1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
2 // vim: ts=8 sw=2 smarttab
4 * Ceph - scalable distributed file system
6 * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
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.
15 #include "include/compat.h"
18 #include "MDBalancer.h"
20 #include "mon/MonClient.h"
28 #include "include/Context.h"
29 #include "msg/Messenger.h"
30 #include "messages/MHeartbeat.h"
39 #include "common/config.h"
40 #include "common/errno.h"
42 #define dout_context g_ceph_context
43 #define dout_subsys ceph_subsys_mds
45 #define dout_prefix *_dout << "mds." << mds->get_nodeid() << ".bal "
49 auto subsys = ceph_subsys_mds;\
50 if ((dout_context)->_conf->subsys.should_gather(ceph_subsys_mds_balancer, lvl)) {\
51 subsys = ceph_subsys_mds_balancer;\
53 dout_impl(dout_context, subsys, lvl) dout_prefix
55 #define dendl dendl_impl; } while (0)
58 #define MIN_LOAD 50 // ??
59 #define MIN_REEXPORT 5 // will automatically reexport
60 #define MIN_OFFLOAD 10 // point at which i stop trying, close enough
63 /* This function DOES put the passed message before returning */
64 int MDBalancer::proc_message(Message
*m
)
66 switch (m
->get_type()) {
68 case MSG_MDS_HEARTBEAT
:
69 handle_heartbeat(static_cast<MHeartbeat
*>(m
));
73 derr
<< " balancer unknown message " << m
->get_type() << dendl_impl
;
74 assert(0 == "balancer unknown message");
80 void MDBalancer::handle_export_pins(void)
82 auto &q
= mds
->mdcache
->export_pin_queue
;
84 dout(20) << "export_pin_queue size=" << q
.size() << dendl
;
85 while (it
!= q
.end()) {
89 mds_rank_t export_pin
= in
->get_export_pin(false);
93 in
->get_dirfrags(dfls
);
94 for (auto dir
: dfls
) {
98 if (export_pin
== MDS_RANK_NONE
) {
99 if (dir
->state_test(CDir::STATE_AUXSUBTREE
)) {
100 if (dir
->is_frozen() || dir
->is_freezing()) {
105 dout(10) << " clear auxsubtree on " << *dir
<< dendl
;
106 dir
->state_clear(CDir::STATE_AUXSUBTREE
);
107 mds
->mdcache
->try_subtree_merge(dir
);
109 } else if (export_pin
== mds
->get_nodeid()) {
110 if (dir
->state_test(CDir::STATE_CREATING
) ||
111 dir
->is_frozen() || dir
->is_freezing()) {
116 if (!dir
->is_subtree_root()) {
117 dir
->state_set(CDir::STATE_AUXSUBTREE
);
118 mds
->mdcache
->adjust_subtree_auth(dir
, mds
->get_nodeid());
119 dout(10) << " create aux subtree on " << *dir
<< dendl
;
120 } else if (!dir
->state_test(CDir::STATE_AUXSUBTREE
)) {
121 dout(10) << " set auxsubtree bit on " << *dir
<< dendl
;
122 dir
->state_set(CDir::STATE_AUXSUBTREE
);
125 mds
->mdcache
->migrator
->export_dir(dir
, export_pin
);
131 in
->state_clear(CInode::STATE_QUEUEDEXPORTPIN
);
136 set
<CDir
*> authsubs
;
137 mds
->mdcache
->get_auth_subtrees(authsubs
);
138 for (auto &cd
: authsubs
) {
139 mds_rank_t export_pin
= cd
->inode
->get_export_pin();
140 dout(10) << "auth tree " << *cd
<< " export_pin=" << export_pin
<< dendl
;
141 if (export_pin
>= 0 && export_pin
!= mds
->get_nodeid()) {
142 dout(10) << "exporting auth subtree " << *cd
->inode
<< " to " << export_pin
<< dendl
;
143 mds
->mdcache
->migrator
->export_dir(cd
, export_pin
);
148 void MDBalancer::tick()
150 static int num_bal_times
= g_conf
->mds_bal_max
;
151 static utime_t first
= ceph_clock_now();
152 utime_t now
= ceph_clock_now();
153 utime_t elapsed
= now
;
156 if (g_conf
->mds_bal_export_pin
) {
157 handle_export_pins();
161 if ((double)now
- (double)last_sample
> g_conf
->mds_bal_sample_interval
) {
162 dout(15) << "tick last_sample now " << now
<< dendl
;
167 if (last_heartbeat
== utime_t())
168 last_heartbeat
= now
;
169 if (mds
->get_nodeid() == 0 &&
170 g_conf
->mds_bal_interval
> 0 &&
172 (g_conf
->mds_bal_max_until
>= 0 &&
173 elapsed
.sec() > g_conf
->mds_bal_max_until
)) &&
175 now
.sec() - last_heartbeat
.sec() >= g_conf
->mds_bal_interval
) {
176 last_heartbeat
= now
;
185 class C_Bal_SendHeartbeat
: public MDSInternalContext
{
187 explicit C_Bal_SendHeartbeat(MDSRank
*mds_
) : MDSInternalContext(mds_
) { }
188 void finish(int f
) override
{
189 mds
->balancer
->send_heartbeat();
194 double mds_load_t::mds_load()
196 switch(g_conf
->mds_bal_mode
) {
199 .8 * auth
.meta_load() +
200 .2 * all
.meta_load() +
205 return req_rate
+ 10.0*queue_len
;
215 mds_load_t
MDBalancer::get_load(utime_t now
)
217 mds_load_t
load(now
);
219 if (mds
->mdcache
->get_root()) {
221 mds
->mdcache
->get_root()->get_dirfrags(ls
);
222 for (list
<CDir
*>::iterator p
= ls
.begin();
225 load
.auth
.add(now
, mds
->mdcache
->decayrate
, (*p
)->pop_auth_subtree_nested
);
226 load
.all
.add(now
, mds
->mdcache
->decayrate
, (*p
)->pop_nested
);
229 dout(20) << "get_load no root, no load" << dendl
;
232 load
.req_rate
= mds
->get_req_rate();
233 load
.queue_len
= messenger
->get_dispatch_queue_len();
235 ifstream
cpu(PROCPREFIX
"/proc/loadavg");
237 cpu
>> load
.cpu_load_avg
;
239 derr
<< "input file " PROCPREFIX
"'/proc/loadavg' not found" << dendl_impl
;
241 dout(15) << "get_load " << load
<< dendl
;
246 * Read synchronously from RADOS using a timeout. We cannot do daemon-local
247 * fallbacks (i.e. kick off async read when we are processing the map and
248 * check status when we get here) with the way the mds is structured.
