Hi Nick, Thanks for the feedback and the pointers. I tried coalescing to fewer partitions and improved the situation dramatically. As you suggested, it is communication overhead dominating the overall runtime.
The training run I mentioned originally had 900 partitions. Each tree aggregation has two stages, one for the original partitions, and then one with the aggregation into a smaller number (at 900 partitions the second stage was 30). The first tree aggregation job (the longer one) uses the MultivariateOnlineSummarizer you mentioned, while the subsequent aggregation jobs use LogisticAggregator (similar problem, though smaller). I've run some tests with fewer partitions on a very similar data set. 400 partitions took 8 hours, 100 partitions took 4 hours, and 10 partitions took 1.4 hours. I put some screenshots from the Spark UI here: http://imgur.com/a/trRJU Still, these numbers seem oddly high. With 10 partitions it's shuffling only some 200 MB per job, but the median "Getting Result Time" is 2.1 minutes. I would expected it to take *seconds* to transfer that data. Anyway, the MultivariateOnlineSummarizer creates several arrays of doubles equal to the size of the vector - arrays of course are inherently dense. While this is only one iteration it is the longest, taking a significant portion of the time by itself. LogisticAggregator meanwhile has fewer arrays, but if you try to pass coefficients as anything other than a dense vector it actually throws an error! Any idea why? Anyone know a reason these aggregators *must* store their data densely, or is just an implementation choice? Perhaps refactoring these classes to store data sparsely would fix the issue. On Wed, Mar 9, 2016 at 7:57 AM, Nick Pentreath <nick.pentre...@gmail.com> wrote: > Hi Daniel > > The bottleneck in Spark ML is most likely (a) the fact that the weight > vector itself is dense, and (b) the related communication via the driver. A > tree aggregation mechanism is used for computing gradient sums (see > https://github.com/apache/spark/blob/master/mllib/src/main/scala/org/apache/spark/ml/classification/LogisticRegression.scala#L1080 > and > https://databricks.com/blog/2014/09/22/spark-1-1-mllib-performance-improvements.html), > which helps efficiency, but ultimately the driver must collect that > gradient vector and re-broadcast the updated weight vector on every > iteration. > > From a quick glance, MultivariateOnlineSummarizer doesn't seem optimized > to sparse data either (though that is only one pass of the dataset so > doubtful it adds too much overhead). > > It would be helpful to understand some further details: > 1. Some more exact timing numbers - if you could provide screenshots / > output from the UI to indicate the stages and call sites where the time is > being spent that would be really useful > 2. Is this part of a pipeline, and if so, what is the contribution of > other parts of that pipeline to overall runtime? > 3. Some stats on input / output data sizes from the critical stages (again > from the UI) > 4. The dataset size (# examples, avg sparsity % per example, etc) > 5. Related to (4), the number of partitions of your dataset > 6. Cluster details (# nodes and spec), as well as Spark version > > If you have a lot of partitions, you could find performance will be better > with fewer partitions because the communication overhead will tend to > dominate the overall runtime. > > Still, 10 hours and >100GB of driver memory seems extreme for a 20 million > size dense weight vector (which should only be a few 100MB memory), so > perhaps something else is going on. > > Nick > > On Tue, 8 Mar 2016 at 22:55 Daniel Siegmann <daniel.siegm...@teamaol.com> > wrote: > >> Just for the heck of it I tried the old MLlib implementation, but it had >> the same scalability problem. >> >> Anyone familiar with the logistic regression implementation who could >> weigh in? >> >> On Mon, Mar 7, 2016 at 5:35 PM, Michał Zieliński < >> zielinski.mich...