10:47 AM, Shangyu Luo <lsyurd@gmail.com> wrote:
>>>
>>>> Hi Walrus,
>>>> Thank you for sharing your solution to your problem. I think I have met
>>>> the similar problem before (i.e., one machine is working while others are
>>>> idle.) and I just waits for a long time and the program will continue
>>>> processing. I am not sure how your program filters an RDD by a locally
>>>> stored set. If the set is a parameter of a function, I assume it should be
>>>> copied to all worker nodes. But it is good that you solved your problem
>>>> with a broadcast variable and the running time seems reasonable!
>>>>
>>>>
>>>> 2013/11/3 Walrus theCat <walrusthecat@gmail.com>
>>>>
>>>>> Hi Shangyu,
>>>>>
>>>>> Thanks for responding.  This is a refactor of other code that isn't
>>>>> completely scalable because it pulls stuff to the driver.  This code
keeps
>>>>> everything on the cluster.  I left it running for 7 hours, and the log
just
>>>>> froze.  I checked ganglia, and only one machine's CPU seemed to be doing
>>>>> anything.  The last output on the log left my code at a spot where it
is
>>>>> filtering an RDD by a locally stored set.  No error was thrown.  I thought
>>>>> that was OK based on the example code, but just in case, I changed it
so
>>>>> it's a broadcast variable.  The un-refactored code (that pulls all the
data
>>>>> to the driver from time to time) runs in minutes.  I've never had the
>>>>> problem before of the log just getting non-responsive, and was wondering
if
>>>>> anyone knew of any heuristics I could check.
>>>>>
>>>>> Thank you
>>>>>
>>>>>
>>>>> On Sat, Nov 2, 2013 at 2:55 PM, Shangyu Luo <lsyurd@gmail.com>
wrote:
>>>>>
>>>>>> Yes, I think so. The running time depends on what work your are doing
>>>>>> and how large it is.
>>>>>>
>>>>>>
>>>>>> 2013/11/1 Walrus theCat <walrusthecat@gmail.com>
>>>>>>
>>>>>>> That's what I thought, too.  So is it not "hanging", just
>>>>>>> recalculating for a very long time?  The log stops updating and
it just
>>>>>>> gives the output I posted.  If there are any suggestions as to
parameters
>>>>>>> to change, or any other data, it would be appreciated.
>>>>>>>
>>>>>>> Thank you, Shangyu.
>>>>>>>
>>>>>>>
>>>>>>> On Fri, Nov 1, 2013 at 11:31 AM, Shangyu Luo <lsyurd@gmail.com>wrote:
>>>>>>>
>>>>>>>> I think the missing parent may be not abnormal. From my
>>>>>>>> understanding, when a Spark task cannot find its parent,
it can use some
>>>>>>>> meta data to find the result of its parent or recalculate
its parent's
>>>>>>>> value. Imaging in a loop, a Spark task tries to find some
value from the
>>>>>>>> last iteration's result.
>>>>>>>>
>>>>>>>>
>>>>>>>> 2013/11/1 Walrus theCat <walrusthecat@gmail.com>
>>>>>>>>
>>>>>>>>> Are there heuristics to check when the scheduler says
it is
>>>>>>>>> "missing parents" and just hangs?
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On Thu, Oct 31, 2013 at 4:56 PM, Walrus theCat <
>>>>>>>>> walrusthecat@gmail.com> wrote:
>>>>>>>>>
>>>>>>>>>> Hi,
>>>>>>>>>>
>>>>>>>>>> I'm not sure what's going on here.  My code seems
to be working
>>>>>>>>>> thus far (map at SparkLR:90 completed.)  What can
I do to help the
>>>>>>>>>> scheduler out here?
>>>>>>>>>>
>>>>>>>>>> Thanks
>>>>>>>>>>
>>>>>>>>>> 13/10/31 02:10:13 INFO scheduler.DAGScheduler: Completed
>>>>>>>>>> ShuffleMapTask(10, 211)
>>>>>>>>>> 13/10/31 02:10:13 INFO scheduler.DAGScheduler: Stage
10 (map at
>>>>>>>>>> SparkLR.scala:90) finished in 0.923 s
>>>>>>>>>> 13/10/31 02:10:13 INFO scheduler.DAGScheduler: looking
for newly
>>>>>>>>>> runnable stages
>>>>>>>>>> 13/10/31 02:10:13 INFO scheduler.DAGScheduler: running:
Set(Stage
>>>>>>>>>> 11)
>>>>>>>>>> 13/10/31 02:10:13 INFO scheduler.DAGScheduler: waiting:
Set(Stage
>>>>>>>>>> 9, Stage 8)
>>>>>>>>>> 13/10/31 02:10:13 INFO scheduler.DAGScheduler: failed:
Set()
>>>>>>>>>> 13/10/31 02:10:16 INFO scheduler.DAGScheduler: Missing
parents
>>>>>>>>>> for Stage 9: List(Stage 11)
>>>>>>>>>> 13/10/31 02:10:16 INFO scheduler.DAGScheduler: Missing
parents
>>>>>>>>>> for Stage 8: List(Stage 9)
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> --
>>>>>>>> --
>>>>>>>>
>>>>>>>> Shangyu, Luo
>>>>>>>> Department of Computer Science
>>>>>>>> Rice University
>>>>>>>>
>>>>>>>> --
>>>>>>>> Not Just Think About It, But Do It!
>>>>>>>> --
>>>>>>>> Success is never final.
>>>>>>>> --
>>>>>>>> Losers always whine about their best
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>> --
>>>>>> --
>>>>>>
>>>>>> Shangyu, Luo
>>>>>> Department of Computer Science
>>>>>> Rice University
>>>>>>
>>>>>> --
>>>>>> Not Just Think About It, But Do It!
>>>>>> --
>>>>>> Success is never final.
>>>>>> --
>>>>>> Losers always whine about their best
>>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>> --
>>>> --
>>>>
>>>> Shangyu, Luo
>>>> Department of Computer Science
>>>> Rice University
>>>>
>>>> --
>>>> Not Just Think About It, But Do It!
>>>> --
>>>> Success is never final.
>>>> --
>>>> Losers always whine about their best
>>>>
>>>
>>>
>>
>>
>> --
>> --
>>
>> Shangyu, Luo
>> Department of Computer Science
>> Rice University
>>
>> --
>> Not Just Think About It, But Do It!
>> --
>> Success is never final.
>> --
>> Losers always whine about their best
>>
>
>


-- 
--

Shangyu, Luo
Department of Computer Science
Rice University

--
Not Just Think About It, But Do It!
--
Success is never final.
--
Losers always whine about their best

Mime
View raw message