SPARK-5016 Perf Testing.md

Cluster

• Spark 1.4 + 1da3c7f
• Databricks Cloud
• 8 Workers, EC2 Spot instances
• Workers: 240 GB Memory, 32 Cores
• Driver: 30 GB Memory, 4 Cores

Data

• 10,000 points, 100 each from N(x,1) where x in [0, 3, 6, ... 297]
• n-dimensional points were generated by n iid draws from N(x,1)

e.g. for 40 features:

val data40D = sc.parallelize(
  (0 to 300 by 3).flatMap { mean =>
    Seq.fill(100)(Vectors.dense((0 until 40).map(_ => rng.nextGaussian() + 100*mean).toArray))
  }
)

Results

Parallel

Num Features	Num Centers	Runtimes
1	10	10.68, 9.14, 8.45
1	50	11.36, 10.08, 9.91
1	100	16.92, 13.53, 16.05
30	10	2.21, 2.23, 2.47
30	50	9.14, 9.16, 8.67
30	100	17.12, 17.13, 16.92
100	10	2.41, 2.59, 2.41, 2.40, 2.36
100	100	14.54, 17.62, 17.11
300	10	16.38, 16.23, 16.21, 16.20, 16.34
300	100	155.86, 147.93, 153.36
40	10000	383.29, 373.50, 372,90

Sequential

Num Features	Num Centers	Runtimes
1	10	5.57, 5.31, 5.57
1	50	8.96, 8.48, 7.18
1	100	10.56, 9.48, 12.88
30	10	2.39, 2.55, 2.33
30	50	9.91, 9.64, 9.55
30	100	18.87, 18.73, 18.64
100	10	3.16, 2.37, 2.31, 2.30, 2.37
100	100	17.86, 21.49, 17.96
300	10	18.44, 17.97, 20.49, 19.00
300	100	183.12, 172.17, 175.38
40	10000	397.97, 396.32, 396.21