250 int MDBalancer::localize_balancer()
252 /* reset everything */
259 /* we assume that balancer is in the metadata pool */
260 object_t oid
= object_t(mds
->mdsmap
->get_balancer());
261 object_locator_t
oloc(mds
->mdsmap
->get_metadata_pool());
262 ceph_tid_t tid
= mds
->objecter
->read(oid
, oloc
, 0, 0, CEPH_NOSNAP
, &lua_src
, 0,
263 new C_SafeCond(&lock
, &cond
, &ack
, &r
));
264 dout(15) << "launched non-blocking read tid=" << tid
265 << " oid=" << oid
<< " oloc=" << oloc
<< dendl
;
267 /* timeout: if we waste half our time waiting for RADOS, then abort! */
268 double t
= ceph_clock_now() + g_conf
->mds_bal_interval
/2;
270 timeout
.set_from_double(t
);
272 int ret_t
= cond
.WaitUntil(lock
, timeout
);
275 /* success: store the balancer in memory and set the version. */
277 if (ret_t
== ETIMEDOUT
) {
278 mds
->objecter
->op_cancel(tid
, -ECANCELED
);
281 bal_code
.assign(lua_src
.to_str());
282 bal_version
.assign(oid
.name
);
283 dout(10) << "localized balancer, bal_code=" << bal_code
<< dendl
;
288 void MDBalancer::send_heartbeat()
290 utime_t now
= ceph_clock_now();
292 if (mds
->is_cluster_degraded()) {
293 dout(10) << "send_heartbeat degraded" << dendl
;
297 if (!mds
->mdcache
->is_open()) {
298 dout(5) << "not open" << dendl
;
299 mds
->mdcache
->wait_for_open(new C_Bal_SendHeartbeat(mds
));
303 if (mds
->get_nodeid() == 0) {
310 mds_load_t load
= get_load(now
);
311 map
<mds_rank_t
, mds_load_t
>::value_type
val(mds
->get_nodeid(), load
);
312 mds_load
.insert(val
);
314 // import_map -- how much do i import from whom
315 map
<mds_rank_t
, float> import_map
;
317 mds
->mdcache
->get_auth_subtrees(authsubs
);
318 for (set
<CDir
*>::iterator it
= authsubs
.begin();
319 it
!= authsubs
.end();
322 mds_rank_t from
= im
->inode
->authority().first
;
323 if (from
== mds
->get_nodeid()) continue;
324 if (im
->get_inode()->is_stray()) continue;
325 import_map
[from
] += im
->pop_auth_subtree
.meta_load(now
, mds
->mdcache
->decayrate
);
327 mds_import_map
[ mds
->get_nodeid() ] = import_map
;
330 dout(5) << "mds." << mds
->get_nodeid() << " epoch " << beat_epoch
<< " load " << load
<< dendl
;
331 for (map
<mds_rank_t
, float>::iterator it
= import_map
.begin();
332 it
!= import_map
.end();
334 dout(5) << " import_map from " << it
->first
<< " -> " << it
->second
<< dendl
;
339 mds
->get_mds_map()->get_up_mds_set(up
);
340 for (set
<mds_rank_t
>::iterator p
= up
.begin(); p
!= up
.end(); ++p
) {
341 if (*p
== mds
->get_nodeid())
343 MHeartbeat
*hb
= new MHeartbeat(load
, beat_epoch
);
344 hb
->get_import_map() = import_map
;
345 messenger
->send_message(hb
,
346 mds
->mdsmap
->get_inst(*p
));
350 /* This function DOES put the passed message before returning */
351 void MDBalancer::handle_heartbeat(MHeartbeat
*m
)
353 typedef map
<mds_rank_t
, mds_load_t
> mds_load_map_t
;
355 mds_rank_t who
= mds_rank_t(m
->get_source().num());
356 dout(25) << "=== got heartbeat " << m
->get_beat() << " from " << m
->get_source().num() << " " << m
->get_load() << dendl
;
358 if (!mds
->is_active())
361 if (!mds
->mdcache
->is_open()) {
362 dout(10) << "opening root on handle_heartbeat" << dendl
;
363 mds
->mdcache
->wait_for_open(new C_MDS_RetryMessage(mds
, m
));
367 if (mds
->is_cluster_degraded()) {
368 dout(10) << " degraded, ignoring" << dendl
;
372 if (mds
->get_nodeid() != 0 && m
->get_beat() > beat_epoch
) {
373 dout(10) << "receive next epoch " << m
->get_beat() << " from mds." << who
<< " before mds0" << dendl
;
375 beat_epoch
= m
->get_beat();
376 // clear the mds load info whose epoch is less than beat_epoch
381 dout(20) << " from mds0, new epoch " << m
->get_beat() << dendl
;
382 if (beat_epoch
!= m
->get_beat()) {
385 beat_epoch
= m
->get_beat();
388 mds
->mdcache
->show_subtrees();
393 mds_load_map_t::value_type
val(who
, m
->get_load());
394 pair
< mds_load_map_t::iterator
, bool > rval (mds_load
.insert(val
));
396 rval
.first
->second
= val
.second
;
399 mds_import_map
[ who
] = m
->get_import_map();
402 unsigned cluster_size
= mds
->get_mds_map()->get_num_in_mds();
403 if (mds_load
.size() == cluster_size
) {
405 //export_empties(); // no!
