diff options
Diffstat (limited to 'src/mds/MDBalancer.cc')
| -rw-r--r-- | src/mds/MDBalancer.cc | 1456 |
1 files changed, 1456 insertions, 0 deletions
diff --git a/src/mds/MDBalancer.cc b/src/mds/MDBalancer.cc new file mode 100644 index 00000000..b6f6bbc1 --- /dev/null +++ b/src/mds/MDBalancer.cc @@ -0,0 +1,1456 @@ +// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- +// vim: ts=8 sw=2 smarttab +/* + * Ceph - scalable distributed file system + * + * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net> + * + * This is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License version 2.1, as published by the Free Software + * Foundation. See file COPYING. + * + */ + +#include "include/compat.h" +#include "mdstypes.h" + +#include "mon/MonClient.h" +#include "MDBalancer.h" +#include "MDSRank.h" +#include "MDSMap.h" +#include "CInode.h" +#include "CDir.h" +#include "MDCache.h" +#include "Migrator.h" +#include "Mantle.h" + +#include "include/Context.h" +#include "msg/Messenger.h" + +#include <fstream> +#include <iostream> +#include <vector> +#include <map> +using std::map; +using std::vector; +using std::chrono::duration_cast; + +#include "common/config.h" +#include "common/errno.h" + +#define dout_context g_ceph_context +#undef dout_prefix +#define dout_prefix *_dout << "mds." << mds->get_nodeid() << ".bal " +#undef dout +#define dout(lvl) \ + do {\ + auto subsys = ceph_subsys_mds;\ + if ((dout_context)->_conf->subsys.should_gather(ceph_subsys_mds_balancer, lvl)) {\ + subsys = ceph_subsys_mds_balancer;\ + }\ + dout_impl(dout_context, ceph::dout::need_dynamic(subsys), lvl) dout_prefix +#undef dendl +#define dendl dendl_impl; } while (0) + + +#define MIN_LOAD 50 // ?? +#define MIN_REEXPORT 5 // will automatically reexport +#define MIN_OFFLOAD 10 // point at which i stop trying, close enough + + +int MDBalancer::proc_message(const Message::const_ref &m) +{ + switch (m->get_type()) { + + case MSG_MDS_HEARTBEAT: + handle_heartbeat(MHeartbeat::msgref_cast(m)); + break; + + default: + derr << " balancer unknown message " << m->get_type() << dendl_impl; + ceph_abort_msg("balancer unknown message"); + } + + return 0; +} + +MDBalancer::MDBalancer(MDSRank *m, Messenger *msgr, MonClient *monc) : + mds(m), messenger(msgr), mon_client(monc) +{ + bal_fragment_dirs = g_conf().get_val<bool>("mds_bal_fragment_dirs"); + bal_fragment_interval = g_conf().get_val<int64_t>("mds_bal_fragment_interval"); +} + +void MDBalancer::handle_conf_change(const std::set<std::string>& changed, const MDSMap& mds_map) +{ + if (changed.count("mds_bal_fragment_dirs")) + bal_fragment_dirs = g_conf().get_val<bool>("mds_bal_fragment_dirs"); + if (changed.count("mds_bal_fragment_interval")) + bal_fragment_interval = g_conf().get_val<int64_t>("mds_bal_fragment_interval"); +} + +void MDBalancer::handle_export_pins(void) +{ + auto &q = mds->mdcache->export_pin_queue; + auto it = q.begin(); + dout(20) << "export_pin_queue size=" << q.size() << dendl; + while (it != q.end()) { + auto cur = it++; + CInode *in = *cur; + ceph_assert(in->is_dir()); + mds_rank_t export_pin = in->get_export_pin(false); + if (export_pin >= mds->mdsmap->get_max_mds()) { + dout(20) << " delay export pin on " << *in << dendl; + in->state_clear(CInode::STATE_QUEUEDEXPORTPIN); + q.erase(cur); + + in->state_set(CInode::STATE_DELAYEDEXPORTPIN); + mds->mdcache->export_pin_delayed_queue.insert(in); + continue; + } + + bool remove = true; + list<CDir*> dfls; + in->get_dirfrags(dfls); + for (auto dir : dfls) { + if (!dir->is_auth()) + continue; + + if (export_pin == MDS_RANK_NONE) { + if (dir->state_test(CDir::STATE_AUXSUBTREE)) { + if (dir->is_frozen() || dir->is_freezing()) { + // try again later + remove = false; + continue; + } + dout(10) << " clear auxsubtree on " << *dir << dendl; + dir->state_clear(CDir::STATE_AUXSUBTREE); + mds->mdcache->try_subtree_merge(dir); + } + } else if (export_pin == mds->get_nodeid()) { + if (dir->state_test(CDir::STATE_CREATING) || + dir->is_frozen() || dir->is_freezing()) { + // try again later + remove = false; + continue; + } + if (!dir->is_subtree_root()) { + dir->state_set(CDir::STATE_AUXSUBTREE); + mds->mdcache->adjust_subtree_auth(dir, mds->get_nodeid()); + dout(10) << " create aux subtree on " << *dir << dendl; + } else if (!dir->state_test(CDir::STATE_AUXSUBTREE)) { + dout(10) << " set auxsubtree bit on " << *dir << dendl; + dir->state_set(CDir::STATE_AUXSUBTREE); + } + } else { + mds->mdcache->migrator->export_dir(dir, export_pin); + remove = false; + } + } + + if (remove) { + in->state_clear(CInode::STATE_QUEUEDEXPORTPIN); + q.erase(cur); + } + } + + std::vector<CDir *> authsubs = mds->mdcache->get_auth_subtrees(); + bool print_auth_subtrees = true; + + if (authsubs.size() > AUTH_TREES_THRESHOLD && + !g_conf()->subsys.should_gather<ceph_subsys_mds, 25>()) { + dout(15) << "number of auth trees = " << authsubs.size() << "; not " + "printing auth trees" << dendl; + print_auth_subtrees = false; + } + + for (auto &cd : authsubs) { + mds_rank_t export_pin = cd->inode->get_export_pin(); + + if (print_auth_subtrees) { + dout(25) << "auth tree " << *cd << " export_pin=" << export_pin << + dendl; + } + + if (export_pin >= 0 && export_pin < mds->mdsmap->get_max_mds() + && export_pin != mds->get_nodeid()) { + mds->mdcache->migrator->export_dir(cd, export_pin); + } + } +} + +void MDBalancer::tick() +{ + static int num_bal_times = g_conf()->mds_bal_max; + auto bal_interval = g_conf().get_val<int64_t>("mds_bal_interval"); + auto bal_max_until = g_conf().get_val<int64_t>("mds_bal_max_until"); + time now = clock::now(); + + if (g_conf()->mds_bal_export_pin) { + handle_export_pins(); + } + + // sample? + if (chrono::duration<double>(now-last_sample).count() > + g_conf()->mds_bal_sample_interval) { + dout(15) << "tick last_sample now " << now << dendl; + last_sample = now; + } + + // We can use duration_cast below, although the result is an int, + // because the values from g_conf are also integers. + // balance? + if (mds->get_nodeid() == 0 + && mds->is_active() + && bal_interval > 0 + && duration_cast<chrono::seconds>(now - last_heartbeat).count() >= bal_interval + && (num_bal_times || (bal_max_until >= 0 && mds->get_uptime().count() > bal_max_until))) { + last_heartbeat = now; + send_heartbeat(); + num_bal_times--; + } + + mds->mdcache->show_subtrees(10, true); +} + + + + +class C_Bal_SendHeartbeat : public MDSInternalContext { +public: + explicit C_Bal_SendHeartbeat(MDSRank *mds_) : MDSInternalContext(mds_) { } + void finish(int f) override { + mds->balancer->send_heartbeat(); + } +}; + + +double mds_load_t::mds_load() const +{ + switch(g_conf()->mds_bal_mode) { + case 0: + return + .8 * auth.meta_load() + + .2 * all.meta_load() + + req_rate + + 10.0 * queue_len; + + case 1: + return req_rate + 10.0*queue_len; + + case 2: + return cpu_load_avg; + + } + ceph_abort(); + return 0; +} + +mds_load_t MDBalancer::get_load() +{ + auto now = clock::now(); + + mds_load_t load{DecayRate()}; /* zero DecayRate! */ + + if (mds->mdcache->get_root()) { + list<CDir*> ls; + mds->mdcache->get_root()->get_dirfrags(ls); + for (auto &d : ls) { + load.auth.add(d->pop_auth_subtree_nested); + load.all.add(d->pop_nested); + } + } else { + dout(20) << "get_load no root, no load" << dendl; + } + + uint64_t num_requests = mds->get_num_requests(); + + uint64_t cpu_time = 1; + { + string stat_path = PROCPREFIX "/proc/self/stat"; + ifstream stat_file(stat_path); + if (stat_file.is_open()) { + vector<string> stat_vec(std::istream_iterator<string>{stat_file}, + std::istream_iterator<string>()); + if (stat_vec.size() >= 15) { + // utime + stime + cpu_time = strtoll(stat_vec[13].c_str(), nullptr, 10) + + strtoll(stat_vec[14].c_str(), nullptr, 10); + } else { + derr << "input file '" << stat_path << "' not resolvable" << dendl_impl; + } + } else { + derr << "input file '" << stat_path << "' not found" << dendl_impl; + } + } + + load.queue_len = messenger->get_dispatch_queue_len(); + + bool update_last = true; + if (last_get_load != clock::zero() && + now > last_get_load) { + double el = std::chrono::duration<double>(now-last_get_load).count(); + if (el >= 1.0) { + if (num_requests > last_num_requests) + load.req_rate = (num_requests - last_num_requests) / el; + if (cpu_time > last_cpu_time) + load.cpu_load_avg = (cpu_time - last_cpu_time) / el; + } else { + auto p = mds_load.find(mds->get_nodeid()); + if (p != mds_load.end()) { + load.req_rate = p->second.req_rate; + load.cpu_load_avg = p->second.cpu_load_avg; + } + if (num_requests >= last_num_requests && cpu_time >= last_cpu_time) + update_last = false; + } + } + + if (update_last) { + last_num_requests = num_requests; + last_cpu_time = cpu_time; + last_get_load = now; + } + + dout(15) << "get_load " << load << dendl; + return load; +} + +/* + * Read synchronously from RADOS using a timeout. We cannot do daemon-local + * fallbacks (i.e. kick off async read when we are processing the map and + * check status when we get here) with the way the mds is structured. + */ +int MDBalancer::localize_balancer() +{ + /* reset everything */ + bool ack = false; + int r = 0; + bufferlist lua_src; + Mutex lock("lock"); + Cond cond; + + /* we assume that balancer is in the metadata pool */ + object_t oid = object_t(mds->mdsmap->get_balancer()); + object_locator_t oloc(mds->mdsmap->get_metadata_pool()); + ceph_tid_t tid = mds->objecter->read(oid, oloc, 0, 0, CEPH_NOSNAP, &lua_src, 0, + new C_SafeCond(&lock, &cond, &ack, &r)); + dout(15) << "launched non-blocking read tid=" << tid + << " oid=" << oid << " oloc=" << oloc << dendl; + + /* timeout: if we waste half our time waiting for RADOS, then abort! */ + auto bal_interval = g_conf().get_val<int64_t>("mds_bal_interval"); + lock.Lock(); + int ret_t = cond.WaitInterval(lock, utime_t(bal_interval / 2, 0)); + lock.Unlock(); + + /* success: store the balancer in memory and set the version. */ + if (!r) { + if (ret_t == ETIMEDOUT) { + mds->objecter->op_cancel(tid, -ECANCELED); + return -ETIMEDOUT; + } + bal_code.assign(lua_src.to_str()); + bal_version.assign(oid.name); + dout(10) << "localized