[Beowulf] Clustering vs Hadoop/spark

Jim Cownie jcownie at gmail.com
Wed Nov 25 09:05:05 UTC 2020


> From: Douglas Eadline <deadline at eadline.org <mailto:deadline at eadline.org>> 
> Sent: 24 November 2020 18:21
…
> So there is a bit of decoupling of parallel compute mechanism and compute code. (I.e. in all these languages the user does think about cores, interconnect, communications, etc) A higher level abstraction.

I think you’re missing a “not”, Doug. "in all these languages the user does not think about cores, interconnect, communications, etc” 

More generally “It's always the data movement.” We see that here, but it also applies to our CPUs. Look at the performance of Fugaku and the A64FX which has optimised for that rather than lots of (generally unusable) FLOPS, or consider why we use roofline models when tuning compute kernels, and,  when we do, how often they show that we’re bandwidth bound.

If we assume that most of the time is I/O here, then it's unsurprising that changing the language used for the processing makes little difference. You can think of it as a variant of Amdahl, where you map I/O time to Amdahl’s serial time (i.e. it’s invariant) and compute to parallel time (I.e. you can change it). 
For example, if the compute is 10% of the time, the best you can hope for even if you could use my new “psychic” language and run the compute in zero time is to get a 1/0.9 = 11% speedup.
In that context it should be no surprise that the precise performance of the compute doesn't matter much, so you get comparable overall performance independent of the choice of language used for the compute operations.

-- Jim
James Cownie <jcownie at gmail.com>
Mob: +44 780 637 7146


> -----Original Message-----
> From: Douglas Eadline <deadline at eadline.org <mailto:deadline at eadline.org>> 
> Sent: 24 November 2020 18:21
> To: Jonathan Aquilina <jaquilina at eagleeyet.net <mailto:jaquilina at eagleeyet.net>>
> Cc: beowulf at beowulf.org <mailto:beowulf at beowulf.org>
> Subject: RE: [Beowulf] Clustering vs Hadoop/spark
> 
> 
> First I am not a Java expert (very far from it).
> 
> Second, Java holds up quite well against Julia as compared to Python. (so does Lisp!)
> 
>  https://benchmarksgame-team.pages.debian.net/benchmarksgame/fastest/julia.html <https://benchmarksgame-team.pages.debian.net/benchmarksgame/fastest/julia.html>
> 
> Some other tings to consider is the underlying Hadoop plumbing is written in Java (or Scala for Spark).
> However, it is possible with Hadoop to
> create mapper and reducing functions in any language (text based std in/out) Similar to Spark, that can use Python, R, Java, or Scala as a front-end.
> 
> So there is a bit of decoupling of parallel compute mechanism and compute code. (I.e. in all these languages the user does think about cores, interconnect, communications, etc) A higher level abstraction.
> 
> Much of the early Hadoop performance was based on running large capability jobs on gobs of data.
> Jobs that could not be run otherwise (except for Google) So any performance was good. Spark come along and says, lets put it all in a redundant distributed memory structure.
> Speed up is much faster then traditional Hadoop, so Hadoop creates Tez API that does the same thing.
> Performance even out.
> 
> Plus analytics jobs are mostly integer. The floating point often comes into play when running the models, which is often not a big data problem (i.e. don't need a big cluster to run)
> 
> --
> Doug
> 
>> Hi Doug,
>> 
>> Appreciate the clarification where I am not clear is given Hadoop and 
>> derivatives are java based where all of this performance all of a 
>> sudden comes from. Is it due to where the data resides?
>> 
>> At one of my previous jobs I worked with Hadoop through Amazon AWS EMR 
>> managed to churn through 5 years' worth of historical data in 1 week. 
>> Data being calculations on vehicular tracking data.
>> 
>> When I learned java as part of my degree I used to see it as clunky 
>> why go for an interpreted language such as java over something more 
>> low level like c/c++ on a traditional cluster?
>> 
>> Regards,
>> Jonathan
>> 
>> -----Original Message-----
>> From: Douglas Eadline <deadline at eadline.org>
>> Sent: 24 November 2020 17:38
>> To: Jonathan Aquilina <jaquilina at eagleeyet.net>
>> Cc: beowulf at beowulf.org
>> Subject: Re: [Beowulf] Clustering vs Hadoop/spark
>> 
>> 
>>> Hi Guys,
>>> 
>>> I am just wondering what advantages does setting up of a cluster have 
>>> in relation to big data analytics vs using something like Hadoop/spark?
>>> 
>> 
>> Long email and the details are important.
>> 
>> It all comes down to filesystems and schedulers. But first remember, 
>> most Data Analytics projects use many different tools and have various 
>> stages that often require iteration and development (e.g. ETL->Feature
>> Matrix->and running models, repeat, and 80% of the work in in first 
>> Matrix->two
>> steps) And, many end-users do not use Java map-reduce APIs. They use 
>> higher level tools.
>> 
>> Filesystems:
>> 
>> 1) Traditional Hadoop filesystem (HDFS) is about slicing large data 
>> files (or large number of files) across multiple servers, then doing 
>> the map phase on all servers at the same time (moving computation to 
>> where the data "live", reduce phase requires a shuffle (data movement) 
>> and final reduction of data.
>> 
>> 2) On-prem HDFS still makes some sense (longer story) however, in the 
>> Cloud there is move to using native cloud storage using Apache Ozone FS.
>> You loose the "data locality," but gain all the cloud Kubernettes stuff.
>> 
>> 3) Both Hadoop Map-Reduce (mostly Hive RDB applications now) and Spark 
>> do "in-memory" map-reduce for performance reasons.
>> In this case, data locality for processing is not as important, 
>> However, loading and storing files on large multi-server memory 
>> resident jobs still gains from HDFS. Very often Spark writes/reads results into Hive tables.
>> 
>> Schedulers:
>> 
>> 1) Map Reduce scheduling is different than traditional HPC scheduling.
>> The primary Hadoop scheduler is called YARN (Yet Another Resource
>> Negotiator) It has two main features not found in most HPC schedulers, 
>> data locality as a resource and dynamic resource allocation.
>> 
>> 2) Data locality is about moving jobs to where the data (slice) lives 
>> on the storage nodes (hyper-converged storage/compute nodes)
>> 
>> 3) Dynamic resource allocation developed because most map-reduce jobs 
>> need a lot of containers for map phase, but much-much less for reduce 
>> phase, so Hadoop map-reduce can give back resources and ask for more 
>> later in other stages of the DAG (multiple map reduce phases are run 
>> as a Directed Acyclic Graph)
>> 
>> Thus, this model is hard to map on to a traditional HPC cluster.
>> There are map-reduce libraries for MPI. Another way to think about it 
>> is Data Analytics is almost always SIMD, all tools language and 
>> platforms are optimized to take advantage of map-reduce SIMD operations and data flow.
>> 
>> 
>> --
>> Doug
>> 
>> 
>> 
>>> 
>>> Regards,
>>> Jonathan
>>> _______________________________________________
>>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin 
>>> Computing To change your subscription (digest mode or unsubscribe) 
>>> visit https://beowulf.org/cgi-bin/mailman/listinfo/beowulf
>>> 
>> 
>> 
>> --
>> Doug
>> 
> 
> 
> --
> Doug
> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org <mailto:Beowulf at beowulf.org> sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit https://beowulf.org/cgi-bin/mailman/listinfo/beowulf <https://beowulf.org/cgi-bin/mailman/listinfo/beowulf>


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://beowulf.org/pipermail/beowulf/attachments/20201125/48b8751c/attachment-0001.htm>


More information about the Beowulf mailing list