You are viewing a plain text version of this content. The canonical link for it is here.
Posted to reviews@spark.apache.org by GitBox <gi...@apache.org> on 2020/11/23 11:20:53 UTC
[GitHub] [spark] zhengruifeng edited a comment on pull request #30468: [SPARK-33518][ML][WIP] Improve performance of ML ALS recommendForAll by GEMV
zhengruifeng edited a comment on pull request #30468:
URL: https://github.com/apache/spark/pull/30468#issuecomment-732093851
[GEMV](https://github.com/apache/spark/pull/30468/commits/b6459683e734af3749be3b9ce6047ca22fabfdd9):
```
val ratings = srcFactorsBlocked.crossJoin(dstFactorsBlocked)
.as[(Array[Int], Array[Float], Array[Int], Array[Float])]
.mapPartitions { iter =>
var buffer: Array[Float] = null
val pq = new BoundedPriorityQueue[(Int, Float)](num)(Ordering.by(_._2))
iter.flatMap { case (srcIds, srcMat, dstIds, dstMat) =>
require(srcMat.length == srcIds.length * rank)
require(dstMat.length == dstIds.length * rank)
val m = srcIds.length
val n = dstIds.length
if (buffer == null || buffer.length < n) {
buffer = Array.ofDim[Float](n)
}
Iterator.range(0, m).flatMap { i =>
BLAS.f2jBLAS.sgemv("T", rank, n, 1.0F, dstMat, 0, rank,
srcMat, i * rank, 1, 0.0F, buffer, 0, 1)
pq.clear()
var j = 0
while (j < n) { pq += dstIds(j) -> buffer(j); j += 1 }
val srcId = srcIds(i)
pq.iterator.map { case (dstId, value) => (srcId, dstId, value) }
}
} ++ {
buffer = null
pq.clear()
Iterator.empty
}
}
```
Then I switch to GEMV, which brings siginificent speedup. The size of `buffer` is reduce to **n**.
![als_gemv_jobs_2020_11_23_16_55_10](https://user-images.githubusercontent.com/7322292/99955591-9236cb00-2dbf-11eb-84f2-3a2421412a48.png)
![als_gemv_exe_2020_11_23_16_57_13](https://user-images.githubusercontent.com/7322292/99955609-99f66f80-2dbf-11eb-8ba6-fa64d6e77fa8.png)
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org