spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Xi Shen <davidshe...@gmail.com>
Subject Re: Why k-means cluster hang for a long time?
Date Mon, 30 Mar 2015 22:55:32 GMT
For the same amount of data, if I set the k=500, the job finished in about
3 hrs. I wonder if I set k=5000, the job could finish in 30 hrs...the
longest time I waited was 12 hrs...

If I use kmeans-random, same amount of data, k=5000, the job finished in
less than 2 hrs.

I think current kmeans|| implementation could not handle large vector
dimensions properly. In my case, my vector has about 350 dimensions. I
found another post complaining about kmeans performance in Spark, and that
guy has vectors of 200 dimensions.

It is possible people never tested large dimension case.


Thanks,
David




On Tue, Mar 31, 2015 at 4:00 AM Xiangrui Meng <mengxr@gmail.com> wrote:

> Hi Xi,
>
> Please create a JIRA if it takes longer to locate the issue. Did you
> try a smaller k?
>
> Best,
> Xiangrui
>
> On Thu, Mar 26, 2015 at 5:45 PM, Xi Shen <davidshen84@gmail.com> wrote:
> > Hi Burak,
> >
> > After I added .repartition(sc.defaultParallelism), I can see from the
> log
> > the partition number is set to 32. But in the Spark UI, it seems all the
> > data are loaded onto one executor. Previously they were loaded onto 4
> > executors.
> >
> > Any idea?
> >
> >
> > Thanks,
> > David
> >
> >
> > On Fri, Mar 27, 2015 at 11:01 AM Xi Shen <davidshen84@gmail.com> wrote:
> >>
> >> How do I get the number of cores that I specified at the command line? I
> >> want to use "spark.default.parallelism". I have 4 executors, each has 8
> >> cores. According to
> >> https://spark.apache.org/docs/1.2.0/configuration.html#
> execution-behavior,
> >> the "spark.default.parallelism" value will be 4 * 8 = 32...I think it
> is too
> >> large, or inappropriate. Please give some suggestion.
> >>
> >> I have already used cache, and count to pre-cache.
> >>
> >> I can try with smaller k for testing, but eventually I will have to use
> k
> >> = 5000 or even large. Because I estimate our data set would have that
> much
> >> of clusters.
> >>
> >>
> >> Thanks,
> >> David
> >>
> >>
> >> On Fri, Mar 27, 2015 at 10:40 AM Burak Yavuz <brkyvz@gmail.com> wrote:
> >>>
> >>> Hi David,
> >>> The number of centroids (k=5000) seems too large and is probably the
> >>> cause of the code taking too long.
> >>>
> >>> Can you please try the following:
> >>> 1) Repartition data to the number of available cores with
> >>> .repartition(numCores)
> >>> 2) cache data
> >>> 3) call .count() on data right before k-means
> >>> 4) try k=500 (even less if possible)
> >>>
> >>> Thanks,
> >>> Burak
> >>>
> >>> On Mar 26, 2015 4:15 PM, "Xi Shen" <davidshen84@gmail.com> wrote:
> >>> >
> >>> > The code is very simple.
> >>> >
> >>> > val data = sc.textFile("very/large/text/file") map { l =>
> >>> >   // turn each line into dense vector
> >>> >   Vectors.dense(...)
> >>> > }
> >>> >
> >>> > // the resulting data set is about 40k vectors
> >>> >
> >>> > KMeans.train(data, k=5000, maxIterations=500)
> >>> >
> >>> > I just kill my application. In the log I found this:
> >>> >
> >>> > 15/03/26 11:42:43 INFO storage.BlockManagerMaster: Updated info of
> >>> > block broadcast_26_piece0
> >>> > 15/03/26 23:02:57 WARN server.TransportChannelHandler: Exception in
> >>> > connection from
> >>> > workernode0.xshe3539-hadoop-sydney.q10.internal.cloudapp.
> net/100.72.84.107:56277
> >>> > java.io.IOException: An existing connection was forcibly closed by
> the
> >>> > remote host
> >>> >
> >>> > Notice the time gap. I think it means the work node did not generate
> >>> > any log at all for about 12hrs...does it mean they are not working
> at all?
> >>> >
> >>> > But when testing with very small data set, my application works and
> >>> > output expected data.
> >>> >
> >>> >
> >>> > Thanks,
> >>> > David
> >>> >
> >>> >
> >>> > On Fri, Mar 27, 2015 at 10:04 AM Burak Yavuz <brkyvz@gmail.com>
> wrote:
> >>> >>
> >>> >> Can you share the code snippet of how you call k-means? Do you
cache
> >>> >> the data before k-means? Did you repartition the data?
> >>> >>
> >>> >> On Mar 26, 2015 4:02 PM, "Xi Shen" <davidshen84@gmail.com>
wrote:
> >>> >>>
> >>> >>> OH, the job I talked about has ran more than 11 hrs without
a
> >>> >>> result...it doesn't make sense.
> >>> >>>
> >>> >>>
> >>> >>> On Fri, Mar 27, 2015 at 9:48 AM Xi Shen <davidshen84@gmail.com>
> >>> >>> wrote:
> >>> >>>>
> >>> >>>> Hi Burak,
> >>> >>>>
> >>> >>>> My iterations is set to 500. But I think it should also
stop of
> the
> >>> >>>> centroid coverages, right?
> >>> >>>>
> >>> >>>> My spark is 1.2.0, working in windows 64 bit. My data set
is about
> >>> >>>> 40k vectors, each vector has about 300 features, all normalised.
> All work
> >>> >>>> node have sufficient memory and disk space.
> >>> >>>>
> >>> >>>> Thanks,
> >>> >>>> David
> >>> >>>>
> >>> >>>>
> >>> >>>> On Fri, 27 Mar 2015 02:48 Burak Yavuz <brkyvz@gmail.com>
wrote:
> >>> >>>>>
> >>> >>>>> Hi David,
> >>> >>>>>
> >>> >>>>> When the number of runs are large and the data is not
properly
> >>> >>>>> partitioned, it seems that K-Means is hanging according
to my
> experience.
> >>> >>>>> Especially setting the number of runs to something
high
> drastically
> >>> >>>>> increases the work in executors. If that's not the
case, can you
> give more
> >>> >>>>> info on what Spark version you are using, your setup,
and your
> dataset?
> >>> >>>>>
> >>> >>>>> Thanks,
> >>> >>>>> Burak
> >>> >>>>>
> >>> >>>>> On Mar 26, 2015 5:10 AM, "Xi Shen" <davidshen84@gmail.com>
> wrote:
> >>> >>>>>>
> >>> >>>>>> Hi,
> >>> >>>>>>
> >>> >>>>>> When I run k-means cluster with Spark, I got this
in the last
> two
> >>> >>>>>> lines in the log:
> >>> >>>>>>
> >>> >>>>>> 15/03/26 11:42:42 INFO spark.ContextCleaner: Cleaned
broadcast
> 26
> >>> >>>>>> 15/03/26 11:42:42 INFO spark.ContextCleaner: Cleaned
shuffle 5
> >>> >>>>>>
> >>> >>>>>>
> >>> >>>>>>
> >>> >>>>>> Then it hangs for a long time. There's no active
job. The driver
> >>> >>>>>> machine is idle. I cannot access the work node,
I am not sure
> if they are
> >>> >>>>>> busy.
> >>> >>>>>>
> >>> >>>>>> I understand k-means may take a long time to finish.
But why no
> >>> >>>>>> active job? no log?
> >>> >>>>>>
> >>> >>>>>>
> >>> >>>>>> Thanks,
> >>> >>>>>> David
> >>> >>>>>>
>

Mime
View raw message