Merge pull request #53332 from zhsgao/mds_overload_epochs

mds: add option mds_bal_overload_epochs

Reviewed-by: Venky Shankar <vshankar@redhat.com>

doc/cephfs/mds-config-ref.rst

@@ -34,6 +34,7 @@
 .. confval:: mds_bal_max_until
 .. confval:: mds_bal_mode
 .. confval:: mds_bal_min_rebalance
+.. confval:: mds_bal_overload_epochs
 .. confval:: mds_bal_min_start
 .. confval:: mds_bal_need_min
 .. confval:: mds_bal_need_max

src/common/options/mds.yaml.in

@@ -819,6 +819,16 @@ options:
   services:
   - mds
   with_legacy: true
+# must be overloaded for more than these epochs before we export anything
+- name: mds_bal_overload_epochs
+  type: int
+  level: dev
+  default: 2
+  services:
+  - mds
+  fmt_desc: The number of epochs the overload lasts before Ceph migrates,
+    setting it to a higher value can avoid frequent migrations caused by load fluctuations.
+  with_legacy: true
 # if we need less than this, we don't do anything
 - name: mds_bal_min_start
   type: float

src/mds/MDBalancer.cc

@@ -766,7 +766,7 @@ void MDBalancer::prep_rebalance(int beat)
       return;
     }
     // am i over long enough?
-    if (last_epoch_under && beat_epoch - last_epoch_under < 2) {
+    if (last_epoch_under && beat_epoch - last_epoch_under < g_conf()->mds_bal_overload_epochs) {
       dout(7) << "  i am overloaded, but only for " << (beat_epoch - last_epoch_under) << " epochs" << dendl;
       return;
     }