balancer, bal_code=" << bal_code << dendl; + } + return r; +} + +void MDBalancer::send_heartbeat() +{ + if (mds->is_cluster_degraded()) { + dout(10) << "send_heartbeat degraded" << dendl; + return; + } + + if (!mds->mdcache->is_open()) { + dout(5) << "not open" << dendl; + mds->mdcache->wait_for_open(new C_Bal_SendHeartbeat(mds)); + return; + } + + if (mds->get_nodeid() == 0) { + beat_epoch++; + mds_load.clear(); + } + + // my load + mds_load_t load = get_load(); + mds->logger->set(l_mds_load_cent, 100 * load.mds_load()); + mds->logger->set(l_mds_dispatch_queue_len, load.queue_len); + + auto em = mds_load.emplace(std::piecewise_construct, std::forward_as_tuple(mds->get_nodeid()), std::forward_as_tuple(load)); + if (!em.second) { + em.first->second = load; + } + + // import_map -- how much do i import from whom + map<mds_rank_t, float> import_map; + for (auto& im : mds->mdcache->get_auth_subtrees()) { + mds_rank_t from = im->inode->authority().first; + if (from == mds->get_nodeid()) continue; + if (im->get_inode()->is_stray()) continue; + import_map[from] += im->pop_auth_subtree.meta_load(); + } + mds_import_map[ mds->get_nodeid() ] = import_map; + + + dout(5) << "mds." << mds->get_nodeid() << " epoch " << beat_epoch << " load " << load << dendl; + for (map<mds_rank_t, float>::iterator it = import_map.begin(); + it != import_map.end(); + ++it) { + dout(5) << " import_map from " << it->first << " -> " << it->second << dendl; + } + + + set<mds_rank_t> up; + mds->get_mds_map()->get_up_mds_set(up); + for (const auto& r : up) { + if (r == mds->get_nodeid()) + continue; + auto hb = MHeartbeat::create(load, beat_epoch); + hb->get_import_map() = import_map; + mds->send_message_mds(hb, r); + } +} + +void MDBalancer::handle_heartbeat(const MHeartbeat::const_ref &m) +{ + mds_rank_t who = mds_rank_t(m->get_source().num()); + dout(25) << "=== got heartbeat " << m->get_beat() << " from " << m->get_source().num() << " " << m->get_load() << dendl; + + if (!mds->is_active()) + return; + + if (!mds->mdcache->is_open()) { + dout(10) << "opening root on handle_heartbeat" << dendl; + mds->mdcache->wait_for_open(new C_MDS_RetryMessage(mds, m)); + return; + } + + if (mds->is_cluster_degraded()) { + dout(10) << " degraded, ignoring" << dendl; + return; + } + + if (mds->get_nodeid() != 0 && m->get_beat() > beat_epoch) { + dout(10) << "receive next epoch " << m->get_beat() << " from mds." << who << " before mds0" << dendl; + + beat_epoch = m->get_beat(); + // clear the mds load info whose epoch is less than beat_epoch + mds_load.clear(); + } + + if (who == 0) { + dout(20) << " from mds0, new epoch " << m->get_beat() << dendl; + if (beat_epoch != m->get_beat()) { + beat_epoch = m->get_beat(); + mds_load.clear(); + } + + send_heartbeat(); + + mds->mdcache->show_subtrees(); + } else if (mds->get_nodeid() == 0) { + if (beat_epoch != m->get_beat()) { + dout(10) << " old heartbeat epoch, ignoring" << dendl; + return; + } + } + + { + auto em = mds_load.emplace(std::piecewise_construct, std::forward_as_tuple(who), std::forward_as_tuple(m->get_load())); + if (!em.second) { + em.first->second = m->get_load(); + } + } + mds_import_map[who] = m->get_import_map(); + + { + unsigned cluster_size = mds->get_mds_map()->get_num_in_mds(); + if (mds_load.size() == cluster_size) { + // let's go! + //export_empties(); // no! + + /* avoid spamming ceph -w if user does not turn mantle on */ + if (mds->mdsmap->get_balancer() != "") { + int r = mantle_prep_rebalance(); + if (!r) return; + mds->clog->warn() << "using old balancer; mantle failed for " + << "balancer=" << mds->mdsmap->get_balancer() + << " : " << cpp_strerror(r); + } + prep_rebalance(m->get_beat()); + } + } +} + +double MDBalancer::try_match(balance_state_t& state, mds_rank_t ex, double& maxex, + mds_rank_t im, double& maxim) +{ + if (maxex <= 0 || maxim <= 0) return 0.0; + + double howmuch = std::min(maxex, maxim); + if (howmuch <= 0) return 0.0; + + dout(5) << " - mds." << ex << " exports " << howmuch << " to mds." << im << dendl; + + if (ex == mds->get_nodeid()) + state.targets[im] += howmuch; + + state.exported[ex] += howmuch; + state.imported[im] += howmuch; + + maxex -= howmuch; + maxim -= howmuch; + + return howmuch; +} + +void MDBalancer::queue_split(const CDir *dir, bool fast) +{ + dout(10) << __func__ << " enqueuing " << *dir + << " (fast=" << fast << ")" << dendl; + + const dirfrag_t frag = dir->dirfrag(); + + auto callback = [this, frag](int r) { + if (split_pending.erase(frag) == 0) { + // Someone beat me to it. This can happen in the fast splitting + // path, because we spawn two contexts, one with mds->timer and + // one with mds->queue_waiter. The loser can safely just drop + // out. + return; + } + + CDir *split_dir = mds->mdcache->get_dirfrag(frag); + if (!split_dir) { + dout(10) << "drop split on " << frag << " because not in cache" << dendl; + return; + } + if (!split_dir->is_auth()) { + dout(10) << "drop split on " << frag << " because