Print this page
4334 Improve ZFS N-way mirror read performance
@@ -23,10 +23,11 @@
* Use is subject to license terms.
*/
/*
* Copyright (c) 2013 by Delphix. All rights reserved.
+ * Copyright (c) 2013 Steven Hartland. All rights reserved.
*/
#include <sys/zfs_context.h>
#include <sys/spa.h>
#include <sys/vdev_impl.h>
@@ -39,87 +40,175 @@
typedef struct mirror_child {
vdev_t *mc_vd;
uint64_t mc_offset;
int mc_error;
+ int mc_load;
uint8_t mc_tried;
uint8_t mc_skipped;
uint8_t mc_speculative;
} mirror_child_t;
typedef struct mirror_map {
+ int *mm_preferred;
+ int mm_preferred_cnt;
int mm_children;
- int mm_replacing;
- int mm_preferred;
- int mm_root;
- mirror_child_t mm_child[1];
+ boolean_t mm_replacing;
+ boolean_t mm_root;
+ mirror_child_t mm_child[];
} mirror_map_t;
-int vdev_mirror_shift = 21;
+int zfs_vdev_mirror_shift = 21;
+
+/*
+ * The load configuration settings below are tuned by default for
+ * the case where all devices are of the same rotational type.
+ *
+ * If there is a mixture of rotating and non-rotating media, setting
+ * non_rotating_seek_inc to 0 may well provide better results as it
+ * will direct more reads to the non-rotating vdevs which are more
+ * likely to have a higher performance.
+ */
+
+/* Rotating media load calculation configuration. */
+/* Rotating media load increment for non-seeking I/O's. */
+int zfs_vdev_mirror_rotating_inc = 0;
+
+/* Rotating media load increment for seeking I/O's. */
+int zfs_vdev_mirror_rotating_seek_inc = 5;
+
+/*
+ * Offset in bytes from the last I/O which triggers a reduced rotating media
+ * seek increment.
+ */
+int zfs_vdev_mirror_rotating_seek_offset = 1 * 1024 * 1024;
+
+/* Non-rotating media load calculation configuration. */
+/* Non-rotating media load increment for non-seeking I/O's. */
+int zfs_vdev_mirror_non_rotating_inc = 0;
+
+/* Non-rotating media load increment for seeking I/O's. */
+int zfs_vdev_mirror_non_rotating_seek_inc = 1;
+
+static inline size_t
+vdev_mirror_map_size(int children)
+{
+ return (offsetof(mirror_map_t, mm_child[children]) +
+ sizeof (int) * children);
+}
+
+static inline mirror_map_t *
+vdev_mirror_map_alloc(int children, boolean_t replacing, boolean_t root)
+{
+ mirror_map_t *mm;
+
+ mm = kmem_zalloc(vdev_mirror_map_size(children), KM_SLEEP);
+ mm->mm_children = children;
+ mm->mm_replacing = replacing;
+ mm->mm_root = root;
+ mm->mm_preferred = (int *)((uintptr_t)mm +
+ offsetof(mirror_map_t, mm_child[children]));
+
+ return (mm);
+}
static void
vdev_mirror_map_free(zio_t *zio)
{
mirror_map_t *mm = zio->io_vsd;
- kmem_free(mm, offsetof(mirror_map_t, mm_child[mm->mm_children]));
+ kmem_free(mm, vdev_mirror_map_size(mm->mm_children));
}
static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
vdev_mirror_map_free,
zio_vsd_default_cksum_report
};
+/*
+ * Calculate and return the load of the specified vdev adjusted for a zio at
+ * the given offset.
+ *
+ * The calcuation takes into account the vdev's:
+ * 1. Rotation rate
+ * 2. The distance of zio_offset from the last queued request
+ */
+static int
+vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset)
+{
+ uint64_t lastoffset;
+ int load;
+
+ /* All DVAs have equal weight at the root. */
+ if (mm->mm_root)
+ return (INT_MAX);
+
+ /*
+ * We don't return INT_MAX if the device is resilvering i.e.
+ * vdev_resilver_txg != 0 as when tested performance was slightly
+ * worse overall when resilvering with compared to without.
