5 lcnt - The Lock Profiler
Internally in the Erlang runtime system locks are used to protect resources from being updated from multiple threads in a fatal way. Locks are necessary to ensure that the runtime system works properly but it also introduces a couple of limitations. Lock contention and locking overhead.
With lock contention we mean when one thread locks a resource and another thread, or threads, tries to acquire the same resource at the same time. The lock will deny the other thread access to the resource and the thread will be blocked from continuing its execution. The second thread has to wait until the first thread has completed its access to the resource and unlocked it. The lcnt tool measures these lock conflicts.
Locks have an inherent cost in execution time and memory space. It takes time initialize, destroy, acquiring or releasing locks. To decrease lock contention it some times necessary to use finer grained locking strategies. This will usually also increase the locking overhead and hence there is a tradeoff between lock contention and overhead. In general, lock contention increases with the number of threads running concurrently. The lcnt tool does not measure locking overhead.
5.1
Enabling lock-counting
For investigation of locks in the emulator we use an internal tool called lcnt (short for lock-count). The VM needs to be compiled with this option enabled. To compile a lock-counting VM along with a normal VM, use:
cd $ERL_TOP ./configure --enable-lock-counter
Start the lock-counting VM like this:
$ERL_TOP/bin/erl -emu_type lcnt
To verify that lock counting is enabled check that [lock-counting] appears in the status text when the VM is started.
Erlang/OTP 20 [erts-9.0] [64-bit] [smp:8:8] [ds:8:8:10] [async-threads:10] [hipe] [kernel-poll:false] [lock-counting]
5.2
Getting started
Once you have a lock counting enabled VM the module lcnt can be used. The module is intended to be used from the current running nodes shell. To access remote nodes use lcnt:clear(Node) and lcnt:collect(Node).
All locks are continuously monitored and its statistics updated. Use lcnt:clear/0 to initially clear all counters before running any specific tests. This command will also reset the duration timer internally.
To retrieve lock statistics information, use lcnt:collect/0,1. The collect operation will start a lcnt server if it not already started. All collected data will be built into an Erlang term and uploaded to the server and a duration time will also be uploaded. This duration is the time between lcnt:clear/0,1 and lcnt:collect/0,1.
Once the data is collected to the server it can be filtered, sorted and printed in many different ways.
See the reference manual for a description of each function.
5.3
Example of usage
From the Erlang shell:
Erlang R13B03 (erts-5.7.4) [source] [smp:8:8] [rq:8] [async-threads:0] [hipe]
[kernel-poll:false] [lock-counting]
1> lcnt:rt_opt({copy_save, true}).
false
2> lcnt:clear(), big:bang(1000), lcnt:collect().
ok
3> lcnt:conflicts().
lock id #tries #collisions collisions [%] time [us] duration [%]
----- --- ------- ------------ --------------- ---------- -------------
alcu_allocator 50 4113692 158921 3.8632 215464 4.4962
pix_lock 256 4007140 4882 0.1218 12221 0.2550
run_queue 8 2287246 6949 0.3038 9825 0.2050
proc_main 1029 3115778 25755 0.8266 1199 0.0250
proc_msgq 1029 2467022 1910 0.0774 1048 0.0219
proc_status 1029 5708439 2435 0.0427 706 0.0147
message_pre_alloc_lock 8 2008569 134 0.0067 90 0.0019
timeofday 1 54065 8 0.0148 22 0.0005
gc_info 1 7071 7 0.0990 5 0.0001
ok
Another way to to profile a specific function is to use lcnt:apply/3 or lcnt:apply/1 which does lcnt:clear/0 before the function and lcnt:collect/0 after its invocation. This method should only be used in micro-benchmarks since it sets copy_save to true for the duration of the function call, which may cause the emulator to run out of memory if attempted under load.
Erlang R13B03 (erts-5.7.4) [source] [smp:8:8] [rq:8] [async-threads:0] [hipe]
[kernel-poll:false] [lock-counting]
1> lcnt:apply(fun() -> big:bang(1000) end).
4384.338
2> lcnt:conflicts().
lock id #tries #collisions collisions [%] time [us] duration [%]
----- --- ------- ------------ --------------- ---------- -------------
alcu_allocator 50 4117913 183091 4.4462 234232 5.1490
run_queue 8 2050398 3801 0.1854 6700 0.1473
pix_lock 256 4007080 4943 0.1234 2847 0.0626
proc_main 1028 3000178 28247 0.9415 1022 0.0225
proc_msgq 1028 2293677 1352 0.0589 545 0.0120
proc_status 1028 5258029 1744 0.0332 442 0.0097
message_pre_alloc_lock 8 2009322 147 0.0073 82 0.0018
timeofday 1 48616 9 0.0185 13 0.0003
gc_info 1 7455 12 0.1610 9 0.0002
ok
The process locks are sorted after its class like all other locks. It is convenient to look at specific processes and ports as classes. We can do this by swapping class and class identifiers with lcnt:swap_pid_keys/0.
