mds: detect small dirs that should be merged, and merge them

Signed-off-by: Sage Weil <sage@newdream.net>

diff --git a/src/mds/CDir.h b/src/mds/CDir.h
index 5a2aae7d2d00c7006786bf2329af29b1fb6addd6..eea38f2b41210c49bf4ee40d213016091b25f9aa 100644 (file)
--- a/src/mds/CDir.h
+++ b/src/mds/CDir.h
@@ -343,6 +343,14 @@ public:
 public:
   void split(int bits, list<CDir*>& subs, list<Context*>& waiters, bool replay);
   void merge(list<CDir*>& subs, list<Context*>& waiters, bool replay);
+
+  bool should_split() {
+    return (int)get_num_head_items() > g_conf.mds_bal_split_size;
+  }
+  bool should_merge() {
+    return (int)get_num_head_items() < g_conf.mds_bal_merge_size;
+  }
+
 private:
   void steal_dentry(CDentry *dn);  // from another dir.  used by merge/split.
   void purge_stolen(list<Context*>& waiters, bool replay);

diff --git a/src/mds/MDBalancer.cc b/src/mds/MDBalancer.cc
index d4e7a7550c82f71beb1394acf23a6f34ea71f51d..da08bf1949ebf34b371c48fc2316ae0ee84f4a19 100644 (file)
--- a/src/mds/MDBalancer.cc
+++ b/src/mds/MDBalancer.cc
@@ -316,25 +316,77 @@ void MDBalancer::queue_split(CDir *dir)
 
 void MDBalancer::do_fragmenting()
 {
-  if (split_queue.empty()) {
+  if (split_queue.empty() && merge_queue.empty()) {
     dout(20) << "do_fragmenting has nothing to do" << dendl;
     return;
   }
 
-  dout(0) << "do_fragmenting " << split_queue.size() << " dirs marked for possible splitting" << dendl;
+  if (!split_queue.empty()) {
+    dout(0) << "do_fragmenting " << split_queue.size() << " dirs marked for possible splitting" << dendl;
   
-  set<dirfrag_t> q;
-  q.swap(split_queue);
-
-  for (set<dirfrag_t>::iterator i = q.begin();
-       i != q.end();
-       i++) {
-    CDir *dir = mds->mdcache->get_dirfrag(*i);
-    if (!dir) continue;
-    if (!dir->is_auth()) continue;
-
-    dout(0) << "do_fragmenting splitting " << *dir << dendl;
-    mds->mdcache->split_dir(dir, g_conf.mds_bal_split_bits);
+    set<dirfrag_t> q;
+    q.swap(split_queue);
+
+    for (set<dirfrag_t>::iterator i = q.begin();
+        i != q.end();
+        i++) {
+      CDir *dir = mds->mdcache->get_dirfrag(*i);
+      if (!dir ||
+         !dir->is_auth())
+       continue;
+      
+      dout(0) << "do_fragmenting splitting " << *dir << dendl;
+      mds->mdcache->split_dir(dir, g_conf.mds_bal_split_bits);
+    }
+  }
+
+  if (!merge_queue.empty()) {
+    dout(0) << "do_fragmenting " << merge_queue.size() << " dirs marked for possible merging" << dendl;
+  
+    set<dirfrag_t> q;
+    q.swap(merge_queue);
+
+    for (set<dirfrag_t>::iterator i = q.begin();
+        i != q.end();
+        i++) {
+      CDir *dir = mds->mdcache->get_dirfrag(*i);
+      if (!dir ||
+         !dir->is_auth() ||
+         dir->get_frag() == frag_t())  // ok who's the joker?
+       continue;
+      
+      dout(0) << "do_fragmenting merging " << *dir << dendl;
+
+      CInode *diri = dir->get_inode();
+
+      frag_t fg = dir->get_frag();
+      while (fg != frag_t()) {
+       frag_t sibfg = fg.get_sibling();
+       list<CDir*> sibs;
+       bool complete = diri->get_dirfrags_under(sibfg, sibs);
+       if (!complete) {
+         dout(10) << "  not all sibs under " << sibfg << " in cache (have " << sibs << ")" << dendl;
+         break;
+       }
+       bool all = true;
+       for (list<CDir*>::iterator p = sibs.begin(); p != sibs.end(); p++) {
+         CDir *sib = *p;
+         if (!sib->is_auth() || !sib->should_merge()) {
+           all = false;
+           break;
+         }
+       }
+       if (!all) {
+         dout(10) << "  not all sibs under " << sibfg << " " << sibs << " should_merge" << dendl;
+         break;
+       }
+       dout(10) << "  all sibs under " << sibfg << " " << sibs << " should merge" << dendl;
+       fg = fg.parent();
+      }
+
+      if (fg != dir->get_frag())
+       mds->mdcache->merge_dir(diri, fg);
+    }
   }
 }
 
