You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@spark.apache.org by Xiangrui Meng <me...@gmail.com> on 2015/06/18 00:58:38 UTC
Re: Parallel parameter tuning: distributed execution of MLlib algorithms
On Fri, May 22, 2015 at 6:15 AM, Hugo Ferreira <hm...@inesctec.pt> wrote:
> Hi,
>
> I am currently experimenting with linear regression (SGD) (Spark + MLlib,
> ver. 1.2). At this point in time I need to fine-tune the hyper-parameters. I
> do this (for now) by an exhaustive grid search of the step size and the
> number of iterations. Currently I am on a dual core that acts as a master
> (local mode for now but will be adding spark worker later). In order to
> maximize throughput I need to execute each execution of the linear
> regression algorithm in parallel.
>
How big is your dataset? If it is small or medium-sized, you might get better
performance by broadcasting the entire dataset and use a single machine solver
on each workers.
> According to the documentation it seems like parallel jobs may be scheduled
> if they are executed in separate threads [1]. So this brings me to my first
> question: does this mean I am CPU bound by the Spark master? In other words
> the maximum number of jobs = maximum number of threads of the OS?
>
We use the driver to collect model updates. Increasing the number of
parallel jobs
also increasing the driver load for both communication and computation. I don't
think you need to worry much about the max number of threads, which is usually
much larger than the number of parallel jobs we can actually run.
> I searched the mailing list but did not find anything regarding MLlib
> itself. I even peaked into the new MLlib API that uses pipelines and has
> support for parameter tuning. However, it looks like each job (instance of
> the learning algorithm) is executed in sequence. Can anyone confirm this?
> This brings me to my 2ndo question: is their any example that shows how one
> can execute MLlib algorithms as parallel jobs?
>
The new API is not optimized for performance yet. There is an example
here for k-means:
https://github.com/apache/spark/blob/master/mllib/src/main/scala/org/apache/spark/mllib/clustering/KMeans.scala#L393
> Finally, is their any general technique I can use to execute an algorithm in
> a distributed manner using Spark? More specifically I would like to have
> several MLlib algorithms run in parallel. Can anyone show me an example of
> sorts to do this?
>
> TIA.
> Hugo F.
>
>
>
>
>
>
>
> [1] https://spark.apache.org/docs/1.2.0/job-scheduling.html
>
>
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
> For additional commands, e-mail: user-help@spark.apache.org
>
---------------------------------------------------------------------
To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
For additional commands, e-mail: user-help@spark.apache.org
Re: Parallel parameter tuning: distributed execution of MLlib algorithms
Posted by Peter Rudenko <pe...@gmail.com>.
Hi, here's how to get Parrallel search pipleine:
package org.apache.spark.ml.pipeline
import org.apache.spark.ml.param.ParamMap
import org.apache.spark.ml.{Pipeline, PipelineModel}
import org.apache.spark.sql._
class ParralelGridSearchPipelineextends Pipeline {
override def fit(dataset: DataFrame, paramMaps: Array[ParamMap]):Seq[PipelineModel] = {
paramMaps.par.map(fit(dataset, _)).toVector
}
}
For this you need: 1) Make sure you have a lot of RAM, since for each
parameter you need to cache label points in LR (i've made a bit
different - first run sequentially for the first param - cache instances
and after run in parralell. You can check prototype here:
https://issues.apache.org/jira/browse/SPARK-5844?focusedCommentId=14323253&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-14323253)
2) Set spark.scheduler.mode="FAIR" othervise tasks submitted within same
context would execute in FIFO mode - so no parrallelizm 3) Also probably
would need to configure pool to use FAIR scheduler also
(http://spark.apache.org/docs/latest/job-scheduling.html#configuring-pool-properties)
What i'm currently looking is to have fork/join pipeline. I have 2
separate branches in my DAG pipeline - to process Numeric columns and to
process categorical columns and after merge everything together with
VectorAssembler. But i want this 2 branches to process in parallel. Also
looking to define a bunch of different crossvalidators that uses other
technics than grid search (random search crossvalidator, bayesian
optimization CV, etc.). Thanks, Peter Rudenko On 2015-06-18 01:58,
Xiangrui Meng wrote:
> On Fri, May 22, 2015 at 6:15 AM, Hugo Ferreira <hm...@inesctec.pt> wrote:
>> Hi,
>>
>> I am currently experimenting with linear regression (SGD) (Spark + MLlib,
>> ver. 1.2). At this point in time I need to fine-tune the hyper-parameters. I
>> do this (for now) by an exhaustive grid search of the step size and the
>> number of iterations. Currently I am on a dual core that acts as a master
>> (local mode for now but will be adding spark worker later). In order to
>> maximize throughput I need to execute each execution of the linear
>> regression algorithm in parallel.
>>
> How big is your dataset? If it is small or medium-sized, you might get better
> performance by broadcasting the entire dataset and use a single machine solver
> on each workers.
>
>> According to the documentation it seems like parallel jobs may be scheduled
>> if they are executed in separate threads [1]. So this brings me to my first
>> question: does this mean I am CPU bound by the Spark master? In other words
>> the maximum number of jobs = maximum number of threads of the OS?
>>
> We use the driver to collect model updates. Increasing the number of
> parallel jobs
> also increasing the driver load for both communication and computation. I don't
> think you need to worry much about the max number of threads, which is usually
> much larger than the number of parallel jobs we can actually run.
>
>> I searched the mailing list but did not find anything regarding MLlib
>> itself. I even peaked into the new MLlib API that uses pipelines and has
>> support for parameter tuning. However, it looks like each job (instance of
>> the learning algorithm) is executed in sequence. Can anyone confirm this?
>> This brings me to my 2ndo question: is their any example that shows how one
>> can execute MLlib algorithms as parallel jobs?
>>
> The new API is not optimized for performance yet. There is an example
> here for k-means:
>
> https://github.com/apache/spark/blob/master/mllib/src/main/scala/org/apache/spark/mllib/clustering/KMeans.scala#L393
>
>> Finally, is their any general technique I can use to execute an algorithm in
>> a distributed manner using Spark? More specifically I would like to have
>> several MLlib algorithms run in parallel. Can anyone show me an example of
>> sorts to do this?
>>
>> TIA.
>> Hugo F.
>>
>>
>>
>>
>>
>>
>>
>> [1] https://spark.apache.org/docs/1.2.0/job-scheduling.html
>>
>>
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
>> For additional commands, e-mail: user-help@spark.apache.org
>>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
> For additional commands, e-mail: user-help@spark.apache.org
>