407 /* avoid spamming ceph -w if user does not turn mantle on */
408 if (mds
->mdsmap
->get_balancer() != "") {
409 int r
= mantle_prep_rebalance();
411 mds
->clog
->warn() << "using old balancer; mantle failed for "
412 << "balancer=" << mds
->mdsmap
->get_balancer()
413 << " : " << cpp_strerror(r
);
415 prep_rebalance(m
->get_beat());
425 void MDBalancer::export_empties()
427 dout(5) << "export_empties checking for empty imports" << dendl
;
429 std::set
<CDir
*> subtrees
;
430 mds
->mdcache
->get_fullauth_subtrees(subtrees
);
431 for (auto &dir
: subtrees
) {
432 if (dir
->is_freezing() || dir
->is_frozen())
435 if (!dir
->inode
->is_base() &&
436 !dir
->inode
->is_stray() &&
437 dir
->get_num_head_items() == 0)
438 mds
->mdcache
->migrator
->export_empty_import(dir
);
444 double MDBalancer::try_match(balance_state_t
& state
, mds_rank_t ex
, double& maxex
,
445 mds_rank_t im
, double& maxim
)
447 if (maxex
<= 0 || maxim
<= 0) return 0.0;
449 double howmuch
= MIN(maxex
, maxim
);
450 if (howmuch
<= 0) return 0.0;
452 dout(5) << " - mds." << ex
<< " exports " << howmuch
<< " to mds." << im
<< dendl
;
454 if (ex
== mds
->get_nodeid())
455 state
.targets
[im
] += howmuch
;
457 state
.exported
[ex
] += howmuch
;
458 state
.imported
[im
] += howmuch
;
466 void MDBalancer::queue_split(const CDir
*dir
, bool fast
)
468 dout(10) << __func__
<< " enqueuing " << *dir
469 << " (fast=" << fast
<< ")" << dendl
;
471 assert(mds
->mdsmap
->allows_dirfrags());
472 const dirfrag_t frag
= dir
->dirfrag();
474 auto callback
= [this, frag
](int r
) {
475 if (split_pending
.erase(frag
) == 0) {
476 // Someone beat me to it. This can happen in the fast splitting
477 // path, because we spawn two contexts, one with mds->timer and
478 // one with mds->queue_waiter. The loser can safely just drop
483 CDir
*split_dir
= mds
->mdcache
->get_dirfrag(frag
);
485 dout(10) << "drop split on " << frag
<< " because not in cache" << dendl
;
488 if (!split_dir
->is_auth()) {
489 dout(10) << "drop split on " << frag
<< " because non-auth" << dendl
;
493 // Pass on to MDCache: note that the split might still not
494 // happen if the checks in MDCache::can_fragment fail.
495 dout(10) << __func__
<< " splitting " << *split_dir
<< dendl
;
496 mds
->mdcache
->split_dir(split_dir
, g_conf
->mds_bal_split_bits
);
500 if (split_pending
.count(frag
) == 0) {
501 split_pending
.insert(frag
);
506 // Do the split ASAP: enqueue it in the MDSRank waiters which are
507 // run at the end of dispatching the current request
508 mds
->queue_waiter(new MDSInternalContextWrapper(mds
,
509 new FunctionContext(callback
)));
511 // Set a timer to really do the split: we don't do it immediately
512 // so that bursts of ops on a directory have a chance to go through
513 // before we freeze it.
514 mds
->timer
.add_event_after(g_conf
->mds_bal_fragment_interval
,
515 new FunctionContext(callback
));
519 void MDBalancer::queue_merge(CDir
*dir
)
521 const auto frag
= dir
->dirfrag();
522 auto callback
= [this, frag
](int r
) {
523 assert(frag
.frag
!= frag_t());
525 // frag must be in this set because only one context is in flight
526 // for a given frag at a time (because merge_pending is checked before
527 // starting one), and this context is the only one that erases it.
