spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jerry Lam <>
Subject Re: Benchmark results between Flink and Spark
Date Tue, 14 Jul 2015 09:02:47 GMT
FYI, another benchmark:

quote: "I have observed a lot of fetch failures while running Spark, which
results in many restarted tasks and, therefore, takes the longest time. I
suspect that executors are incapable of serving shuffle data due to JVMs
doing long garbage-collection (I also tried large numbers for
spark.core.connection.ack.wait.timeout). Flink seems to be irrelevant to GC
issues thanks to its own internal memory management. MapReduce and Tez
execute each task in a separate process and rely on an external auxiliary
service for shuffling. Although the shuffle service could exhibit fetch
failures for other reasons, it works without any fetch failure in this
experiment for Hadoop MapReduce and Tez."

On Mon, Jul 6, 2015 at 3:13 AM, Jan-Paul Bultmann <>

> Sorry, that should be shortest path, and diameter of the graph.
> I shouldn't write emails before I get my morning coffee...
> On 06 Jul 2015, at 09:09, Jan-Paul Bultmann <>
> wrote:
> I would guess the opposite is true for highly iterative benchmarks (common
> in graph processing and data-science).
> Spark has a pretty large overhead per iteration, more optimisations and
> planning only makes this worse.
> Sure people implemented things like dijkstra's algorithm in spark
> (a problem where the number of iterations is bounded by the circumference
> of the input graph),
> but all the datasets I've seen it running on had a very small
> circumference (which is common for e.g. social networks).
> Take sparkSQL for example. Catalyst is a really good query optimiser, but
> it introduces significant overhead.
> Since spark has no iterative semantics on its own (unlike flink),
> one has to materialise the intermediary dataframe at each iteration
> boundary to determine if a termination criterion is reached.
> This causes a huge amount of planning, especially since it looks like
> catalyst will try to optimise the dependency graph
> regardless of caching. A dependency graph that grows in the number of
> iterations and thus the size of the input dataset.
> In flink on the other hand, you can describe you entire iterative program
> through transformations without ever calling an action.
> This means that the optimiser will only have to do planing once.
> Just my 2 cents :)
> Cheers, Jan
> On 06 Jul 2015, at 06:10, wrote:
> Maybe some flink benefits from some pts they outline here:
> Probably if re-ran the benchmarks with 1.5/tungsten line would close the
> gap a bit(or a lot) with spark moving towards similar style off-heap memory
> mgmt, more planning optimizations
> *From:* Jerry Lam [ <>]
> *Sent:* Sunday, July 5, 2015 6:28 PM
> *To:* Ted Yu
> *Cc:* Slim Baltagi; user
> *Subject:* Re: Benchmark results between Flink and Spark
> Hi guys,
> I just read the paper too. There is no much information regarding why
> Flink is faster than Spark for data science type of workloads in the
> benchmark. It is very difficult to generalize the conclusion of a benchmark
> from my point of view. How much experience the author has with Spark is in
> comparisons to Flink is one of the immediate questions I have. It would be
> great if they have the benchmark software available somewhere for other
> people to experiment.
> just my 2 cents,
> Jerry
> On Sun, Jul 5, 2015 at 4:35 PM, Ted Yu <> wrote:
> There was no mentioning of the versions of Flink and Spark used in
> benchmarking.
> The size of cluster is quite small.
> Cheers
> On Sun, Jul 5, 2015 at 10:24 AM, Slim Baltagi <> wrote:
> Hi
> Apache Flink outperforms Apache Spark in processing machine learning &
> graph
> algorithms and relational queries but not in batch processing!
> The results were published in the proceedings of the 18th International
> Conference, Business Information Systems 2015, PoznaƄ, Poland, June 24-26,
> 2015.
> Thanks to our friend Google, Chapter 3: 'Evaluating New Approaches of Big
> Data Analytics Frameworks' by Norman Spangenberg, Martin Roth and Bogdan
> Franczyk is available for preview at on pages 28-37.
> Enjoy!
> Slim Baltagi
> <>
> --
> View this message in context:
> Sent from the Apache Spark User List mailing list archive at
> <>.
> ---------------------------------------------------------------------
> To unsubscribe, e-mail:
> For additional commands, e-mail:

View raw message