LU-10460 osd-zfs: Add tunables to disable sync

This patch allows replacing the call to txg_wait_synced(),
which blocks waiting for a full pool sync, with a smaller
tunable delay.  This delay is intended to stand in for the time
it would have taken to synchronously write the dirty data to
the intent log.

This allows testing ZFS behaviour as if there were a low-latency
ZIL device enabled to handle sync IO operations.  Setting the
delay to zero disables sync operations on the server completely.
However, be aware that no data is guaranteed to be written to
disk if the tunables are enabled, and this patch is solely for
performance analysis.  By default the tunables are set to -1,
which leaves the system using the normal sync behaviour.

Two new tunables are introduced to control the delay, the
osd_object_sync_delay_us and osd_txg_sync_delay_us module options.
These values default to -1 which preserves the safe full sync
pool behavior.  Setting these values to zero or larger will
replace the pool sync with a delay of N microseconds.

The initial test results obtained by running sanityN test 16
(fsx) are encouraging.  If the zil_commit() time can be kept to
less than 10ms we should see a significant performance improvement.
These tests were run in a pristine centos 6.4 VM and the results
are averaged over four runs.

osd_txg_sync_delay_us     -1    -1     -1     -1      -1
osd_obj_sync_delay_us     -1     0   1000  10000  100000
--------------------------------------------------------
SanityN test 16 (secs)  24.3   7.3    7.6   10.1    34.4

Change-Id: Iff9b66888edc79a5e1585fa3ce8377be068748f2
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Darby Vicker <darby.vicker-1@nasa.gov>
Reviewed-on: https://review.whamcloud.com/7761
Tested-by: Jenkins
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: Andreas Dilger <andreas.dilger@intel.com>
Reviewed-by: Giuseppe Di Natale <dinatale2@llnl.gov>

diff --git a/lustre/osd-zfs/osd_handler.c b/lustre/osd-zfs/osd_handler.c
index 2575460..74db348 100644 (file)
--- a/lustre/osd-zfs/osd_handler.c
+++ b/lustre/osd-zfs/osd_handler.c
@@ -66,6 +66,8 @@
 
 struct lu_context_key  osd_key;
 
+static int osd_txg_sync_delay_us = -1;
+
 /* Slab for OSD object allocation */
 struct kmem_cache *osd_object_kmem;
 
@@ -315,8 +317,12 @@ static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
 
        osd_unlinked_list_emptify(env, osd, &unlinked, true);
 
-       if (sync)
-               txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
+       if (sync) {
+               if (osd_txg_sync_delay_us < 0)
+                       txg_wait_synced(dmu_objset_pool(osd->od_os), txg);
+               else
+                       udelay(osd_txg_sync_delay_us);
+       }
 
        RETURN(rc);
 }
@@ -1550,6 +1556,10 @@ extern unsigned int osd_oi_count;
 module_param(osd_oi_count, int, 0444);
 MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem.");
 
+module_param(osd_txg_sync_delay_us, int, 0644);
+MODULE_PARM_DESC(osd_txg_sync_delay_us,
+                "When zero or larger delay N usec instead of doing TXG sync");
+
 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
 MODULE_VERSION(LUSTRE_VERSION_STRING);

diff --git a/lustre/osd-zfs/osd_object.c b/lustre/osd-zfs/osd_object.c
index 508cab7..00b3f4c 100644 (file)
--- a/lustre/osd-zfs/osd_object.c
+++ b/lustre/osd-zfs/osd_object.c
@@ -62,6 +62,7 @@
 #include <sys/txg.h>
 
 char *osd_obj_tag = "osd_object";
+static int osd_object_sync_delay_us = -1;
 
 static struct dt_object_operations osd_obj_ops;
 static struct lu_object_operations osd_lu_obj_ops;
@@ -1928,8 +1929,12 @@ static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
         * support ZIL.  If the object tracked the txg that it was last
         * modified in, it could pass that txg here instead of "0".  Maybe
         * the changes are already committed, so no wait is needed at all? */
-       if (!osd->od_dt_dev.dd_rdonly)
-               txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
+       if (!osd->od_dt_dev.dd_rdonly) {
+               if (osd_object_sync_delay_us < 0)
+                       txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
+               else
+                       udelay(osd_object_sync_delay_us);
+       }
 
        RETURN(0);
 }
@@ -1989,3 +1994,7 @@ static struct dt_object_operations osd_obj_otable_it_ops = {
        .do_attr_get            = osd_otable_it_attr_get,
        .do_index_try           = osd_index_try,
 };
+
+module_param(osd_object_sync_delay_us, int, 0644);
+MODULE_PARM_DESC(osd_object_sync_delay_us,
+                "If zero or larger delay N usec instead of doing object sync");