528 merge_pending
.erase(frag
);
530 CDir
*dir
= mds
->mdcache
->get_dirfrag(frag
);
532 dout(10) << "drop merge on " << frag
<< " because not in cache" << dendl
;
535 assert(dir
->dirfrag() == frag
);
537 if(!dir
->is_auth()) {
538 dout(10) << "drop merge on " << *dir
<< " because lost auth" << dendl
;
542 dout(10) << "merging " << *dir
<< dendl
;
544 CInode
*diri
= dir
->get_inode();
546 frag_t fg
= dir
->get_frag();
547 while (fg
!= frag_t()) {
548 frag_t sibfg
= fg
.get_sibling();
550 bool complete
= diri
->get_dirfrags_under(sibfg
, sibs
);
552 dout(10) << " not all sibs under " << sibfg
<< " in cache (have " << sibs
<< ")" << dendl
;
556 for (list
<CDir
*>::iterator p
= sibs
.begin(); p
!= sibs
.end(); ++p
) {
558 if (!sib
->is_auth() || !sib
->should_merge()) {
564 dout(10) << " not all sibs under " << sibfg
<< " " << sibs
<< " should_merge" << dendl
;
567 dout(10) << " all sibs under " << sibfg
<< " " << sibs
<< " should merge" << dendl
;
571 if (fg
!= dir
->get_frag())
572 mds
->mdcache
->merge_dir(diri
, fg
);
575 if (merge_pending
.count(frag
) == 0) {
576 dout(20) << __func__
<< " enqueued dir " << *dir
<< dendl
;
577 merge_pending
.insert(frag
);
578 mds
->timer
.add_event_after(g_conf
->mds_bal_fragment_interval
,
579 new FunctionContext(callback
));
581 dout(20) << __func__
<< " dir already in queue " << *dir
<< dendl
;
585 void MDBalancer::prep_rebalance(int beat
)
587 balance_state_t state
;
589 if (g_conf
->mds_thrash_exports
) {
590 //we're going to randomly export to all the mds in the cluster
591 set
<mds_rank_t
> up_mds
;
592 mds
->get_mds_map()->get_up_mds_set(up_mds
);
593 for (const auto &rank
: up_mds
) {
594 state
.targets
[rank
] = 0.0;
597 int cluster_size
= mds
->get_mds_map()->get_num_in_mds();
598 mds_rank_t whoami
= mds
->get_nodeid();
599 rebalance_time
= ceph_clock_now();
601 dout(5) << " prep_rebalance: cluster loads are" << dendl
;
603 mds
->mdcache
->migrator
->clear_export_queue();
605 // rescale! turn my mds_load back into meta_load units
606 double load_fac
= 1.0;
607 map
<mds_rank_t
, mds_load_t
>::iterator m
= mds_load
.find(whoami
);
608 if ((m
!= mds_load
.end()) && (m
->second
.mds_load() > 0)) {
609 double metald
= m
->second
.auth
.meta_load(rebalance_time
, mds
->mdcache
->decayrate
);
610 double mdsld
= m
->second
.mds_load();
611 load_fac
= metald
/ mdsld
;
612 dout(7) << " load_fac is " << load_fac
613 << " <- " << m
->second
.auth
<< " " << metald
618 double total_load
= 0.0;
619 multimap
<double,mds_rank_t
> load_map
;
620 for (mds_rank_t i
=mds_rank_t(0); i
< mds_rank_t(cluster_size
); i
++) {
621 map
<mds_rank_t
, mds_load_t
>::value_type
val(i
, mds_load_t(ceph_clock_now()));
622 std::pair
< map
<mds_rank_t
, mds_load_t
>::iterator
, bool > r(mds_load
.insert(val
));
623 mds_load_t
&load(r
.first
->second
);
625 double l
= load
.mds_load() * load_fac
;
626 mds_meta_load
[i
] = l
;
629 dout(5) << " mds." << i
631 << " = " << load
.mds_load()
632 << " ~ " << l
<< dendl
;
634 if (whoami
== i
) my_load
= l
;
637 load_map
.insert(pair
<double,mds_rank_t
>( l
, i
));
641 target_load
= total_load
/ (double)cluster_size
;
642 dout(5) << "prep_rebalance: my load " << my_load
643 << " target " << target_load
644 << " total " << total_load
648 if (my_load
< target_load
* (1.0 + g_conf
->mds_bal_min_rebalance
)) {
649 dout(5) << " i am underloaded or barely overloaded, doing nothing." << dendl
;
650 last_epoch_under
= beat_epoch
;
651 mds
->mdcache
->show_subtrees();
655 // am i over long enough?
