ceph/src/mds/MDBalancer.h

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3 /*
   4  * Ceph - scalable distributed file system
   5  *
   6  * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
   7  *
   8  * This is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License version 2.1, as published by the Free Software
  11  * Foundation.  See file COPYING.
  12  *
  13  */
  14
  15
  16
  17 #ifndef CEPH_MDBALANCER_H
  18 #define CEPH_MDBALANCER_H
  19
  20 #include <list>
  21 #include <map>
  22 using std::list;
  23 using std::map;
  24
  25 #include "include/types.h"
  26 #include "common/Clock.h"
  27 #include "common/Cond.h"
  28
  29 class MDSRank;
  30 class Message;
  31 class MHeartbeat;
  32 class CInode;
  33 class CDir;
  34 class Messenger;
  35 class MonClient;
  36
  37 class MDBalancer {
  38   friend class C_Bal_SendHeartbeat;
  39 public:
  40   MDBalancer(MDSRank *m, Messenger *msgr, MonClient *monc) :
  41     mds(m), messenger(msgr), mon_client(monc) { }
  42
  43   int proc_message(Message *m);
  44
  45   /**
  46    * Regularly called upkeep function.
  47    *
  48    * Sends MHeartbeat messages to the mons.
  49    */
  50   void tick();
  51
  52   void subtract_export(CDir *ex, utime_t now);
  53   void add_import(CDir *im, utime_t now);
  54   void adjust_pop_for_rename(CDir *pdir, CDir *dir, utime_t now, bool inc);
  55
  56   void hit_inode(const utime_t& now, CInode *in, int type, int who=-1);
  57   void hit_dir(const utime_t& now, CDir *dir, int type, int who=-1, double amount=1.0);
  58
  59   void queue_split(const CDir *dir, bool fast);
  60   void queue_merge(CDir *dir);
  61
  62   /**
  63    * Based on size and configuration, decide whether to issue a queue_split
  64    * or queue_merge for this CDir.
  65    *
  66    * \param hot whether the directory's temperature is enough to split it
  67    */
  68   void maybe_fragment(CDir *dir, bool hot);
  69
  70   void handle_mds_failure(mds_rank_t who);
  71
  72   int dump_loads(Formatter *f);
  73
  74 private:
  75   typedef struct {
  76     std::map<mds_rank_t, double> targets;
  77     std::map<mds_rank_t, double> imported;
  78     std::map<mds_rank_t, double> exported;
  79   } balance_state_t;
  80
  81   //set up the rebalancing targets for export and do one if the
  82   //MDSMap is up to date
  83   void prep_rebalance(int beat);
  84   int mantle_prep_rebalance();
  85
  86   void handle_export_pins(void);
  87
  88   mds_load_t get_load(utime_t now);
  89   int localize_balancer();
  90   void send_heartbeat();
  91   void handle_heartbeat(MHeartbeat *m);
  92   void find_exports(CDir *dir,
  93                     double amount,
  94                     list<CDir*>& exports,
  95                     double& have,
  96                     set<CDir*>& already_exporting);
  97
  98   double try_match(balance_state_t &state,
  99                    mds_rank_t ex, double& maxex,
 100                    mds_rank_t im, double& maxim);
 101
 102   double get_maxim(balance_state_t &state, mds_rank_t im) {
 103     return target_load - mds_meta_load[im] - state.imported[im];
 104   }
 105   double get_maxex(balance_state_t &state, mds_rank_t ex) {
 106     return mds_meta_load[ex] - target_load - state.exported[ex];
 107   }
 108
 109   /**
 110    * Try to rebalance.
 111    *
 112    * Check if the monitor has recorded the current export targets;
 113    * if it has then do the actual export. Otherwise send off our
 114    * export targets message again.
 115    */
 116   void try_rebalance(balance_state_t& state);
 117
 118   MDSRank *mds;
 119   Messenger *messenger;
 120   MonClient *mon_client;
 121   int beat_epoch = 0;
 122
 123   string bal_code;
 124   string bal_version;
 125
 126   utime_t last_heartbeat;
 127   utime_t last_sample;
 128   utime_t rebalance_time; //ensure a consistent view of load for rebalance
 129
 130   utime_t last_get_load;
 131   uint64_t last_num_requests = 0;
 132
 133   // Dirfrags which are marked to be passed on to MDCache::[split|merge]_dir
 134   // just as soon as a delayed context comes back and triggers it.
 135   // These sets just prevent us from spawning extra timer contexts for
 136   // dirfrags that already have one in flight.
 137   set<dirfrag_t>   split_pending, merge_pending;
 138
 139   // per-epoch scatter/gathered info
 140   map<mds_rank_t, mds_load_t>  mds_load;
 141   map<mds_rank_t, double>       mds_meta_load;
 142   map<mds_rank_t, map<mds_rank_t, float> > mds_import_map;
 143   map<mds_rank_t, int> mds_last_epoch_under_map;
 144
 145   // per-epoch state
 146   double my_load = 0;
 147   double target_load = 0;
 148 };
 149
 150 #endif