[SPARK-10329] Cost RDD in k-means|| initialization is not storage-efficient

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Later
• Affects Version/s: 1.3.1, 1.4.1, 1.5.0
• Component/s: MLlib
• Labels: clustering
• Target Version/s: 1.6.0
• Assignee: hujiayin
• Reporter: Xiangrui Meng
• Shepherd: Xiangrui Meng
• Created: 28/Aug/15 03:39
• Updated: 21/Dec/15 15:36
• Resolved: 06/Nov/15 19:40

Description

Currently we use `RDD[Vector]` to store point cost during k-means|| initialization, where each `Vector` has size `runs`. This is not storage-efficient because `runs` is usually 1 and then each record is a Vector of size 1. What we need is just the 8 bytes to store the cost, but we introduce two objects (DenseVector and its values array), which could cost 16 bytes. That is 200% overhead. Thanks Grace Huang and Jiayin Hu from Intel for reporting this issue!

There are several solutions:

1. Use `RDD[Array[Double]]` instead of `RDD[Vector]`, which saves 8 bytes per record.
2. Use `RDD[Array[Double]]` but batch the values for storage, e.g. each `Array[Double]` object covers 1024 instances, which could remove most of the overhead.

Besides, using MEMORY_AND_DISK instead of MEMORY_ONLY could prevent cost RDDs kicking out the training dataset from memory.

Issue Links

• is related to: SPARK-10354 First cost RDD shouldn't be cached in k-means|| and the following cost RDD should use MEMORY_AND_DISK
• is related to: SPARK-12450 Un-persist broadcasted variables in KMeans
• relates to: SPARK-11358 Deprecate `runs` in k-means
• links to: [Github] Pull Request #8546 (HuJiayin)