diff --git a/mllib/src/main/scala/org/apache/spark/mllib/recommendation/ALS.scala b/mllib/src/main/scala/org/apache/spark/mllib/recommendation/ALS.scala
index 102742c7c56755a4d15d2ed9b97cb15d3f7c9927..1f5c746a3457cd6d85cac8c3b8eb02640466e548 100644
--- a/mllib/src/main/scala/org/apache/spark/mllib/recommendation/ALS.scala
+++ b/mllib/src/main/scala/org/apache/spark/mllib/recommendation/ALS.scala
@@ -269,7 +269,7 @@ class ALS private (
   private def computeYtY(factors: RDD[(Int, Array[Array[Double]])]) = {
     val n = rank * (rank + 1) / 2
     val LYtY = factors.values.aggregate(new DoubleMatrix(n))( seqOp = (L, Y) => {
-      Y.foreach(y => dspr(1.0, new DoubleMatrix(y), L))
+      Y.foreach(y => dspr(1.0, wrapDoubleArray(y), L))
       L
     }, combOp = (L1, L2) => {
       L1.addi(L2)
@@ -304,6 +304,15 @@ class ALS private (
     }
   }
 
+  /**
+   * Wrap a double array in a DoubleMatrix without creating garbage.
+   * This is a temporary fix for jblas 1.2.3; it should be safe to move back to the
+   * DoubleMatrix(double[]) constructor come jblas 1.2.4.
+   */
+  private def wrapDoubleArray(v: Array[Double]): DoubleMatrix = {
+    new DoubleMatrix(v.length, 1, v: _*)
+  }
+
   /**
    * Flatten out blocked user or product factors into an RDD of (id, factor vector) pairs
    */
@@ -457,7 +466,7 @@ class ALS private (
     // block
     for (productBlock <- 0 until numBlocks) {
       for (p <- 0 until blockFactors(productBlock).length) {
-        val x = new DoubleMatrix(blockFactors(productBlock)(p))
+        val x = wrapDoubleArray(blockFactors(productBlock)(p))
         tempXtX.fill(0.0)
         dspr(1.0, x, tempXtX)
         val (us, rs) = inLinkBlock.ratingsForBlock(productBlock)(p)