@gmail.com> wrote: >> >>> We're using SparseVector columns in a DataFrame, so they are definitely >>> supported. But maybe for LR some implicit magic is happening inside. >>> >>> On 7 March 2016 at 23:04, Devin Jones <devin.jo...@columbia.edu> wrote: >>> >>>> I could be wrong but its possible that toDF populates a dataframe which >>>> I understand do not support sparsevectors at the moment. >>>> >>>> If you use the MlLib logistic regression implementation (not ml) you >>>> can pass the RDD[LabeledPoint] data type directly to the learner. >>>> >>>> >>>> http://spark.apache.org/docs/latest/api/scala/index.html#org.apache.spark.mllib.classification.LogisticRegressionWithLBFGS >>>> >>>> Only downside is that you can't use the pipeline framework from spark >>>> ml. >>>> >>>> Cheers, >>>> Devin >>>> >>>> >>>> >>>> On Mon, Mar 7, 2016 at 4:54 PM, Daniel Siegmann < >>>> daniel.siegm...@teamaol.com> wrote: >>>> >>>>> Yes, it is a SparseVector. Most rows only have a few features, and >>>>> all the rows together only have tens of thousands of features, but the >>>>> vector size is ~ 20 million because that is the largest feature. >>>>> >>>>> On Mon, Mar 7, 2016 at 4:31 PM, Devin Jones <devin.jo...@columbia.edu> >>>>> wrote: >>>>> >>>>>> Hi, >>>>>> >>>>>> Which data structure are you using to train the model? If you haven't >>>>>> tried yet, you should consider the SparseVector >>>>>> >>>>>> >>>>>> http://spark.apache.org/docs/latest/api/scala/index.html#org.apache.spark.mllib.linalg.SparseVector >>>>>> >>>>>> >>>>>> On Mon, Mar 7, 2016 at 4:03 PM, Daniel Siegmann < >>>>>> daniel.siegm...@teamaol.com> wrote: >>>>>> >>>>>>> I recently tried to a model using >>>>>>> org.apache.spark.ml.classification.LogisticRegression on a data set >>>>>>> where the feature vector size was around ~20 million. It did *not* >>>>>>> go well. It took around 10 hours to train on a substantial cluster. >>>>>>> Additionally, it pulled a lot data back to the driver - I eventually >>>>>>> set --conf >>>>>>> spark.driver.memory=128g --conf spark.driver.maxResultSize=112g >>>>>>> when submitting. >>>>>>> >>>>>>> Attempting the same application on the same cluster with the feature >>>>>>> vector size reduced to 100k took only ~ 9 minutes. Clearly there is an >>>>>>> issue with scaling to large numbers of features. I'm not doing anything >>>>>>> fancy in my app, here's the relevant code: >>>>>>> >>>>>>> val lr = new LogisticRegression().setRegParam(1) >>>>>>> val model = lr.fit(trainingSet.toDF()) >>>>>>> >>>>>>> In comparison, a coworker trained a logistic regression model on her >>>>>>> *laptop* using the Java library liblinear in just a few minutes. >>>>>>> That's with the ~20 million-sized feature vectors. This suggests to me >>>>>>> there is some issue with Spark ML's implementation of logistic >>>>>>> regression >>>>>>> which is limiting its scalability. >>>>>>> >>>>>>> Note that my feature vectors are *very* sparse. The maximum feature >>>>>>> is around 20 million, but I think there are only 10's of thousands of >>>>>>> features. >>>>>>> >>>>>>> Has anyone run into this? Any idea where the bottleneck is or how >>>>>>> this problem might be solved? >>>>>>> >>>>>>> One solution of course is to implement some dimensionality >>>>>>> reduction. I'd really like to avoid this, as it's just another thing to >>>>>>> deal with - not so hard to put it into the trainer, but then anything >>>>>>> doing >>>>>>> scoring will need the same logic. Unless Spark ML supports this out of >>>>>>> the >>>>>>> box? An easy way to save / load a model along with the dimensionality >>>>>>> reduction logic so when transform is called on the model it will handle >>>>>>> the >>>>>>> dimensionality reduction transparently? >>>>>>> >>>>>>> Any advice would be appreciated. >>>>>>> >>>>>>> ~Daniel Siegmann >>>>>>> >>>>>> >>>>>> >>>>> >>>> >>> >>
