Index: sys/cddl/compat/opensolaris/sys/time.h
===================================================================
--- sys/cddl/compat/opensolaris/sys/time.h	(revision 253925)
+++ sys/cddl/compat/opensolaris/sys/time.h	(working copy)
@@ -50,6 +50,8 @@
 #define	SEC_TO_TICK(sec)	((sec) * hz)
 #define	NSEC_TO_TICK(usec)	((usec) / (NANOSEC / hz))
 
+#define NSEC_PER_USEC	1000
+
 #ifdef _KERNEL
 static __inline hrtime_t
 gethrtime(void) {
Index: sys/cddl/contrib/opensolaris/uts/common/fs/zfs/vdev_mirror.c
===================================================================
--- sys/cddl/contrib/opensolaris/uts/common/fs/zfs/vdev_mirror.c	(revision 253925)
+++ sys/cddl/contrib/opensolaris/uts/common/fs/zfs/vdev_mirror.c	(working copy)
@@ -33,6 +33,8 @@
 #include <sys/zio.h>
 #include <sys/fs/zfs.h>
 
+SYSCTL_DECL(_vfs_zfs_vdev);
+
 /*
  * Virtual device vector for mirroring.
  */
@@ -41,6 +43,7 @@
 	vdev_t		*mc_vd;
 	uint64_t	mc_offset;
 	int		mc_error;
+	int		mc_pending;
 	uint8_t		mc_tried;
 	uint8_t		mc_skipped;
 	uint8_t		mc_speculative;
@@ -54,7 +57,20 @@
 	mirror_child_t	mm_child[1];
 } mirror_map_t;
 
-int vdev_mirror_shift = 21;
+/*
+ * When the children are equally busy queue incoming requests to a single
+ * child for N microseconds.
+ * Otherwise, requests are queued to the least busy device.
+ *
+ * For fast SSDs it may make sense to decrease zfs_vdev_mirror_switch_us
+ * significantly to bound the worst case latencies.  It would probably be
+ * ideal to calculate a decaying average of the last observed latencies and
+ * use that to dynamically adjust the zfs_vdev_mirror_switch_us time.
+ */
+int zfs_vdev_mirror_switch_us = 10000;
+TUNABLE_INT("vfs.zfs.vdev.mirror_switch_us", &zfs_vdev_mirror_switch_us);
+SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, mirror_switch_us, CTLFLAG_RW,
+    &zfs_vdev_mirror_switch_us, 0, "Switch mirrors every N usecs");
 
 static void
 vdev_mirror_map_free(zio_t *zio)
@@ -69,6 +85,19 @@
 	zio_vsd_default_cksum_report
 };
 
+static int
+vdev_mirror_pending(vdev_t *vd)
+{
+	vdev_queue_t *vq = &vd->vdev_queue;
+	int pending;
+
+	mutex_enter(&vq->vq_lock);
+	pending = avl_numnodes(&vq->vq_pending_tree);
+	mutex_exit(&vq->vq_lock);
+
+	return (pending);
+}
+
 static mirror_map_t *
 vdev_mirror_map_alloc(zio_t *zio)
 {
@@ -108,6 +137,9 @@
 			mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
 		}
 	} else {
+		int lowest_pending = INT_MAX;
+		int lowest_nr = 1;
+
 		c = vd->vdev_children;
 
 		mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
@@ -114,8 +146,7 @@
 		mm->mm_children = c;
 		mm->mm_replacing = (vd->vdev_ops == &vdev_replacing_ops ||
 		    vd->vdev_ops == &vdev_spare_ops);
-		mm->mm_preferred = mm->mm_replacing ? 0 :
-		    (zio->io_offset >> vdev_mirror_shift) % c;
+		mm->mm_preferred = 0;
 		mm->mm_root = B_FALSE;
 
 		for (c = 0; c < mm->mm_children; c++) {
@@ -122,7 +153,40 @@
 			mc = &mm->mm_child[c];
 			mc->mc_vd = vd->vdev_child[c];
 			mc->mc_offset = zio->io_offset;
+
+			if (mm->mm_replacing)
+				continue;
+
+			if (!vdev_readable(mc->mc_vd)) {
+				mc->mc_error = ENXIO;
+				mc->mc_tried = 1;
+				mc->mc_skipped = 1;
+				mc->mc_pending = INT_MAX;
+				continue;
+			}
+
+			mc->mc_pending = vdev_mirror_pending(mc->mc_vd);
+			if (mc->mc_pending < lowest_pending) {
+				lowest_pending = mc->mc_pending;
+				lowest_nr = 1;
+			} else if (mc->mc_pending == lowest_pending) {
+				lowest_nr++;
+			}
 		}
+
+		d = gethrtime() / (NSEC_PER_USEC * zfs_vdev_mirror_switch_us);
+		d = (d % lowest_nr) + 1;
+
+		for (c = 0; c < mm->mm_children; c++) {
+			mc = &mm->mm_child[c];
+
+			if (mm->mm_child[c].mc_pending == lowest_pending) {
+				if (--d == 0) {
+					mm->mm_preferred = c;
+					break;
+				}
+			}
+		}
 	}
 
 	zio->io_vsd = mm;