+ */
+
+ /* Standard load based on pending queue length. */
+ load = vdev_queue_length(vd);
+ lastoffset = vdev_queue_last_queued_offset(vd);
+
+ if (vd->vdev_rotation_rate == VDEV_RATE_NON_ROTATING) {
+ /* Non-rotating media. */
+ if (lastoffset == zio_offset)
+ return (load + zfs_vdev_mirror_non_rotating_inc);
+
+ /*
+ * Apply a seek penalty even for non-rotating devices as
+ * sequential I/O'a can be aggregated into fewer operations
+ * on the device, thus avoiding unnecessary per-command
+ * overhead and boosting performance.
+ */
+ return (load + zfs_vdev_mirror_non_rotating_seek_inc);
+ }
+
+ /* Rotating media I/O's which directly follow the last I/O. */
+ if (lastoffset == zio_offset)
+ return (load + zfs_vdev_mirror_rotating_inc);
+
+ /*
+ * Apply half the seek increment to I/O's within seek offset
+ * of the last I/O queued to this vdev as they should incure less
+ * of a seek increment.
+ */
+ if (ABS(lastoffset - zio_offset) <
+ zfs_vdev_mirror_rotating_seek_offset)
+ return (load + (zfs_vdev_mirror_rotating_seek_inc / 2));
+
+ /* Apply the full seek increment to all other I/O's. */
+ return (load + zfs_vdev_mirror_rotating_seek_inc);
+}
+
+
static mirror_map_t *
-vdev_mirror_map_alloc(zio_t *zio)
+vdev_mirror_map_init(zio_t *zio)
{
mirror_map_t *mm = NULL;
mirror_child_t *mc;
vdev_t *vd = zio->io_vd;
- int c, d;
+ int c;
if (vd == NULL) {
dva_t *dva = zio->io_bp->blk_dva;
spa_t *spa = zio->io_spa;
- c = BP_GET_NDVAS(zio->io_bp);
-
- mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
- mm->mm_children = c;
- mm->mm_replacing = B_FALSE;
- mm->mm_preferred = spa_get_random(c);
- mm->mm_root = B_TRUE;
-
- /*
- * Check the other, lower-index DVAs to see if they're on
- * the same vdev as the child we picked. If they are, use
- * them since they are likely to have been allocated from
- * the primary metaslab in use at the time, and hence are
- * more likely to have locality with single-copy data.
- */
- for (c = mm->mm_preferred, d = c - 1; d >= 0; d--) {
- if (DVA_GET_VDEV(&dva[d]) == DVA_GET_VDEV(&dva[c]))
- mm->mm_preferred = d;
- }
-
+ mm = vdev_mirror_map_alloc(BP_GET_NDVAS(zio->io_bp), B_FALSE,
+ B_TRUE);
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
-
mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
}
} else {
- c = vd->vdev_children;
-
- mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
- 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_root = B_FALSE;
-
+ mm = vdev_mirror_map_alloc(vd->vdev_children,
+ (vd->vdev_ops == &vdev_replacing_ops ||
+ vd->vdev_ops == &vdev_spare_ops), B_FALSE);
for (c = 0; c < mm->mm_children; c++) {
mc = &mm->mm_child[c];
mc->mc_vd = vd->vdev_child[c];
mc->mc_offset = zio->io_offset;
}
@@ -207,54 +296,125 @@
mc->mc_tried = 1;
mc->mc_skipped = 0;
}
/*
- * Try to find a child whose DTL doesn't contain the block we want to read.
+ * Check the other, lower-index DVAs to see if they're on the same
+ * vdev as the child we picked. If they are, use them since they
+ * are likely to have been allocated from the primary metaslab in
+ * use at the time, and hence are more likely to have locality with
+ * single-copy data.