656 if (last_epoch_under
&& beat_epoch
- last_epoch_under
< 2) {
657 dout(5) << " i am overloaded, but only for " << (beat_epoch
- last_epoch_under
) << " epochs" << dendl
;
661 dout(5) << " i am sufficiently overloaded" << dendl
;
664 // first separate exporters and importers
665 multimap
<double,mds_rank_t
> importers
;
666 multimap
<double,mds_rank_t
> exporters
;
667 set
<mds_rank_t
> importer_set
;
668 set
<mds_rank_t
> exporter_set
;
670 for (multimap
<double,mds_rank_t
>::iterator it
= load_map
.begin();
671 it
!= load_map
.end();
673 if (it
->first
< target_load
) {
674 dout(15) << " mds." << it
->second
<< " is importer" << dendl
;
675 importers
.insert(pair
<double,mds_rank_t
>(it
->first
,it
->second
));
676 importer_set
.insert(it
->second
);
678 dout(15) << " mds." << it
->second
<< " is exporter" << dendl
;
679 exporters
.insert(pair
<double,mds_rank_t
>(it
->first
,it
->second
));
680 exporter_set
.insert(it
->second
);
685 // determine load transfer mapping
688 // analyze import_map; do any matches i can
690 dout(15) << " matching exporters to import sources" << dendl
;
692 // big -> small exporters
693 for (multimap
<double,mds_rank_t
>::reverse_iterator ex
= exporters
.rbegin();
694 ex
!= exporters
.rend();
696 double maxex
= get_maxex(state
, ex
->second
);
697 if (maxex
<= .001) continue;
699 // check importers. for now, just in arbitrary order (no intelligent matching).
700 for (map
<mds_rank_t
, float>::iterator im
= mds_import_map
[ex
->second
].begin();
701 im
!= mds_import_map
[ex
->second
].end();
703 double maxim
= get_maxim(state
, im
->first
);
704 if (maxim
<= .001) continue;
705 try_match(state
, ex
->second
, maxex
, im
->first
, maxim
);
706 if (maxex
<= .001) break;
713 dout(15) << " matching big exporters to big importers" << dendl
;
714 // big exporters to big importers
715 multimap
<double,mds_rank_t
>::reverse_iterator ex
= exporters
.rbegin();
716 multimap
<double,mds_rank_t
>::iterator im
= importers
.begin();
717 while (ex
!= exporters
.rend() &&
718 im
!= importers
.end()) {
719 double maxex
= get_maxex(state
, ex
->second
);
720 double maxim
= get_maxim(state
, im
->second
);
721 if (maxex
< .001 || maxim
< .001) break;
722 try_match(state
, ex
->second
, maxex
, im
->second
, maxim
);
723 if (maxex
<= .001) ++ex
;
724 if (maxim
<= .001) ++im
;
727 dout(15) << " matching small exporters to big importers" << dendl
;
728 // small exporters to big importers
729 multimap
<double,mds_rank_t
>::iterator ex
= exporters
.begin();
730 multimap
<double,mds_rank_t
>::iterator im
= importers
.begin();
731 while (ex
!= exporters
.end() &&
732 im
!= importers
.end()) {
733 double maxex
= get_maxex(state
, ex
->second
);
734 double maxim
= get_maxim(state
, im
->second
);
735 if (maxex
< .001 || maxim
< .001) break;
736 try_match(state
, ex
->second
, maxex
, im
->second
, maxim
);
737 if (maxex
<= .001) ++ex
;
738 if (maxim
<= .001) ++im
;
742 try_rebalance(state
);
745 int MDBalancer::mantle_prep_rebalance()
747 balance_state_t state
;
749 /* refresh balancer if it has changed */
750 if (bal_version
!= mds
->mdsmap
->get_balancer()) {
751 bal_version
.assign("");
752 int r
= localize_balancer();
755 /* only spam the cluster log from 1 mds on version changes */
756 if (mds
->get_nodeid() == 0)
757 mds
->clog
->info() << "mantle balancer version changed: " << bal_version
;
760 /* prepare for balancing */
761 int cluster_size
= mds
->get_mds_map()->get_num_in_mds();
762 rebalance_time
= ceph_clock_now();
763 mds
->mdcache
->migrator
->clear_export_queue();
765 /* fill in the metrics for each mds by grabbing load struct */
766 vector
< map
<string
, double> > metrics (cluster_size
);
767 for (mds_rank_t i
=mds_rank_t(0);
768 i
< mds_rank_t(cluster_size
);
770 map
<mds_rank_t
, mds_load_t
>::value_type
val(i
, mds_load_t(ceph_clock_now()));
771 std::pair
< map
<mds_rank_t
, mds_load_t
>::iterator
, bool > r(mds_load
.insert(val
));
772 mds_load_t
&load(r
.first
->second
);
774 metrics
[i
] = {{"auth.meta_load", load
.auth
.meta_load()},
775 {"all.meta_load", load
.all
.meta_load()},
776 {"req_rate", load
.req_rate
},
777 {"queue_len", load
.queue_len
},
778 {"cpu_load_avg", load
.cpu_load_avg
}};
781 /* execute the balancer */
783 int ret
= mantle
.balance(bal_code
, mds
->get_nodeid(), metrics
, state
.targets
);
784 dout(2) << " mantle decided that new targets=" << state
.targets
<< dendl
;
786 /* mantle doesn't know about cluster size, so check target len here */
787 if ((int) state
.targets
.size() != cluster_size
)
792 try_rebalance(state
);
798 void MDBalancer::try_rebalance(balance_state_t
& state
)
800 if (g_conf
->mds_thrash_exports
) {
801 dout(5) << "mds_thrash is on; not performing standard rebalance operation!"