@@ -909,21 +961,32 @@ void MDBalancer::hit_dir(utime_t now, CDir *dir, int type, int who, double amoun
   //if (dir->ino() == inodeno_t(0x10000000000))
   //dout(0) << "hit_dir " << type << " pop " << v << " in " << *dir << dendl;
 
-  // split?
+  // split/merge
   if (g_conf.mds_bal_frag && g_conf.mds_bal_fragment_interval > 0 &&
       !dir->inode->is_base() &&        // not root/base (for now at least)
-      dir->is_auth() &&
-      
-      ((g_conf.mds_bal_split_size > 0 &&
-       dir->get_num_head_items() > (unsigned)g_conf.mds_bal_split_size) ||
-       (v > g_conf.mds_bal_split_rd && type == META_POP_IRD) ||
-       (v > g_conf.mds_bal_split_wr && type == META_POP_IWR)) &&
-      split_queue.count(dir->dirfrag()) == 0) {
-    dout(0) << "hit_dir " << type << " pop is " << v << ", putting in split_queue: " << *dir << dendl;
-    split_queue.insert(dir->dirfrag());
-  }
-  // merge?
-  
+      dir->is_auth()) {
+
+    dout(20) << "hit_dir " << type << " pop is " << v << ", frag " << dir->get_frag()
+            << " size " << dir->get_num_head_items() << dendl;
+
+    // split
+    if (g_conf.mds_bal_split_size > 0 &&
+       ((dir->get_num_head_items() > (unsigned)g_conf.mds_bal_split_size) ||
+        (v > g_conf.mds_bal_split_rd && type == META_POP_IRD) ||
+        (v > g_conf.mds_bal_split_wr && type == META_POP_IWR)) &&
+       split_queue.count(dir->dirfrag()) == 0) {
+      dout(1) << "hit_dir " << type << " pop is " << v << ", putting in split_queue: " << *dir << dendl;
+      split_queue.insert(dir->dirfrag());
+    }
+
+    // merge?
+    if (dir->get_frag() != frag_t() &&
+       (dir->get_num_head_items() < (unsigned)g_conf.mds_bal_merge_size) &&
+       merge_queue.count(dir->dirfrag()) == 0) {
+      dout(1) << "hit_dir " << type << " pop is " << v << ", putting in merge_queue: " << *dir << dendl;
+      merge_queue.insert(dir->dirfrag());
+    }
+  }  
   
   // replicate?
   if (type == META_POP_IRD && who >= 0) {

diff --git a/src/mds/MDBalancer.h b/src/mds/MDBalancer.h
index bbf3cfc89d30c3cea86ff287ca8f6dbe633b4ef6..e4eb530b4494afc03483e91c7e0b7a9eec92dc05 100644 (file)
--- a/src/mds/MDBalancer.h
+++ b/src/mds/MDBalancer.h
@@ -48,7 +48,7 @@ class MDBalancer {
   utime_t rebalance_time; //ensure a consistent view of load for rebalance
 
   // todo
-  set<dirfrag_t>   split_queue;
+  set<dirfrag_t>   split_queue, merge_queue;
 
   // per-epoch scatter/gathered info
   map<int, mds_load_t>  mds_load;