+/*
+ * When the children are equally busy queue incoming requests to a single
+ * child for N microseconds. This is done to maximize the likelihood that
+ * the Linux elevator will be able to merge requests while it is plugged.
+ * Otherwise, requests are queued to the least busy device.
+ *
+ * For rotational disks the Linux elevator will plug for 10ms which is
+ * why zfs_vdev_mirror_switch_us is set to 10ms by default. For non-
+ * rotational disks the elevator will not plug, but 10ms is still a small
+ * enough value that the requests will get spread over all the children.
+ *
+ * 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;