mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Dmitriy Lyubimov <>
Subject Re: SSVD + PCA
Date Tue, 21 Aug 2012 17:10:27 GMT
On Aug 21, 2012 8:52 AM, "Pat Ferrel" <> wrote:
> RE:  the -t param. I understand your point but the requirement seems
counter to the philosophy of Hadoop where it's the responsibility of the
hadoop cluster admin to determine the number of jops, tasks, mappers, and
reducers that can run on the cluster or any node. Some tweaking is required
for specific jobs but that is not why you have the -t. As far as I know
there is nothing special about the SSVD job reducers.

Having too many reducers may still cause qr blocking deficiency in reduce
tasks of power iterations and potentially mappers of the next stack. This
needs to be controlled explicitly esp as i said in large clusters.

Requiring -t forces a user to change their scripts (is it required in the
API too? now that would be bad)

No the api does not require it set

every time the cluster config changes or when running on a different
cluster. And as you say, if people don't understand the use of multiple
reducers they will not understand the -t anyway.
> I'd vote to make the param optional especially in the API. I personally
would rather leave it up to the hadoop config to determine.
> BTW if anyone else is reading this, the SSVD ran remarkably fast on my
micro cluster (8 cores in two machines) for 295625 docs and 337258 terms
even with worst case parameters. I don't think it took a complete quarter
of football to finish (50 minutes actually), which gave me something to
cheer about  :-P

Yes it is usually quite fast and accurate on modestly sized problems. One
known bottleneck is mapper side matrix multiplication during power
iterations on super sparse matrices. If your input dimensions are quite
large and the problem is supersparse it may become sensitive to available
memory in mappers of power iterations (q>0). Once it starts scratching the
swap partition in disk due to memory pressures it may actually run quite
slowly in power iteration phase(ABt job)

> On Aug 20, 2012, at 8:23 AM, Dmitriy Lyubimov <> wrote:
> On Aug 19, 2012 1:06 AM, "Pat Ferrel" <> wrote:
> >
> > -t Param
> >
> > I'm no hadoop expert but there are a couple parameters for each node in
> cluster that specifies the default number of mappers and reducers for that
> node. There is a rule of thumb about how many mappers and reducers per
> core. You can tweak them either way depending on your typical jobs.
> >
> > No idea what you mean about the total reducers being 1 for most configs.
> My very small cluster at home with 10 cores in three machines is
> to produce a conservative 10 mappers and 10 reducers, which is about what
> happens with balanced jobs. The reducers = 1 is probably for a
> non-clustered one machine setup.
> Yes i agree i was thinking the same and relying on people doing the right
> thing initially. And the life proved me wrong. Absolutely all crews who
> tried the method, not only did they not have reducers set up in their
> client conf, but also they failed to use t parameter to fix it. Also they
> all failed to diagnose it on their own (i.e. simply noticing it in the job
> stats). I think it has something to do with a typical background of our
> customer.
> >
> > I'm suspicious that the -t  parameter is not needed but would definitely
> defer to a hadoop master. In any case I set it to 10 for my mini cluster.
> Recommended value is 95% of the cluster capacity to leave space for
> opportunistic execution. Although on bigger clustes, i am far from sure
> that too many reducers may be that beneficial for a particular problem.
> Hence again override of default in command line may be useful.
> Also one usually ha more than 1 task capacity per node, so i would expect
> your cluster to be able to run up to 40 reducers, typically
> >
> > Variance Retained
> >
> > If one batch of data yields a greatly different estimate of VR than
> another, it would be worth noticing, even if we don't know the actual
> in it. To say that your estimate of VR is valueless would require that we
> have some experience with it, no?
> I am not saying it is valueless. Actually i am hoping it is useful, or i
> wouldnt inckude it in the howto. I am just saying it is something i leave
> outside the scope of the method itself.
> >
> > On Aug 18, 2012, at 10:39 AM, Dmitriy Lyubimov <>
> >
> > On Aug 18, 2012 8:32 AM, "Pat Ferrel" <> wrote:
> >>
> >> Switching from API to CLI
> >>
> >> the parameter -t is described in the PDF
> >>
> >> --reduceTasks <int-value> optional. The number of reducers to use
> > applicable): depends on the size of the hadoop cluster. At this point it
> > could also be overwritten by a standard hadoop property using -D option
> >> 4. Probably always needs to be speciļ¬ed as by default Hadoop would set
> it
> > to 1, which is certainly far below the cluster capacity. Recommended
> > for this option ~ 95% or ~190% of available reducer capacity to allow
> > opportunistic executions.
> >>
> >> The description above seems to say it will be taken from the hadoop
> > config if not specified, which is probably all most people would every
> > want. I am unclear why this is needed? I cannot run SSVD without
> specifying
> > it, in other words it does not seem to be optional?
> >
> > This parameter was made mandatory because people were repeatedly
> forgetting
> > set the number of reducers and kept coming back with questions like why
> > is running so slow. So there was an issue in 0.7 where i made it
> mandatory.
> > I am actually not sure now other mahout methods ensure reducer
> > specification is always specified other than 1
> >
> >>
> >> As a first try using the CLI I'm running with 295625 rows and 337258
> > columns using the following parameters to get a sort of worst case run
> time
> > result with best case data output. The parameters will be tweaked later
> > get better dimensional reduction and runtime.
> >>
> >>   mahout ssvd -i b2/matrix -k 500 -q 1 -pca -o b2/ssvd-out -t (depends
> > on cluster)
> >>
> >> Is there work being done to calculate the variance retained for the
> > output or should I calculate it myself?
> >
> > No theres no work done since it implies your are building your own
> pipeline
> > for a particular purpose. It also takes a lot of assumptions that may or
> > may not hold in a  particular case, such that you do something
> > and corpuses are of similar nature. Also, i know no paper that would do
> > exactly the way i described, so theres no error estimate on either
> > inequality approach or any sort of decay interpolation.
> >
> > It is not very difficult to experiment a little with your data though
> > a subset of the corpus and see what may work.
> >

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message