[Beowulf] Beowulf Cluster VS Hadoop/Spark

John Hanks griznog at gmail.com
Fri Dec 30 08:47:03 PST 2016


I suspect that you can take any hadoop/spark application and give it to a
good C/C++/OpenMp/MPI coder and in six months, a year, two years,..., you
will end up with a much faster and much more efficient application.
Meanwhile the original question the application was answering very likely
won't matter to those who originally used hadoop/spark to answer it.

It's worth keeping in mind that a lot (maybe most) of the "big data"
analysis being done is done in Microsoft Excel. Python and R cover a big
chunk of it, often running on laptops. I assume anyone reading a post on
this list, like me, suffers from "cluster bias" which causes us to forget
that the bulk of computational work taking place in the world happens
outside, far outside, of the top 100 machines. And much of that work is
done by people who care more about the total time to solution and will
happily trade a little additional CPU time for a better, easier and more
powerful abstraction to use to ask questions.

Consider also the increasing amount of this work being done by training a
deep learning framework after which the researcher may or may not be able
to explain how/why the thing works. Port that to C :)

In general I always get suspicious at any suggestion of a pure approach to
anything. As with "centralization" efforts in the world of IT, a pure
approach is often code for "arbitrary boundaries for you that we are
comfortable with." Hadoop/spark are great data exploration tools and
someone who understands their data and knows Python can do wonderful things
in a Python notebook backed by an appropriately sized spark cluster and
then be off to the next question before "hello world" can be compiled in C.
I for one welcome our new big data overlords, unless they demand to run
Excel on the cluster.

Good thing it's close to my bedtime, I have exhausted my daily buzzword
quota.

jbh

On Fri, Dec 30, 2016, 11:00 AM Jonathan Aquilina <jaquilina at eagleeyet.net>
wrote:

> Thanks John for your reply. Very interesting food for thought here. What I
> do understand between hadoop and spark is that spark is intended, i could
> be wrong here, as a replacement to hadoop as it performs better and faster
> then hadoop.
>
> Is spark also java based? I never thought java to be so high performant. I
> know when i started learning to program in java (java6) it was slow and
> clunky. Wouldnt it be better to stick with a pure beowulf cluster and build
> yoru apps in c or c++ something that is closer to the machine language then
> the use of an interpreted language such as java? I think where I fall short
> to understand is how with hadoop and spark have they made java so quick
> compared to a compiled language.
>
>
>
> On 2016-12-30 08:47, John Hanks wrote:
>
> This often gets presented as an either/or proposition and it's really not.
> We happily use SLURM to schedule the setup, run and teardown of spark
> clusters. At the end of the day it's all software, even the kernel and OS.
> The big secret of HPC is that in a job scheduler we have an amazingly
> powerful tool to manage resources. Once you are scheduling spark clusters,
> hadoop clusters, VMs as jobs, containers, long running web services, ....,
> you begin to feel sorry for those poor "cloud" people trapped in buzzword
> land.
>
> But, directly to your question what we are learning as we dive deeper into
> spark (interest in hadoop here seems to be minimal and fading) is that it
> is just as hard or maybe harder to tune for than MPI and the people who
> want to use it tend to have a far looser grasp of how to tune it than those
> using MPI. In the short term I think it is beneficial as a sysadmin to
> spend some time learning the inner squishy bits to compensate for that. A
> simple wordcount example or search can show that wc and grep can often
> outperform spark and it takes some experience to understand when a
> particular approach is the better one for a given problem. (Where better is
> measured by efficiency, not by the number of cool new technical toys were
> employed :)
>
> jbh
>
> On Fri, Dec 30, 2016, 9:32 AM Jonathan Aquilina <jaquilina at eagleeyet.net>
> wrote:
>
> Hi All,
>
> Seeing the new activity about new clusters for 2017, this sparked a
> thought in my mind here. Beowulf Cluster vs hadoop/spark
>
> In this day and age given that there is the technology with hadoop and
> spark to crunch large data sets, why build a cluster of pc's instead of use
> something like hadoop/spark?
>
>
>
> Happy New Year
>
> Jonathan Aquilina
>
> Owner EagleEyeT
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit
> http://www.beowulf.org/mailman/listinfo/beowulf
>
> --
> '[A] talent for following the ways of yesterday, is not sufficient to
> improve the world of today.'
>  - King Wu-Ling, ruler of the Zhao state in northern China, 307 BC
>
> --
‘[A] talent for following the ways of yesterday, is not sufficient to
improve the world of today.’
 - King Wu-Ling, ruler of the Zhao state in northern China, 307 BC
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.beowulf.org/pipermail/beowulf/attachments/20161230/61282f41/attachment-0001.html>


More information about the Beowulf mailing list