806 // make a sorted list of my imports
807 map
<double,CDir
*> import_pop_map
;
808 multimap
<mds_rank_t
,CDir
*> import_from_map
;
809 set
<CDir
*> fullauthsubs
;
811 mds
->mdcache
->get_fullauth_subtrees(fullauthsubs
);
812 for (set
<CDir
*>::iterator it
= fullauthsubs
.begin();
813 it
!= fullauthsubs
.end();
816 if (im
->get_inode()->is_stray()) continue;
818 double pop
= im
->pop_auth_subtree
.meta_load(rebalance_time
, mds
->mdcache
->decayrate
);
819 if (g_conf
->mds_bal_idle_threshold
> 0 &&
820 pop
< g_conf
->mds_bal_idle_threshold
&&
821 im
->inode
!= mds
->mdcache
->get_root() &&
822 im
->inode
->authority().first
!= mds
->get_nodeid()) {
823 dout(5) << " exporting idle (" << pop
<< ") import " << *im
824 << " back to mds." << im
->inode
->authority().first
826 mds
->mdcache
->migrator
->export_dir_nicely(im
, im
->inode
->authority().first
);
830 import_pop_map
[ pop
] = im
;
831 mds_rank_t from
= im
->inode
->authority().first
;
832 dout(15) << " map: i imported " << *im
<< " from " << from
<< dendl
;
833 import_from_map
.insert(pair
<mds_rank_t
,CDir
*>(from
, im
));
839 set
<CDir
*> already_exporting
;
841 for (auto &it
: state
.targets
) {
842 mds_rank_t target
= it
.first
;
843 double amount
= it
.second
;
845 if (amount
< MIN_OFFLOAD
) continue;
846 if (amount
/ target_load
< .2) continue;
848 dout(5) << "want to send " << amount
<< " to mds." << target
849 //<< " .. " << (*it).second << " * " << load_fac
851 << dendl
;//" .. fudge is " << fudge << dendl;
855 mds
->mdcache
->show_subtrees();
857 // search imports from target
858 if (import_from_map
.count(target
)) {
859 dout(5) << " aha, looking through imports from target mds." << target
<< dendl
;
860 pair
<multimap
<mds_rank_t
,CDir
*>::iterator
, multimap
<mds_rank_t
,CDir
*>::iterator
> p
=
861 import_from_map
.equal_range(target
);
862 while (p
.first
!= p
.second
) {
863 CDir
*dir
= (*p
.first
).second
;
864 dout(5) << "considering " << *dir
<< " from " << (*p
.first
).first
<< dendl
;
865 multimap
<mds_rank_t
,CDir
*>::iterator plast
= p
.first
++;
867 if (dir
->inode
->is_base() ||
868 dir
->inode
->is_stray())
870 if (dir
->is_freezing() || dir
->is_frozen()) continue; // export pbly already in progress
871 double pop
= dir
->pop_auth_subtree
.meta_load(rebalance_time
, mds
->mdcache
->decayrate
);
872 assert(dir
->inode
->authority().first
== target
); // cuz that's how i put it in the map, dummy
874 if (pop
<= amount
-have
) {
875 dout(5) << "reexporting " << *dir
877 << " back to mds." << target
<< dendl
;
878 mds
->mdcache
->migrator
->export_dir_nicely(dir
, target
);
880 import_from_map
.erase(plast
);
881 import_pop_map
.erase(pop
);
883 dout(5) << "can't reexport " << *dir
<< ", too big " << pop
<< dendl
;
885 if (amount
-have
< MIN_OFFLOAD
) break;
888 if (amount
-have
< MIN_OFFLOAD
) {
894 for (map
<double,CDir
*>::iterator import
= import_pop_map
.begin();
895 import
!= import_pop_map
.end();
897 CDir
*imp
= (*import
).second
;
898 if (imp
->inode
->is_base() ||
899 imp
->inode
->is_stray())
902 double pop
= (*import
).first
;
903 if (pop
< amount
-have
|| pop
< MIN_REEXPORT
) {
904 dout(5) << "reexporting " << *imp
906 << " back to mds." << imp
->inode
->authority()
909 mds
->mdcache
->migrator
->export_dir_nicely(imp
, imp
->inode
->authority().first
);
911 if (amount
-have
< MIN_OFFLOAD
) break;
913 if (amount
-have
< MIN_OFFLOAD
) {
914 //fudge = amount-have;
918 // okay, search for fragments of my workload
919 set
<CDir
*> candidates
;
920 mds
->mdcache
->get_fullauth_subtrees(candidates
);
924 for (set
<CDir
*>::iterator pot
= candidates
.begin();
925 pot
!= candidates
.end();
927 if ((*pot
)->get_inode()->is_stray()) continue;
928 find_exports(*pot
, amount
, exports
, have
, already_exporting
);
929 if (have
> amount
-MIN_OFFLOAD
)
932 //fudge = amount - have;
934 for (list
<CDir
*>::iterator it
= exports
.begin(); it
!= exports
.end(); ++it
) {
935 dout(5) << " - exporting "
936 << (*it
)->pop_auth_subtree
938 << (*it
)->pop_auth_subtree
.meta_load(rebalance_time
, mds
->mdcache
->decayrate
)
939 << " to mds." << target
942 mds
->mdcache
->migrator
->export_dir_nicely(*it
, target
);
946 dout(5) << "rebalance done" << dendl
;
947 mds
->mdcache
->show_subtrees();
950 void MDBalancer::find_exports(CDir
*dir
,
952 list
<CDir
*>& exports
,
954 set
<CDir
*>& already_exporting
)
956 double need
= amount
- have
;
957 if (need
< amount
* g_conf
->mds_bal_min_start
)
958 return; // good enough!