non-auth" << dendl; + return; + } + + // Pass on to MDCache: note that the split might still not + // happen if the checks in MDCache::can_fragment fail. + dout(10) << __func__ << " splitting " << *split_dir << dendl; + mds->mdcache->split_dir(split_dir, g_conf()->mds_bal_split_bits); + }; + + bool is_new = false; + if (split_pending.count(frag) == 0) { + split_pending.insert(frag); + is_new = true; + } + + if (fast) { + // Do the split ASAP: enqueue it in the MDSRank waiters which are + // run at the end of dispatching the current request + mds->queue_waiter(new MDSInternalContextWrapper(mds, + new FunctionContext(callback))); + } else if (is_new) { + // Set a timer to really do the split: we don't do it immediately + // so that bursts of ops on a directory have a chance to go through + // before we freeze it. + mds->timer.add_event_after(bal_fragment_interval, + new FunctionContext(callback)); + } +} + +void MDBalancer::queue_merge(CDir *dir) +{ + const auto frag = dir->dirfrag(); + auto callback = [this, frag](int r) { + ceph_assert(frag.frag != frag_t()); + + // frag must be in this set because only one context is in flight + // for a given frag at a time (because merge_pending is checked before + // starting one), and this context is the only one that erases it. + merge_pending.erase(frag); + + CDir *dir = mds->mdcache->get_dirfrag(frag); + if (!dir) { + dout(10) << "drop merge on " << frag << " because not in cache" << dendl; + return; + } + ceph_assert(dir->dirfrag() == frag); + + if(!dir->is_auth()) { + dout(10) << "drop merge on " << *dir << " because lost auth" << dendl; + return; + } + + dout(10) << "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); + }; + + if (merge_pending.count(frag) == 0) { + dout(20) << __func__ << " enqueued dir " << *dir << dendl; + merge_pending.insert(frag); + mds->timer.add_event_after(bal_fragment_interval, + new FunctionContext(callback)); + } else { + dout(20) << __func__ << " dir already in queue " << *dir << dendl; + } +} + +void MDBalancer::prep_rebalance(int beat) +{ + balance_state_t state; + + if (g_conf()->mds_thrash_exports) { + //we're going to randomly export to all the mds in the cluster + set<mds_rank_t> up_mds; + mds->get_mds_map()->get_up_mds_set(up_mds); + for (const auto &rank : up_mds) { + state.targets[rank] = 0.0; + } + } else { + int cluster_size = mds->get_mds_map()->get_num_in_mds(); + mds_rank_t whoami = mds->get_nodeid(); + rebalance_time = clock::now(); + + dout(5) << " prep_rebalance: cluster loads are" << dendl; + + mds->mdcache->migrator->clear_export_queue(); + + // rescale! turn my mds_load back into meta_load units + double load_fac = 1.0; + map<mds_rank_t, mds_load_t>::iterator m = mds_load.find(whoami); + if ((m != mds_load.end()) && (m->second.mds_load() > 0)) { + double metald = m->second.auth.meta_load(); + double mdsld = m->second.mds_load(); + load_fac = metald / mdsld; + dout(7) << " load_fac is " << load_fac + << " <- " << m->second.auth << " " << metald + << " / " << mdsld + << dendl; + } + + mds_meta_load.clear(); + + double total_load = 0.0; + multimap<double,mds_rank_t> load_map; + for (mds_rank_t i=mds_rank_t(0); i < mds_rank_t(cluster_size); i++) { + mds_load_t& load = mds_load.at(i); + + double l = load.mds_load() * load_fac; + mds_meta_load[i] = l; + + if (whoami == 0) + dout(5) << " mds." << i + << " " << load + << " = " << load.mds_load() + << " ~ " << l << dendl; + + if (whoami == i) my_load = l; + total_load += l; + + load_map.insert(pair<double,mds_rank_t>( l, i )); + } + + // target load + target_load = total_load / (double)cluster_size; + dout(5) << "prep_rebalance: my load " << my_load + << " target " << target_load + << " total " << total_load + << dendl; + + // under or over? + for (auto p : load_map) { + if (p.first < target_load * (1.0 + g_conf()->mds_bal_min_rebalance)) { + dout(5) << " mds." << p.second << " is underloaded or barely overloaded." << dendl; + mds_last_epoch_under_map[p.second] = beat_epoch; + } + } + + int last_epoch_under = mds_last_epoch_under_map[whoami]; + if (last_epoch_under == beat_epoch) { + dout(5) << " i am underloaded or barely overloaded, doing nothing." << dendl; + return; + } + // am i over long enough? + if (last_epoch_under && beat_epoch - last_epoch_under < 2) { + dout(5) << " i am overloaded, but only for " << (beat_epoch - last_epoch_under) << " epochs" << dendl; + return; + } + + dout(5) << " i am sufficiently overloaded" << dendl; + + + // first separate exporters and importers + multimap<double,mds_rank_t> importers; + multimap<double,mds_rank_t> exporters; + set<mds_rank_t> importer_set; + set<mds_rank_t> exporter_set; + + for (multimap<double,mds_rank_t>::iterator it = load_map.begin(); + it != load_map.end(); + ++it) { + if (it->first < target_load) { + dout(15) << " mds." << it->second << " is importer" << dendl; + importers.insert(pair<double,mds_rank_t>(it->first,it->second)); + importer_set.insert(it->second); + } else { + int mds_last_epoch_under = mds_last_epoch_under_map[it->second]; + if (!