Marko Asplund created SPARK-10791:
-------------------------------------
Summary: Optimize MLlib LDA topic distribution query performance
Key: SPARK-10791
URL: https://issues.apache.org/jira/browse/SPARK-10791
Project: Spark
Issue Type: Improvement
Components: MLlib
Affects Versions: 1.5.0
Environment: Ubuntu 13.10, Oracle Java 8
Reporter: Marko Asplund
I've been testing MLlib LDA training with 100 topics, 105 K vocabulary size and
~3.4 M documents using EMLDAOptimizer.
Training the model took ~2.5 hours with MLlib, whereas with Vowpal Wabbit
training with the same data and on the same system set took ~5 minutes. Loading
the persisted model from disk (~2 minutes), as well as querying LDA model topic
distributions (~4 seconds for one document) are also quite slow operations.
Our application is querying LDA model topic distribution (for one doc at a
time) as part of end-user operation execution flow, so a ~4 second execution
time is very problematic.
The log includes the following message, which AFAIK, should mean that
netlib-java is using machine optimised native implementation:
"com.github.fommil.jni.JniLoader - successfully loaded
/tmp/jniloader4682745056459314976netlib-native_system-linux-x86_64.so"
My test code can be found here:
https://github.com/marko-asplund/tech-protos/blob/08e9819a2108bf6bd4d878253c4aa32510a0a9ce/mllib-lda/src/main/scala/fi/markoa/proto/mllib/LDADemo.scala#L56-L57
I also tried using the OnlineLDAOptimizer, but there wasn't a noticeable change
in training performance. Model loading time was reduced to ~ 5 seconds from ~ 2
minutes (now persisted as LocalLDAModel). However, query / prediction time was
unchanged.
Unfortunately, this is the critical performance characteristic in our case.
I did some profiling for my LDA prototype code that requests topic
distributions from a model. According to Java Mission Control more than 80 % of
execution time during sample interval is spent in the following methods:
- org.apache.commons.math3.util.FastMath.log(double); count: 337; 47.07%
- org.apache.commons.math3.special.Gamma.digamma(double); count: 164; 22.91%
- org.apache.commons.math3.util.FastMath.log(double, double[]); count: 50;
6.98%
- java.lang.Double.valueOf(double); count: 31; 4.33%
Is there any way of using the API more optimally?
Are there any opportunities for optimising the "topicDistributions" code
path in MLlib?
My query test code looks like this essentially:
// executed once
val model = LocalLDAModel.load(ctx, ModelFileName)
// executed four times
val samples = Transformers.toSparseVectors(vocabularySize,
ctx.parallelize(Seq(input))) // fast
model.topicDistributions(samples.zipWithIndex.map(_.swap)) // <== this
seems to take about 4 seconds to execute
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