959 double needmax
= need
* g_conf
->mds_bal_need_max
;
960 double needmin
= need
* g_conf
->mds_bal_need_min
;
961 double midchunk
= need
* g_conf
->mds_bal_midchunk
;
962 double minchunk
= need
* g_conf
->mds_bal_minchunk
;
964 list
<CDir
*> bigger_rep
, bigger_unrep
;
965 multimap
<double, CDir
*> smaller
;
967 double dir_pop
= dir
->pop_auth_subtree
.meta_load(rebalance_time
, mds
->mdcache
->decayrate
);
968 dout(7) << " find_exports in " << dir_pop
<< " " << *dir
<< " need " << need
<< " (" << needmin
<< " - " << needmax
<< ")" << dendl
;
970 double subdir_sum
= 0;
971 for (auto it
= dir
->begin(); it
!= dir
->end(); ++it
) {
972 CInode
*in
= it
->second
->get_linkage()->get_inode();
974 if (!in
->is_dir()) continue;
977 in
->get_dirfrags(dfls
);
978 for (list
<CDir
*>::iterator p
= dfls
.begin();
982 if (!subdir
->is_auth()) continue;
983 if (already_exporting
.count(subdir
)) continue;
985 if (subdir
->is_frozen()) continue; // can't export this right now!
988 double pop
= subdir
->pop_auth_subtree
.meta_load(rebalance_time
, mds
->mdcache
->decayrate
);
990 dout(15) << " subdir pop " << pop
<< " " << *subdir
<< dendl
;
992 if (pop
< minchunk
) continue;
995 if (pop
> needmin
&& pop
< needmax
) {
996 exports
.push_back(subdir
);
997 already_exporting
.insert(subdir
);
1003 if (subdir
->is_rep())
1004 bigger_rep
.push_back(subdir
);
1006 bigger_unrep
.push_back(subdir
);
1008 smaller
.insert(pair
<double,CDir
*>(pop
, subdir
));
1011 dout(15) << " sum " << subdir_sum
<< " / " << dir_pop
<< dendl
;
1013 // grab some sufficiently big small items
1014 multimap
<double,CDir
*>::reverse_iterator it
;
1015 for (it
= smaller
.rbegin();
1016 it
!= smaller
.rend();
1019 if ((*it
).first
< midchunk
)
1022 dout(7) << " taking smaller " << *(*it
).second
<< dendl
;
1024 exports
.push_back((*it
).second
);
1025 already_exporting
.insert((*it
).second
);
1026 have
+= (*it
).first
;
1031 // apprently not enough; drill deeper into the hierarchy (if non-replicated)
1032 for (list
<CDir
*>::iterator it
= bigger_unrep
.begin();
1033 it
!= bigger_unrep
.end();
1035 dout(15) << " descending into " << **it
<< dendl
;
1036 find_exports(*it
, amount
, exports
, have
, already_exporting
);
1041 // ok fine, use smaller bits
1043 it
!= smaller
.rend();
1045 dout(7) << " taking (much) smaller " << it
->first
<< " " << *(*it
).second
<< dendl
;
1047 exports
.push_back((*it
).second
);
1048 already_exporting
.insert((*it
).second
);
1049 have
+= (*it
).first
;
1054 // ok fine, drill into replicated dirs
1055 for (list
<CDir
*>::iterator it
= bigger_rep
.begin();
1056 it
!= bigger_rep
.end();
1058 dout(7) << " descending into replicated " << **it
<< dendl
;
1059 find_exports(*it
, amount
, exports
, have
, already_exporting
);
1066 void MDBalancer::hit_inode(utime_t now
, CInode
*in
, int type
, int who
)
1069 in
->pop
.get(type
).hit(now
, mds
->mdcache
->decayrate
);
1071 if (in
->get_parent_dn())
1072 hit_dir(now
, in
->get_parent_dn()->get_dir(), type
, who
);
1075 void MDBalancer::maybe_fragment(CDir
*dir
, bool hot
)
1078 if (g_conf
->mds_bal_frag
&& g_conf
->mds_bal_fragment_interval
> 0 &&
1079 !dir
->inode
->is_base() && // not root/base (for now at least)
1083 if (g_conf
->mds_bal_split_size
> 0 &&
1084 mds
->mdsmap
->allows_dirfrags() &&
1085 (dir
->should_split() || hot
))
1087 if (split_pending
.count(dir
->dirfrag()) == 0) {
1088 queue_split(dir
, false);
1090 if (dir
->should_split_fast()) {
1091 queue_split(dir
, true);
1093 dout(10) << __func__
<< ": fragment already enqueued to split: "
1100 if (dir
->get_frag() != frag_t() && dir
->should_merge() &&
1101 merge_pending
.count(dir
->dirfrag()) == 0) {
1107 void MDBalancer::hit_dir(utime_t now
, CDir
*dir
, int type
, int who
, double amount
)
1110 double v
= dir
->pop_me
.get(type
).hit(now
, mds
->mdcache
->decayrate
, amount
);
1112 const bool hot
= (v
> g_conf
->mds_bal_split_rd
&& type
== META_POP_IRD
) ||
1113 (v
> g_conf
->mds_bal_split_wr
&& type
== META_POP_IWR
);
1115 dout(20) << "hit_dir " << type
<< " pop is " << v
<< ", frag " << dir
->get_frag()
1116 << " size " << dir
->get_frag_size() << dendl
;
1118 maybe_fragment(dir
, hot
);
1121 if (type
== META_POP_IRD
&& who
>= 0) {
1122 dir
->pop_spread
.hit(now
, mds
->mdcache
->decayrate
, who
);
1125 double rd_adj
= 0.0;
1126 if (type
== META_POP_IRD
&&
1127 dir
->last_popularity_sample
< last_sample
) {
1128 double dir_pop
= dir
->pop_auth_subtree
.get(type
).get(now
, mds
->mdcache
->decayrate
); // hmm??