3> lcnt:swap_pid_keys().
ok
4> lcnt:conflicts([{print, [name, tries, ratio, time]}]).
lock #tries collisions [%] time [us]
----- ------- --------------- ----------
alcu_allocator 4117913 4.4462 234232
run_queue 2050398 0.1854 6700
pix_lock 4007080 0.1234 2847
message_pre_alloc_lock 2009322 0.0073 82
<nonode@nohost.660.0> 13493 1.4452 41
<nonode@nohost.724.0> 13504 1.1404 36
<nonode@nohost.803.0> 13181 1.6235 35
<nonode@nohost.791.0> 13534 0.8202 22
<nonode@nohost.37.0> 8744 5.8326 22
<nonode@nohost.876.0> 13335 1.1174 19
<nonode@nohost.637.0> 13452 1.3678 19
<nonode@nohost.799.0> 13497 1.8745 18
<nonode@nohost.469.0> 11009 2.5343 18
<nonode@nohost.862.0> 13131 1.2566 16
<nonode@nohost.642.0> 13216 1.7327 15
<nonode@nohost.582.0> 13156 1.1098 15
<nonode@nohost.622.0> 13420 0.7303 14
<nonode@nohost.596.0> 13141 1.6437 14
<nonode@nohost.592.0> 13346 1.2064 13
<nonode@nohost.526.0> 13076 1.1701 13
ok
5.4
Example with Mnesia Transaction Benchmark
From the Erlang shell:
Erlang R13B03 (erts-5.7.4) [source] [smp:8:8] [rq:8] [async-threads:0] [hipe]
[kernel-poll:false] [lock-counting]
Eshell V5.7.4 (abort with ^G)
1> Conf=[{db_nodes, [node()]}, {driver_nodes, [node()]}, {replica_nodes, [node()]},
{n_drivers_per_node, 10}, {n_branches, 1000}, {n_accounts_per_branch, 10},
{replica_type, ram_copies}, {stop_after, 60000}, {reuse_history_id, true}].
[{db_nodes,[nonode@nohost]},
{driver_nodes,[nonode@nohost]},
{replica_nodes,[nonode@nohost]},
{n_drivers_per_node,10},
{n_branches,1000},
{n_accounts_per_branch,10},
{replica_type,ram_copies},
{stop_after,60000},
{reuse_history_id,true}]
2> mnesia_tpcb:init([{use_running_mnesia, false}|Conf]).
ignore
Initial configuring of the benchmark is done. It is time to profile the actual benchmark and Mnesia
3> lcnt:apply(fun() -> {ok,{time, Tps,_,_,_,_}} = mnesia_tpcb:run([{use_running_mnesia,
true}|Conf]), Tps/60 end).
12037.483333333334
ok
4> lcnt:swap_pid_keys().
ok
The id header represents the number of unique identifiers under a class when the option {combine, true} is used (which is on by default). It will otherwise show the specific identifier. The db_tab listing shows 722287 unique locks, it is one for each ets-table created and Mnesia creates one for each transaction.
5> lcnt:conflicts().
lock id #tries #collisions collisions [%] time [us] duration [%]
----- --- ------- ------------ --------------- ---------- -------------
alcu_allocator 50 56355118 732662 1.3001 2934747 4.8862
db_tab 722287 94513441 63203 0.0669 1958797 3.2613
timeofday 1 2701048 175854 6.5106 1746079 2.9071
pix_lock 256 24306168 163214 0.6715 918309 1.5289
run_queue 8 11813811 152637 1.2920 357040 0.5945
message_pre_alloc_lock 8 17671449 57203 0.3237 263043 0.4380
mnesia_locker 4 17477633 1618548 9.2607 97092 0.1617
mnesia_tm 4 9891408 463788 4.6888 86353 0.1438
gc_info 1 823460 628 0.0763 24826 0.0413
meta_main_tab_slot 16 41393400 7193 0.0174 11393 0.0190
<nonode@nohost.1108.0> 4 4331412 333 0.0077 7148 0.0119
timer_wheel 1 203185 30 0.0148 3108 0.0052
<nonode@nohost.1110.0> 4 4291098 210 0.0049 885 0.0015
<nonode@nohost.1114.0> 4 4294702 288 0.0067 442 0.0007
<nonode@nohost.1113.0> 4 4346066 235 0.0054 390 0.0006
<nonode@nohost.1106.0> 4 4348159 287 0.0066 379 0.0006
<nonode@nohost.1111.0> 4 4279309 290 0.0068 325 0.0005
<nonode@nohost.1107.0> 4 4292190 302 0.0070 315 0.0005
<nonode@nohost.1112.0> 4 4208858 265 0.0063 276 0.0005
<nonode@nohost.1109.0> 4 4377502 267 0.0061 276 0.0005
ok
The listing shows mnesia_locker, a process, has highly contended locks.