(mds_last_epoch_under && beat_epoch - mds_last_epoch_under < 2)) { + dout(15) << " mds." << it->second << " is exporter" << dendl; + exporters.insert(pair<double,mds_rank_t>(it->first,it->second)); + exporter_set.insert(it->second); + } + } + } + + + // determine load transfer mapping + + if (true) { + // analyze import_map; do any matches i can + + dout(15) << " matching exporters to import sources" << dendl; + + // big -> small exporters + for (multimap<double,mds_rank_t>::reverse_iterator ex = exporters.rbegin(); + ex != exporters.rend(); + ++ex) { + double maxex = get_maxex(state, ex->second); + if (maxex <= .001) continue; + + // check importers. for now, just in arbitrary order (no intelligent matching). + for (map<mds_rank_t, float>::iterator im = mds_import_map[ex->second].begin(); + im != mds_import_map[ex->second].end(); + ++im) { + double maxim = get_maxim(state, im->first); + if (maxim <= .001) continue; + try_match(state, ex->second, maxex, im->first, maxim); + if (maxex <= .001) break; + } + } + } + + // old way + if (beat % 2 == 1) { + dout(15) << " matching big exporters to big importers" << dendl; + // big exporters to big importers + multimap<double,mds_rank_t>::reverse_iterator ex = exporters.rbegin(); + multimap<double,mds_rank_t>::iterator im = importers.begin(); + while (ex != exporters.rend() && + im != importers.end()) { + double maxex = get_maxex(state, ex->second); + double maxim = get_maxim(state, im->second); + if (maxex < .001 || maxim < .001) break; + try_match(state, ex->second, maxex, im->second, maxim); + if (maxex <= .001) ++ex; + if (maxim <= .001) ++im; + } + } else { // new way + dout(15) << " matching small exporters to big importers" << dendl; + // small exporters to big importers + multimap<double,mds_rank_t>::iterator ex = exporters.begin(); + multimap<double,mds_rank_t>::iterator im = importers.begin(); + while (ex != exporters.end() && + im != importers.end()) { + double maxex = get_maxex(state, ex->second); + double maxim = get_maxim(state, im->second); + if (maxex < .001 || maxim < .001) break; + try_match(state, ex->second, maxex, im->second, maxim); + if (maxex <= .001) ++ex; + if (maxim <= .001) ++im; + } + } + } + try_rebalance(state); +} + +int MDBalancer::mantle_prep_rebalance() +{ + balance_state_t state; + + /* refresh balancer if it has changed */ + if (bal_version != mds->mdsmap->get_balancer()) { + bal_version.assign(""); + int r = localize_balancer(); + if (r) return r; + + /* only spam the cluster log from 1 mds on version changes */ + if (mds->get_nodeid() == 0) + mds->clog->info() << "mantle balancer version changed: " << bal_version; + } + + /* prepare for balancing */ + int cluster_size = mds->get_mds_map()->get_num_in_mds(); + rebalance_time = clock::now(); + mds->mdcache->migrator->clear_export_queue(); + + /* fill in the metrics for each mds by grabbing load struct */ + vector < map<string, double> > metrics (cluster_size); + for (mds_rank_t i=mds_rank_t(0); i < mds_rank_t(cluster_size); i++) { + mds_load_t& load = mds_load.at(i); + + metrics[i] = {{"auth.meta_load", load.auth.meta_load()}, + {"all.meta_load", load.all.meta_load()}, + {"req_rate", load.req_rate}, + {"queue_len", load.queue_len}, + {"cpu_load_avg", load.cpu_load_avg}}; + } + + /* execute the balancer */ + Mantle mantle; + int ret = mantle.balance(bal_code, mds->get_nodeid(), metrics, state.targets); + dout(5) << " mantle decided that new targets=" << state.targets << dendl; + + /* mantle doesn't know about cluster size, so check target len here */ + if ((int) state.targets.size() != cluster_size) + return -EINVAL; + else if (ret) + return ret; + + try_rebalance(state); + return 0; +} + + + +void MDBalancer::try_rebalance(balance_state_t& state) +{ + if (g_conf()->mds_thrash_exports) { + dout(5) << "mds_thrash is on; not performing standard rebalance operation!" + << dendl; + return; + } + + // make a sorted list of my imports + multimap<double, CDir*> import_pop_map; + multimap<mds_rank_t, pair<CDir*, double> > import_from_map; + + for (auto& dir : mds->mdcache->get_fullauth_subtrees()) { + CInode *diri = dir->get_inode(); + if (diri->is_mdsdir()) + continue; + if (diri->get_export_pin(false) != MDS_RANK_NONE) + continue; + if (dir->is_freezing() || dir->is_frozen()) + continue; // export pbly already in progress + + mds_rank_t from = diri->authority().first; + double pop = dir->pop_auth_subtree.meta_load(); + if (g_conf()->mds_bal_idle_threshold > 0 && + pop < g_conf()->mds_bal_idle_threshold && + diri != mds->mdcache->get_root() && + from != mds->get_nodeid()) { + dout(5) << " exporting idle (" << pop << ") import " << *dir + << " back to mds." << from << dendl; + mds->mdcache->migrator->export_dir_nicely(dir, from); + continue; + } + + dout(15) << " map: i imported " << *dir << " from " << from << dendl; + import_pop_map.insert(make_pair(pop, dir)); + import_from_map.insert(make_pair(from, make_pair(dir, pop))); + } + + // do my exports! + map<mds_rank_t, double> export_pop_map; + + for (auto &it : state.targets) { + mds_rank_t target = it.first; + double amount = it.second; + + if (amount < MIN_OFFLOAD) + continue; + if (amount * 10 * state.targets.size() < target_load) + continue; + + dout(5) << "want to send " << amount << " to mds." << target + //<< " .. " << (*it).second << " * " << load_fac + << " -> " << amount + << dendl;//" .. fudge is " << fudge << dendl; + + double& have = export_pop_map[target]; + + mds->mdcache->show_subtrees(); + + // search imports from target + if (import_from_map.count(target)) { + dout(5) << " aha, looking through imports from target mds." << target << dendl; + for (auto p = import_from_map.equal_range(target); + p.first != p.second; ) { + CDir *dir = p.first->second.first; + double pop = p.first->second.second; + dout(5) << "considering " << *dir << " from " << (*p.first).first << dendl; + auto plast = p.first++; + + if (dir->inode->is_base()) + continue; + ceph_assert(dir->inode->authority().first == target); // cuz that's how i put it in the map, dummy + + if (pop <= amount-have) { + dout(5) << "reexporting " << *dir << " pop " << pop + << " back to mds." << target << dendl; + mds->mdcache->migrator->export_dir_nicely(dir, target); + have += pop; + import_from_map.erase(plast); + for (auto q = import_pop_map.equal_range(pop); + q.first != q.second; ) { + if (q.first->second == dir) { + import_pop_map.erase(q.first); + break; + } + q.first++; + } + } else { + dout(5) << "can't reexport " << *dir << ", too big " << pop << dendl; + } + if (amount-have < MIN_OFFLOAD) + break; + } + } + } + + // any other imports + for (auto &it : state.targets) { + mds_rank_t target = it.first; + double amount = it.second; + + if (!export_pop_map.count(target)) + continue; + double& have = export_pop_map[target]; + if (amount-have < MIN_OFFLOAD) + continue; + + for (auto p = import_pop_map.begin(); + p != import_pop_map.end(); ) { + CDir *dir = p->second; + if (dir->inode->is_base()) { + ++p; + continue; + } + + double pop = p->first; + if (pop <= amount-have && pop > MIN_REEXPORT) { + dout(0) << "reexporting " << *dir << " pop " << pop + << " to mds." << target << dendl; + have += pop; + mds->mdcache->migrator->export_dir_nicely(dir, target); + import_pop_map.erase(p++); + } else { + ++p; + } + if (amount-have < MIN_OFFLOAD) + break; + } + } + + set<CDir*> already_exporting; + + for (auto &it : state.targets) { + mds_rank_t target = it.first; + double amount = it.second; + + if (!export_pop_map.count(target)) + continue; + double& have = export_pop_map[target]; + if (amount-have < MIN_OFFLOAD) + continue; + + // okay, search for fragments of my workload + list<CDir*> exports; + + for (auto p = import_pop_map.rbegin(); + p != import_pop_map.rend(); + ++p) { + CDir *dir = p->second; + find_exports(dir, amount, exports, have, already_exporting); + if (amount-have < MIN_OFFLOAD) + break; + } + //fudge = amount - have; + + for (auto dir : exports) { + dout(5) << " - exporting " << dir->pop_auth_subtree + << " " << dir->pop_auth_subtree.meta_load() + << " to mds." << target << " " << *dir << dendl; + mds->mdcache->migrator->export_dir_nicely(dir, target); + } + } + + dout(5) << "rebalance done" << dendl; + mds->mdcache->show_subtrees(); +} + +void MDBalancer::find_exports(CDir *dir, + double amount, + list<CDir*>& exports, + double& have, + set<CDir*>& already_exporting) +{ + auto now = clock::now(); + auto duration = std::chrono::duration<double>(now-rebalance_time).count(); + if (duration > 0.1) { + derr << " balancer runs too long" << dendl_impl; + have = amount; + return; + } + + ceph_assert(dir->is_auth()); + + double need = amount - have; + if (need < amount * g_conf()->mds_bal_min_start) + return; // good enough! + + double needmax = need * g_conf()->mds_bal_need_max; + double needmin = need * g_conf()->mds_bal_need_min; + double midchunk = need * g_conf()->mds_bal_midchunk; + double minchunk = need * g_conf()->mds_bal_minchunk; + + list<CDir*> bigger_rep, bigger_unrep; + multimap<double, CDir*> smaller; + + double dir_pop = dir->pop_auth_subtree.meta_load(); + dout(7) << " find_exports in " << dir_pop << " " << *dir << " need " << need << " (" << needmin << " - " << needmax << ")" << dendl; + + double subdir_sum = 0; + for (elist<CInode*>::iterator it = dir->pop_lru_subdirs.begin_use_current(); + !it.end(); ) { + CInode *in = *it; + ++it; + + ceph_assert(in->is_dir()); + ceph_assert(in->get_parent_dir() == dir); + + list<CDir*> dfls; + in->get_nested_dirfrags(dfls); + + size_t num_idle_frags = 0; + for (list<CDir*>::iterator p = dfls.begin(); + p != dfls.end(); + ++p) { + CDir *subdir = *p; + if (already_exporting.count(subdir)) + continue; + + // we know all ancestor dirfrags up to subtree root are not freezing or frozen. + // It's more efficient to use CDir::is_{freezing,frozen}_tree_root() + if (subdir->is_frozen_dir() || subdir->is_frozen_tree_root() || + subdir->is_freezing_dir() || subdir->is_freezing_tree_root()) + continue; // can't export this right now! + + // how