Github user MLnick commented on a diff in the pull request: https://github.com/apache/spark/pull/17742#discussion_r114712062 --- Diff: mllib/src/main/scala/org/apache/spark/mllib/recommendation/MatrixFactorizationModel.scala --- @@ -274,46 +275,62 @@ object MatrixFactorizationModel extends Loader[MatrixFactorizationModel] { srcFeatures: RDD[(Int, Array[Double])], dstFeatures: RDD[(Int, Array[Double])], num: Int): RDD[(Int, Array[(Int, Double)])] = { - val srcBlocks = blockify(rank, srcFeatures) - val dstBlocks = blockify(rank, dstFeatures) - val ratings = srcBlocks.cartesian(dstBlocks).flatMap { - case ((srcIds, srcFactors), (dstIds, dstFactors)) => - val m = srcIds.length - val n = dstIds.length - val ratings = srcFactors.transpose.multiply(dstFactors) - val output = new Array[(Int, (Int, Double))](m * n) - var k = 0 - ratings.foreachActive { (i, j, r) => - output(k) = (srcIds(i), (dstIds(j), r)) - k += 1 + val srcBlocks = blockify(srcFeatures) + val dstBlocks = blockify(dstFeatures) + /** + * The previous approach used for computing top-k recommendations aimed to group + * individual factor vectors into blocks, so that Level 3 BLAS operations (gemm) could + * be used for efficiency. However, this causes excessive GC pressure due to the large + * arrays required for intermediate result storage, as well as a high sensitivity to the + * block size used. + * The following approach still groups factors into blocks, but instead computes the + * top-k elements per block, using Level 1 BLAS (dot) and an efficient + * BoundedPriorityQueue. This avoids any large intermediate data structures and results + * in significantly reduced GC pressure as well as shuffle data, which far outweighs + * any cost incurred from not using Level 3 BLAS operations. + */ + val ratings = srcBlocks.cartesian(dstBlocks).flatMap { case (srcIter, dstIter) => + val m = srcIter.size + val n = math.min(dstIter.size, num) + val output = new Array[(Int, (Int, Double))](m * n) + var j = 0 + val pq = new BoundedPriorityQueue[(Int, Double)](n)(Ordering.by(_._2)) + srcIter.foreach { case (srcId, srcFactor) => + dstIter.foreach { case (dstId, dstFactor) => + /** + * Compared with BLAS.dot, the hand-written version below is more efficient than a call --- End diff -- Let's match the form of comment from https://github.com/apache/spark/pull/17845/files#diff-be65dd1d6adc53138156641b610fcadaR403
--- If your project is set up for it, you can reply to this email and have your reply appear on GitHub as well. If your project does not have this feature enabled and wishes so, or if the feature is enabled but not working, please contact infrastructure at infrastruct...@apache.org or file a JIRA ticket with INFRA. --- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org