6> lcnt:inspect(mnesia_locker).
lock id #tries #collisions collisions [%] time [us] duration [%]
----- --- ------- ------------ --------------- ---------- -------------
mnesia_locker proc_msgq 5449930 59374 1.0894 69781 0.1162
mnesia_locker proc_main 4462782 1487374 33.3284 14398 0.0240
mnesia_locker proc_status 7564921 71800 0.9491 12913 0.0215
mnesia_locker proc_link 0 0 0.0000 0 0.0000
ok
Listing without class combiner.
7> lcnt:conflicts([{combine, false}, {print, [name, id, tries, ratio, time]}]).
lock id #tries collisions [%] time [us]
----- --- ------- --------------- ----------
db_tab mnesia_transient_decision 722250 3.9463 1856852
timeofday undefined 2701048 6.5106 1746079
alcu_allocator ets_alloc 7490696 2.2737 692655
alcu_allocator ets_alloc 7081771 2.3294 664522
alcu_allocator ets_alloc 7047750 2.2520 658495
alcu_allocator ets_alloc 5883537 2.3177 610869
pix_lock 58 11011355 1.1924 564808
pix_lock 60 4426484 0.7120 262490
alcu_allocator ets_alloc 1897004 2.4248 219543
message_pre_alloc_lock undefined 4211267 0.3242 128299
run_queue 3 2801555 1.3003 116792
run_queue 2 2799988 1.2700 100091
run_queue 1 2966183 1.2712 78834
mnesia_locker proc_msgq 5449930 1.0894 69781
message_pre_alloc_lock undefined 3495672 0.3262 65773
message_pre_alloc_lock undefined 4189752 0.3174 58607
mnesia_tm proc_msgq 2094144 1.7184 56361
run_queue 4 2343585 1.3115 44300
db_tab branch 1446529 0.5229 38244
gc_info undefined 823460 0.0763 24826
ok
In this scenario the lock that protects ets-table mnesia_transient_decision has spent most of its waiting for. That is 1.8 seconds in a test that run for 60 seconds. The time is also spread on eight different scheduler threads.
8> lcnt:inspect(db_tab, [{print, [name, id, tries, colls, ratio, duration]}]).
lock id #tries #collisions collisions [%] duration [%]
----- --- ------- ------------ --------------- -------------
db_tab mnesia_transient_decision 722250 28502 3.9463 3.0916
db_tab branch 1446529 7564 0.5229 0.0637
db_tab account 1464500 8203 0.5601 0.0357
db_tab teller 1464529 8110 0.5538 0.0291
db_tab history 722250 3767 0.5216 0.0232
db_tab mnesia_stats 750332 7057 0.9405 0.0180
db_tab mnesia_trans_store 61 0 0.0000 0.0000
db_tab mnesia_trans_store 61 0 0.0000 0.0000
db_tab mnesia_trans_store 53 0 0.0000 0.0000
db_tab mnesia_trans_store 53 0 0.0000 0.0000
db_tab mnesia_trans_store 53 0 0.0000 0.0000
db_tab mnesia_trans_store 53 0 0.0000 0.0000
db_tab mnesia_trans_store 53 0 0.0000 0.0000
db_tab mnesia_trans_store 53 0 0.0000 0.0000
db_tab mnesia_trans_store 53 0 0.0000 0.0000
db_tab mnesia_trans_store 53 0 0.0000 0.0000
db_tab mnesia_trans_store 53 0 0.0000 0.0000
db_tab mnesia_trans_store 53 0 0.0000 0.0000
db_tab mnesia_trans_store 53 0 0.0000 0.0000
db_tab mnesia_trans_store 53 0 0.0000 0.0000
ok
5.5
Deciphering the output
Typically high time values are bad and this is often the thing to look for. However, one should also look for high lock acquisition frequencies (#tries) since locks generate overhead and because high frequency could become problematic if they begin to have conflicts even if it is not shown in a particular test.