popular? + double pop = subdir->pop_auth_subtree.meta_load(); + subdir_sum += pop; + dout(15) << " subdir pop " << pop << " " << *subdir << dendl; + + if (pop < minchunk) { + num_idle_frags++; + continue; + } + + // lucky find? + if (pop > needmin && pop < needmax) { + exports.push_back(subdir); + already_exporting.insert(subdir); + have += pop; + return; + } + + if (pop > need) { + if (subdir->is_rep()) + bigger_rep.push_back(subdir); + else + bigger_unrep.push_back(subdir); + } else + smaller.insert(pair<double,CDir*>(pop, subdir)); + } + if (dfls.size() == num_idle_frags) + in->item_pop_lru.remove_myself(); + } + dout(15) << " sum " << subdir_sum << " / " << dir_pop << dendl; + + // grab some sufficiently big small items + multimap<double,CDir*>::reverse_iterator it; + for (it = smaller.rbegin(); + it != smaller.rend(); + ++it) { + + if ((*it).first < midchunk) + break; // try later + + dout(7) << " taking smaller " << *(*it).second << dendl; + + exports.push_back((*it).second); + already_exporting.insert((*it).second); + have += (*it).first; + if (have > needmin) + return; + } + + // apprently not enough; drill deeper into the hierarchy (if non-replicated) + for (list<CDir*>::iterator it = bigger_unrep.begin(); + it != bigger_unrep.end(); + ++it) { + dout(15) << " descending into " << **it << dendl; + find_exports(*it, amount, exports, have, already_exporting); + if (have > needmin) + return; + } + + // ok fine, use smaller bits + for (; + it != smaller.rend(); + ++it) { + dout(7) << " taking (much) smaller " << it->first << " " << *(*it).second << dendl; + + exports.push_back((*it).second); + already_exporting.insert((*it).second); + have += (*it).first; + if (have > needmin) + return; + } + + // ok fine, drill into replicated dirs + for (list<CDir*>::iterator it = bigger_rep.begin(); + it != bigger_rep.end(); + ++it) { + dout(7) << " descending into replicated " << **it << dendl; + find_exports(*it, amount, exports, have, already_exporting); + if (have > needmin) + return; + } +} + +void MDBalancer::hit_inode(CInode *in, int type, int who) +{ + // hit inode + in->pop.get(type).hit(); + + if (in->get_parent_dn()) + hit_dir(in->get_parent_dn()->get_dir(), type, who); +} + +void MDBalancer::maybe_fragment(CDir *dir, bool hot) +{ + // split/merge + if (bal_fragment_dirs && bal_fragment_interval > 0 && + dir->is_auth() && + !dir->inode->is_base() && // not root/mdsdir (for now at least) + !dir->inode->is_stray()) { // not straydir + + // split + if (g_conf()->mds_bal_split_size > 0 && (dir->should_split() || hot)) { + if (split_pending.count(dir->dirfrag()) == 0) { + queue_split(dir, false); + } else { + if (dir->should_split_fast()) { + queue_split(dir, true); + } else { + dout(10) << __func__ << ": fragment already enqueued to split: " + << *dir << dendl; + } + } + } + + // merge? + if (dir->get_frag() != frag_t() && dir->should_merge() && + merge_pending.count(dir->dirfrag()) == 0) { + queue_merge(dir); + } + } +} + +void MDBalancer::hit_dir(CDir *dir, int type, int who, double amount) +{ + // hit me + double v = dir->pop_me.get(type).hit(amount); + + const bool hot = (v > g_conf()->mds_bal_split_rd && type == META_POP_IRD) || + (v > g_conf()->mds_bal_split_wr && type == META_POP_IWR); + + dout(20) << "hit_dir " << type << " pop is " << v << ", frag " << dir->get_frag() + << " size " << dir->get_frag_size() << " " << dir->pop_me << dendl; + + maybe_fragment(dir, hot); + + // replicate? + if (type == META_POP_IRD && who >= 0) { + dir->pop_spread.hit(who); + } + + double rd_adj = 0.0; + if (type == META_POP_IRD && + dir->last_popularity_sample < last_sample) { + double dir_pop = dir->pop_auth_subtree.get(type).get(); // hmm?? + dir->last_popularity_sample = last_sample; + double pop_sp = dir->pop_spread.get(); + dir_pop += pop_sp * 10; + + //if (dir->ino() == inodeno_t(0x10000000002)) + if (pop_sp > 0) { + dout(20) << "hit_dir " << type << " pop " << dir_pop << " spread " << pop_sp + << " " << dir->pop_spread.last[0] + << " " << dir->pop_spread.last[1] + << " " << dir->pop_spread.last[2] + << " " << dir->pop_spread.last[3] + << " in " << *dir << dendl; + } + + if (dir->is_auth() && !dir->is_ambiguous_auth()) { + if (!dir->is_rep() && + dir_pop >= g_conf()->mds_bal_replicate_threshold) { + // replicate + double rdp = dir->pop_me.get(META_POP_IRD).get(); + rd_adj = rdp / mds->get_mds_map()->get_num_in_mds() - rdp; + rd_adj /= 2.0; // temper somewhat + + dout(5) << "replicating dir " << *dir << " pop " << dir_pop << " .. rdp " << rdp << " adj " << rd_adj << dendl; + + dir->dir_rep = CDir::REP_ALL; + mds->mdcache->send_dir_updates(dir, true); + + // fixme this should adjust the whole pop hierarchy + dir->pop_me.get(META_POP_IRD).adjust(rd_adj); + dir->pop_auth_subtree.get(META_POP_IRD).adjust(rd_adj); + } + + if (dir->ino() != 1 && + dir->is_rep() && + dir_pop < g_conf()->mds_bal_unreplicate_threshold) { + // unreplicate + dout(5) << "unreplicating dir " << *dir << " pop " << dir_pop << dendl; + + dir->dir_rep = CDir::REP_NONE; + mds->mdcache->send_dir_updates(dir); + } + } + } + + // adjust ancestors + bool hit_subtree = dir->is_auth(); // current auth subtree (if any) + bool hit_subtree_nested = dir->is_auth(); // all nested auth subtrees + + while (true) { + CDir *pdir = dir->inode->get_parent_dir(); + dir->pop_nested.get(type).hit(amount); + if (rd_adj != 0.0) + dir->pop_nested.get(META_POP_IRD).adjust(rd_adj); + + if (hit_subtree) { + dir->pop_auth_subtree.get(type).hit(amount); + + if (rd_adj != 0.0) + dir->pop_auth_subtree.get(META_POP_IRD).adjust(rd_adj); + + if (dir->is_subtree_root()) + hit_subtree = false; // end of auth domain, stop hitting auth counters. + else if (pdir) + pdir->pop_lru_subdirs.push_front(&dir->get_inode()->item_pop_lru); + } + + if (hit_subtree_nested) { + dir->pop_auth_subtree_nested.get(type).hit(amount); + if (rd_adj != 0.0) + dir->pop_auth_subtree_nested.get(META_POP_IRD).adjust(rd_adj); + } + if (!pdir) break; + dir = pdir; + } +} + + +/* + * subtract off an exported chunk. + * this excludes *dir itself (encode_export_dir should have take care of that) + * we _just_ do the parents' nested counters. + * + * NOTE: call me _after_ forcing *dir into a subtree root, + * but _before_ doing the encode_export_dirs. + */ +void MDBalancer::subtract_export(CDir *dir) +{ + dirfrag_load_vec_t subload = dir->pop_auth_subtree; + + while (true) { + dir = dir->inode->get_parent_dir(); + if (!dir) break; + + dir->pop_nested.sub(subload); + dir->pop_auth_subtree_nested.sub(subload); + } +} + + +void MDBalancer::add_import(CDir *dir) +{ + dirfrag_load_vec_t subload = dir->pop_auth_subtree; + + while (true) { + dir = dir->inode->get_parent_dir(); + if (!dir) break; + + dir->pop_nested.add(subload); + dir->pop_auth_subtree_nested.add(subload); + } +} + +void MDBalancer::adjust_pop_for_rename(CDir *pdir, CDir *dir, bool inc) +{ + bool adjust_subtree_nest = dir->is_auth(); + bool adjust_subtree = adjust_subtree_nest && !dir->is_subtree_root(); + CDir *cur = dir; + while (true) { + if (inc) { + pdir->pop_nested.add(dir->pop_nested); + if (adjust_subtree) { + pdir->pop_auth_subtree.add(dir->pop_auth_subtree); + pdir->pop_lru_subdirs.push_front(&cur->get_inode()->item_pop_lru); + } + + if (adjust_subtree_nest) + pdir->pop_auth_subtree_nested.add(dir->pop_auth_subtree_nested); + } else { + pdir->pop_nested.sub(dir->pop_nested); + if (adjust_subtree) + pdir->pop_auth_subtree.sub(dir->pop_auth_subtree); + + if (adjust_subtree_nest) + pdir->pop_auth_subtree_nested.sub(dir->pop_auth_subtree_nested); + } + + if (pdir->is_subtree_root()) + adjust_subtree = false; + cur = pdir; + pdir = pdir->inode->get_parent_dir(); + if (!pdir) break; + } +} + +void MDBalancer::handle_mds_failure(mds_rank_t who) +{ + if (0 == who) { + mds_last_epoch_under_map.clear(); + } +} + +int MDBalancer::dump_loads(Formatter *f) +{ + list<CDir*> dfs; + if (mds->mdcache->get_root()) { + mds->mdcache->get_root()->get_dirfrags(dfs); + } else { + dout(5) << "dump_load no root" << dendl; + } + + f->open_object_section("loads"); + + f->open_array_section("dirfrags"); + while (!dfs.empty()) { + CDir *dir = dfs.front(); + dfs.pop_front(); + + f->open_object_section("dir"); + dir->dump_load(f); + f->close_section(); + + for (auto it = dir->begin(); it != dir->end(); ++it) { + CInode *in = it->second->get_linkage()->get_inode(); + if (!in || !in->is_dir()) + continue; + + list<CDir*> ls; + in->get_dirfrags(ls); + for (auto subdir : ls) { + if (subdir->pop_nested.meta_load() < .001) + continue; + dfs.push_back(subdir); + } + } + } + f->close_section(); // dirfrags array + + f->open_object_section("mds_load"); + { + + auto dump_mds_load = [f](mds_load_t& load) { + f->dump_float("request_rate", load.req_rate); + f->dump_float("cache_hit_rate", load.cache_hit_rate); + f->dump_float("queue_length", load.queue_len); + f->dump_float("cpu_load", load.cpu_load_avg); + f->dump_float("mds_load", load.mds_load()); + + f->open_object_section("auth_dirfrags"); + load.auth.dump(f); + f->close_section(); + f->open_object_section("all_dirfrags"); + load.all.dump(f); + f->close_section(); + }; + + for (auto p : mds_load) { + stringstream name; + name << "mds." << p.first; + f->open_object_section(name.str().c_str()); + dump_mds_load(p.second); + f->close_section(); + } + } + f->close_section(); // mds_load + + f->open_object_section("mds_meta_load"); + for (auto p : mds_meta_load) { + stringstream name; + name << "mds." << p.first; + f->dump_float(name.str().c_str(), p.second); + } + f->close_section(); // mds_meta_load + + f->open_object_section("mds_import_map"); + for (auto p : mds_import_map) { + stringstream name1; + name1 << "mds." << p.first; + f->open_array_section(name1.str().c_str()); + for (auto q : p.second) { + f->open_object_section("from"); + stringstream name2; + name2 << "mds." << q.first; + f->dump_float(name2.str().c_str(), q.second); + f->close_section(); + } + f->close_section(); // mds.? array + } + f->close_section(); // mds_import_map + + f->close_section(); // loads + return 0; +} |