+ */
+static int
+vdev_mirror_dva_select(zio_t *zio, int preferred)
+{
+ dva_t *dva = zio->io_bp->blk_dva;
+ int c;
+
+ for (c = preferred - 1; c >= 0; c--) {
+ if (DVA_GET_VDEV(&dva[c]) == DVA_GET_VDEV(&dva[preferred]))
+ preferred = c;
+ }
+ return (preferred);
+}
+
+static int
+vdev_mirror_preferred_child_randomize(zio_t *zio)
+{
+ mirror_map_t *mm = zio->io_vsd;
+ int p;
+
+ if (mm->mm_root) {
+ p = spa_get_random(mm->mm_preferred_cnt);
+ return (vdev_mirror_dva_select(zio, mm->mm_preferred[p]));
+ }
+
+ /*
+ * To ensure we don't always favour the first matching vdev,
+ * which could lead to wear leveling issues on SSD's, we
+ * use the I/O offset as a pseudo random seed into the vdevs
+ * which have the lowest load.
+ */
+ p = (zio->io_offset >> zfs_vdev_mirror_shift) % mm->mm_preferred_cnt;
+ return (mm->mm_preferred[p]);
+}
+
+/*
+ * Try to find a vdev whose DTL doesn't contain the block we want to read
+ * prefering vdevs based on determined load.
+ *
* If we can't, try the read on any vdev we haven't already tried.
*/
static int
vdev_mirror_child_select(zio_t *zio)
{
mirror_map_t *mm = zio->io_vsd;
- mirror_child_t *mc;
uint64_t txg = zio->io_txg;
- int i, c;
+ int c, lowest_load;
ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg);
- /*
- * Try to find a child whose DTL doesn't contain the block to read.
- * If a child is known to be completely inaccessible (indicated by
- * vdev_readable() returning B_FALSE), don't even try.
- */
- for (i = 0, c = mm->mm_preferred; i < mm->mm_children; i++, c++) {
- if (c >= mm->mm_children)
- c = 0;
+ lowest_load = INT_MAX;
+ mm->mm_preferred_cnt = 0;
+ for (c = 0; c < mm->mm_children; c++) {
+ mirror_child_t *mc;
+
mc = &mm->mm_child[c];
if (mc->mc_tried || mc->mc_skipped)
continue;
+
if (!vdev_readable(mc->mc_vd)) {
mc->mc_error = SET_ERROR(ENXIO);
mc->mc_tried = 1; /* don't even try */
mc->mc_skipped = 1;
continue;
}
- if (!vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1))
- return (c);
+
+ if (vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1)) {
mc->mc_error = SET_ERROR(ESTALE);
mc->mc_skipped = 1;
mc->mc_speculative = 1;
+ continue;
+ }
+
+ mc->mc_load = vdev_mirror_load(mm, mc->mc_vd, mc->mc_offset);
+ if (mc->mc_load > lowest_load)
+ continue;
+
+ if (mc->mc_load < lowest_load) {
+ lowest_load = mc->mc_load;
+ mm->mm_preferred_cnt = 0;
+ }
+ mm->mm_preferred[mm->mm_preferred_cnt] = c;
+ mm->mm_preferred_cnt++;
+ }
+
+ if (mm->mm_preferred_cnt == 1) {
+ vdev_queue_register_last_queued_offset(
+ mm->mm_child[mm->mm_preferred[0]].mc_vd, zio);
+ return (mm->mm_preferred[0]);
+ }
+
+ if (mm->mm_preferred_cnt > 1) {
+ int c = vdev_mirror_preferred_child_randomize(zio);
+
+ vdev_queue_register_last_queued_offset(mm->mm_child[c].mc_vd,
+ zio);
+ return (c);
}
/*
* Every device is either missing or has this txg in its DTL.
* Look for any child we haven't already tried before giving up.
*/
- for (c = 0; c < mm->mm_children; c++)
- if (!mm->mm_child[c].mc_tried)
+ for (c = 0; c < mm->mm_children; c++) {
+ if (!mm->mm_child[c].mc_tried) {
+ vdev_queue_register_last_queued_offset(
+ mm->mm_child[c].mc_vd, zio);
return (c);
+ }
+ }
/*
* Every child failed. There's no place left to look.
*/
return (-1);
@@ -265,11 +425,11 @@
{
mirror_map_t *mm;
mirror_child_t *mc;
int c, children;
- mm = vdev_mirror_map_alloc(zio);
+ mm = vdev_mirror_map_init(zio);
if (zio->io_type == ZIO_TYPE_READ) {
if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_replacing) {
/*
* For scrubbing reads we need to allocate a read