spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Timothy Perrigo <>
Subject Re: Help with Initial Cluster Configuration / Tuning
Date Tue, 22 Oct 2013 18:02:42 GMT
As the newbie who started the conversation, I'd like to thank everyone for
the feedback and the subsequent discussion.  I certainly understand the
point that there's no magic rule book that can take the place of learning
the ins-and-outs of distributed / cluster computing-- a certain amount of
pain is to be expected.  I'd like to add, too, that so far, with Spark,
this pain has been surprisingly minimal, thanks in no small part to the
information I've gleaned (directly or indirectly) from this mailing list.

However, any additional information is always welcome.  In my own case,
what I think I would really benefit from would be a start-to-finish example
of a problem that works on a large-ish dataset.  In particular, it would be
helpful to know what parameters have to be considered, what they are set
to, and the rationale behind how those values were obtained, as well as a
discussion about determining a "good" cluster size / configuration for the
example problem.  (In fact, if anyone knows of such an example, I would be
very appreciative!).  This certainly won't make everything completely
painless, but would be invaluable and certainly seems feasible.

Thanks again everyone for you help and advice.


On Tue, Oct 22, 2013 at 12:01 PM, Mark Hamstra <>wrote:

> Yes, there are certainly rough spots and sharp edges that we can work at
> polishing out and rounding over; and there are people working on such
> things.  Don't get me wrong, feedback from users about what they are
> finding to difficult, opaque or impenetrable is useful; but I don't think
> that the expectation that working with a framework like Spark should be
> smooth and easy can be completely met.  Even when all of the documentation,
> guidance, instrumentation and user interface are in place, there will still
> be a lot for users to come to terms with.
> On Tue, Oct 22, 2013 at 9:50 AM, Aaron Davidson <>wrote:
>> On the other hand, I totally agree that memory usage in Spark is rather
>> opaque, and is one area where we could do a lot better at in terms of
>> communicating issues, through both docs and instrumentation. At least with
>> serialization and such, you can get meaningful exceptions (hopefully), but
>> OOMs are just blanket "something wasn't right somewhere." Debugging them
>> empirically would require deep diving into Spark's heap allocations, which
>> requires a lot more knowledge of Spark internals than should be required
>> for general usage.
>> On Tue, Oct 22, 2013 at 9:22 AM, Mark Hamstra <>wrote:
>>> Yes, but that also illustrates the problem faced by anyone trying to
>>> write a "little white paper or guide lines" to make newbies' experience
>>> painless.  Distributed computing clusters are necessarily complex things,
>>> and problems can crop up in multiple locations, layers or subsystems.  It's
>>> just not feasible to quickly bring up to speed someone with no experience
>>> in distributed programming and cluster systems.  It takes a lot of
>>> knowledge, both broad and deep.  Very few people have the complete scope of
>>> knowledge and experience required, so creating, debugging and maintaining a
>>> cluster computing application almost always has to be a team effort.
>>> Support organizations and communities can replace some of the need for a
>>> knowledgeable and well-functioning team, but not all of it; and at some
>>> point you have to expect that debugging is going to take a considerable
>>> amount of painstaking, systematic effort -- including a close reading of
>>> the available docs.
>>> Several people are working on making more and better reference and
>>> training material available, and some of that will include trouble-shooting
>>> guidance, but that doesn't mean that there can ever be "one little paper"
>>> to solve newbies' (or more experienced developers') problems or provide
>>> adequate guidance.  There's just too much to cover and too many different
>>> kinds or levels of initial-user knowledge to make that completely feasible.
>>> On Tue, Oct 22, 2013 at 8:50 AM, Shay Seng <> wrote:
>>>> Hey Mark, I didn't mean to say that the information isn't out there --
>>>> just that when something goes wrong with spark, the scope of what could be
>>>> wrong is so large - some bad setting with JVM, serializer, akka, badly
>>>> written scala code, algorithm wrong, check worker logs, check executor
>>>> stderrs, ....
>>>> When I looked at this post this morning, my initial thought wasn't that
>>>> "countByValue" would be at fault. ...probably since I've only been using
>>>> Scala/Spark for a month or so.
>>>> It was just a suggestion to help newbies come up to speed more quickly
>>>> and gain insights into how to debug issues.