1129 dir
->last_popularity_sample
= last_sample
;
1130 double pop_sp
= dir
->pop_spread
.get(now
, mds
->mdcache
->decayrate
);
1131 dir_pop
+= pop_sp
* 10;
1133 //if (dir->ino() == inodeno_t(0x10000000002))
1135 dout(20) << "hit_dir " << type
<< " pop " << dir_pop
<< " spread " << pop_sp
1136 << " " << dir
->pop_spread
.last
[0]
1137 << " " << dir
->pop_spread
.last
[1]
1138 << " " << dir
->pop_spread
.last
[2]
1139 << " " << dir
->pop_spread
.last
[3]
1140 << " in " << *dir
<< dendl
;
1143 if (dir
->is_auth() && !dir
->is_ambiguous_auth()) {
1144 if (!dir
->is_rep() &&
1145 dir_pop
>= g_conf
->mds_bal_replicate_threshold
) {
1147 double rdp
= dir
->pop_me
.get(META_POP_IRD
).get(now
, mds
->mdcache
->decayrate
);
1148 rd_adj
= rdp
/ mds
->get_mds_map()->get_num_in_mds() - rdp
;
1149 rd_adj
/= 2.0; // temper somewhat
1151 dout(5) << "replicating dir " << *dir
<< " pop " << dir_pop
<< " .. rdp " << rdp
<< " adj " << rd_adj
<< dendl
;
1153 dir
->dir_rep
= CDir::REP_ALL
;
1154 mds
->mdcache
->send_dir_updates(dir
, true);
1156 // fixme this should adjust the whole pop hierarchy
1157 dir
->pop_me
.get(META_POP_IRD
).adjust(rd_adj
);
1158 dir
->pop_auth_subtree
.get(META_POP_IRD
).adjust(rd_adj
);
1161 if (dir
->ino() != 1 &&
1163 dir_pop
< g_conf
->mds_bal_unreplicate_threshold
) {
1165 dout(5) << "unreplicating dir " << *dir
<< " pop " << dir_pop
<< dendl
;
1167 dir
->dir_rep
= CDir::REP_NONE
;
1168 mds
->mdcache
->send_dir_updates(dir
);
1174 bool hit_subtree
= dir
->is_auth(); // current auth subtree (if any)
1175 bool hit_subtree_nested
= dir
->is_auth(); // all nested auth subtrees
1178 dir
->pop_nested
.get(type
).hit(now
, mds
->mdcache
->decayrate
, amount
);
1180 dir
->pop_nested
.get(META_POP_IRD
).adjust(now
, mds
->mdcache
->decayrate
, rd_adj
);
1183 dir
->pop_auth_subtree
.get(type
).hit(now
, mds
->mdcache
->decayrate
, amount
);
1185 dir
->pop_auth_subtree
.get(META_POP_IRD
).adjust(now
, mds
->mdcache
->decayrate
, rd_adj
);
1188 if (hit_subtree_nested
) {
1189 dir
->pop_auth_subtree_nested
.get(type
).hit(now
, mds
->mdcache
->decayrate
, amount
);
1191 dir
->pop_auth_subtree_nested
.get(META_POP_IRD
).adjust(now
, mds
->mdcache
->decayrate
, rd_adj
);
1194 if (dir
->is_subtree_root())
1195 hit_subtree
= false; // end of auth domain, stop hitting auth counters.
1197 if (dir
->inode
->get_parent_dn() == 0) break;
1198 dir
= dir
->inode
->get_parent_dn()->get_dir();
1204 * subtract off an exported chunk.
1205 * this excludes *dir itself (encode_export_dir should have take care of that)
1206 * we _just_ do the parents' nested counters.
1208 * NOTE: call me _after_ forcing *dir into a subtree root,
1209 * but _before_ doing the encode_export_dirs.
1211 void MDBalancer::subtract_export(CDir
*dir
, utime_t now
)
1213 dirfrag_load_vec_t subload
= dir
->pop_auth_subtree
;
1216 dir
= dir
->inode
->get_parent_dir();
1219 dir
->pop_nested
.sub(now
, mds
->mdcache
->decayrate
, subload
);
1220 dir
->pop_auth_subtree_nested
.sub(now
, mds
->mdcache
->decayrate
, subload
);
1225 void MDBalancer::add_import(CDir
*dir
, utime_t now
)
1227 dirfrag_load_vec_t subload
= dir
->pop_auth_subtree
;
1230 dir
= dir
->inode
->get_parent_dir();
1233 dir
->pop_nested
.add(now
, mds
->mdcache
->decayrate
, subload
);
1234 dir
->pop_auth_subtree_nested
.add(now
, mds
->mdcache
->decayrate
, subload
);
1238 void MDBalancer::handle_mds_failure(mds_rank_t who
)
1241 last_epoch_under
= 0;