[SPARK-19007] Speedup and optimize the GradientBoostedTrees in the "data>memory" scene

• Type: Improvement
• Status: Closed
• Priority: Minor
• Resolution: Later
• Affects Version/s: 2.0.1, 2.0.2, 2.1.0
• Component/s: ML

Environment

A CDH cluster consists of 3 redhat server ,(120G memory、40 cores、43TB disk per server).

• Assignee: Unassigned
• Reporter: zhangdenghui
• Created: 27/Dec/16 03:04
• Updated: 20/Oct/17 17:33
• Resolved: 12/Jan/17 02:41

Description

Test data：80G CTR training data from criteolabs(http://criteolabs.wpengine.com/downloads/download-terabyte-click-logs/ ) ,I used 1 of the 24 days' data.Some features needed to be repalced by new generated continuous features，the way to generate the new features refers to the way mentioned in the xgboost's paper.

Recource allocated: spark on yarn, 20 executors, 8G memory and 2 cores per executor.

Parameters: numIterations 10, maxdepth 8, the rest parameters are default

I tested the GradientBoostedTrees algorithm in mllib using 80G CTR data mentioned above.

It totally costs 1.5 hour, and i found many task failures after 6 or 7 GBT rounds later.Without these task failures and task retry it can be much faster ,which can save about half the time. I think it's caused by the RDD named predError in the while loop of the boost method at GradientBoostedTrees.scala,because the lineage of the RDD named predError is growing after every GBT round, and then it caused failures like this :

(ExecutorLostFailure (executor 6 exited caused by one of the running tasks) Reason: Container killed by YARN for exceeding memory limits. 10.2 GB of 10 GB physical memory used. Consider boosting spark.yarn.executor.memoryOverhead.).

I tried to boosting spark.yarn.executor.memoryOverhead but the meomry it needed is too much (even increase half the memory can't solve the problem) so i think it's not a proper method.

Although it can set the predCheckpoint Interval smaller to cut the line of the lineage but it increases IO cost a lot.

I tried another way to solve this problem.I persisted the RDD named predError every round and use pre_predError to record the previous RDD and unpersist it because it's useless anymore.

Finally it costs about 45 min after i tried my method and no task failure occured and no more memeory added.

So when the data is much larger than memory, my little improvement can speedup the GradientBoostedTrees 1~2 times.

Issue Links

• contains: SPARK-19063 Add parameter for storage levels to LDA
• links to: [Github] Pull Request #16415 (zdh2292390)