>>>> On Tue, Oct 22, 2013 at 8:14 AM, Mark Hamstra <>wrote:
>>>>> There's no need to guess at that.  The docs tell you directly:
>>>>> def countByValue(): Map[T, Long]
>>>>> Return the count of each unique value in this RDD as a map of (value,
>>>>> count) pairs. The final combine step happens locally on the master,
>>>>> equivalent to running a single reduce task.
>>>>> On Tue, Oct 22, 2013 at 7:22 AM, Shay Seng <>
>>>>>> Hi Matei,
>>>>>> I've seen several memory tuning queries on this mailing list, and
>>>>>> also heard the same kinds of queries at the spark meetup. In fact
the last
>>>>>> bullet point in Josh Carver(?) slides, the guy from Bizo, was "memory
>>>>>> tuning is still a mystery".
>>>>>> I certainly had lots of issues in when I first started. From memory
>>>>>> issues to gc issues, things seem to run fine until you try something
>>>>>> 500GB of data etc.
>>>>>> I was wondering if you could write up a little white paper or some
>>>>>> guide lines on how to set memory values, and what to look at when
>>>>>> goes wrong? Eg. I would never gave guessed that countByValue happens
on a
>>>>>> single machine etc.
>>>>>>  On Oct 21, 2013 6:18 PM, "Matei Zaharia" <>
>>>>>> wrote:
>>>>>>> Hi there,
>>>>>>> The problem is that countByValue happens in only a single reduce
>>>>>>> task -- this is probably something we should fix but it's basically
>>>>>>> designed for lots of values. Instead, do the count in parallel
as follows:
>>>>>>> val counts = => (str, 1)).reduceByKey((a, b)
=> a + b)
>>>>>>> If this still has trouble, you can also increase the level of
>>>>>>> parallelism of reduceByKey by passing it a second parameter for
the number
>>>>>>> of tasks (e.g. 100).
>>>>>>> BTW one other small thing with your code, flatMap should actually
>>>>>>> work fine if your function returns an Iterator to Traversable,
so there's
>>>>>>> no need to call toList and return a Seq in ngrams; you can just
return an
>>>>>>> Iterator[String].
>>>>>>> Matei
>>>>>>> On Oct 21, 2013, at 1:05 PM, Timothy Perrigo <>
>>>>>>> wrote:
>>>>>>> > Hi everyone,
>>>>>>> > I am very new to Spark, so as a learning exercise I've set
up a
>>>>>>> small cluster consisting of 4 EC2 m1.large instances (1 master,
3 slaves),
>>>>>>> which I'm hoping to use to calculate ngram frequencies from text
files of
>>>>>>> various sizes (I'm not doing anything with them; I just thought
this would
>>>>>>> be slightly more interesting than the usual 'word count' example).
>>>>>>>  Currently, I'm trying to work with a 1GB text file, but running
>>>>>>> memory issues.  I'm wondering what parameters I should be setting
>>>>>>> in order to properly utilize the cluster.  Right
now, I'd be
>>>>>>> happy just to have the process complete successfully with the
1 gig file,
>>>>>>> so I'd really appreciate any suggestions you all might have.
>>>>>>> >
>>>>>>> > Here's a summary of the code I'm running through the spark
>>>>>>> on the master:
>>>>>>> >
>>>>>>> > def ngrams(s: String, n: Int = 3): Seq[String] = {
>>>>>>> >   (s.split("\\s+").sliding(n)).filter(_.length ==
>>>>>>> n).map(_.mkString(" ")).map(_.trim).toList
>>>>>>> > }
>>>>>>> >
>>>>>>> > val text = sc.textFile("s3n://my-bucket/my-1gb-text-file")
>>>>>>> >
>>>>>>> > val mapped = text.filter(_.trim.length > 0).flatMap(ngrams(_,
>>>>>>> >
>>>>>>> > So far so good; the problems come during the reduce phase.
>>>>>>> small files, I was able to issue the following to calculate the
>>>>>>> frequently occurring trigram:
>>>>>>> >
>>>>>>> > val topNgram = (mapped countByValue) reduce((a:(String,
>>>>>>> b:(String, Long)) => if (a._2 > b._2) a else b)
>>>>>>> >
>>>>>>> > With the 1 gig file, though, I've been running into OutOfMemory
>>>>>>> errors, so I decided to split the reduction to several steps,
starting with
>>>>>>> simply issuing countByValue of my "mapped" RDD, but I have yet
to get it to
>>>>>>> complete successfully.
>>>>>>> >
>>>>>>> > SPARK_MEM is currently set to 6154m.  I also bumped up the
>>>>>>> spark.akka.framesize setting to 500 (though at this point, I
was grasping
>>>>>>> at straws; I'm not sure what a "proper" value would be).  What
>>>>>>> should I be setting for a job of this size on a cluster of 3
>>>>>>> slaves? (The cluster was initially configured using the spark-ec2
>>>>>>>  Also, programmatically, what should I be doing differently?
 (For example,
>>>>>>> should I be setting the minimum number of splits when reading
the text
>>>>>>> file?  If so, what would be a good default?).
>>>>>>> >
>>>>>>> > I apologize for what I'm sure are very naive questions.
 I think
>>>>>>> Spark is a fantastic project and have enjoyed working with it,
but I'm
>>>>>>> still very much a newbie and would appreciate any help you all
can provide
>>>>>>> (as well as any 'rules-of-thumb' or best practices I should be
>>>>>>> >
>>>>>>> > Thanks,
>>>>>>> > Tim Perrigo

View raw message