View Source lcnt (tools v4.0)
A runtime system Lock Profiling tool.
The lcnt
module is used to profile the internal ethread locks in the Erlang
Runtime System. With lcnt
enabled, internal counters in the runtime system are
updated each time a lock is taken. The counters stores information about the
number of acquisition tries and the number of collisions that has occurred
during the acquisition tries. The counters also record the waiting time a lock
has caused for a blocked thread when a collision has occurred.
The data produced by the lock counters will give an estimate on how well the runtime system will behave from a parallelizable view point for the scenarios tested. This tool was mainly developed to help Erlang runtime developers iron out potential and generic bottlenecks.
Locks in the emulator are named after what type of resource they protect and
where in the emulator they are initialized, those are lock 'classes'. Most of
those locks are also instantiated several times, and given unique identifiers,
to increase locking granularity. Typically an instantiated lock protects a
disjunct set of the resource, for example ets tables, processes or ports. In
other cases it protects a specific range of a resource, for example pix_lock
which protects index to process mappings, and is given a unique number within
the class. A unique lock in lcnt
is referenced by a name (class) and an
identifier: {Name, Id}
.
Some locks in the system are static and protects global resources, for example
bif_timers
and the run_queue
locks. Other locks are dynamic and not
necessarily long lived, for example process locks and ets-table locks. The
statistics data from short lived locks can be stored separately when the locks
are deleted. This behavior is by default turned off to save memory but can be
turned on via lcnt:rt_opt({copy_save, true})
. The lcnt:apply/1,2,3
functions
enables this behavior during profiling.
See Also
Summary
Convenience functions
Same as apply(Fun, [])
.
Clears the lock counters and then setups the instrumentation to save all
destroyed locks. After setup the function is called, passing the elements in
Args
as arguments. When the function returns the statistics are immediately
collected to the server. After the collection the instrumentation is returned to
its previous behavior. The result of the applied function is returned.
Same as pid(node(), Id, Serial)
.
Creates a process id with creation 0.
Same as port(node(), Id)
.
Creates a port id with creation 0.
Internal runtime lock counter controllers
Same as rt_clear(node())
.
Clear the internal counters. Same as lcnt:clear(Node)
.
Same as rt_collect(node())
.
Returns a list of raw lock counter data.
Same as rt_mask(node())
.
Refer to rt_mask/2
. for a list of valid categories. All categories are enabled
by default.
Sets the lock category mask to the given categories.
Option description
Functions
Same as clear(node())
.
Clears the internal lock statistics from the runtime system. This does not clear the data on the server only on runtime system. All counters for static locks are zeroed, all dynamic locks currently alive are zeroed and all saved locks now destroyed are removed. It also resets the duration timer.
Same as collect(node())
.
Collects lock statistics from the runtime system. The function starts a server if it is not already started. It then populates the server with lock statistics. If the server held any lock statistics data before the collect then that data is lost.
Same as conflicts([])
.
Prints a list of internal locks and its statistics.
Prints lcnt server state and generic information about collected lock statistics.
Same as inspect(Lock, [])
.
Prints a list of internal lock counters for a specific lock.
Restores previously saved data to the server.
Same as locations([])
.
Prints a list of internal lock counters by source code locations.
Saves the collected data to file.
Starts the lock profiler server. The server only act as a medium for the user
and performs filtering and printing of data collected by lcnt:collect/1
.
Stops the lock profiler server.
Swaps places on Name
and Id
space for ports and processes.
Types
-type category_atom() :: atom().
-type lock_counter_data() :: term().
-type option() :: {sort, Sort :: sort()} | {reverse, boolean()} | {locations, boolean()} | {thresholds, Thresholds :: [threshold()]} | {print, PrintOptions :: [print() | {print(), non_neg_integer()}]} | {max_locks, MaxLocks :: non_neg_integer() | none} | {combine, boolean()}.
-type print() :: colls | duration | entry | id | name | ratio | time | tries | type.
-type sort() :: colls | entry | id | name | ratio | time | tries | type.
-type threshold() :: {colls, non_neg_integer()} | {time, non_neg_integer()} | {tries, non_neg_integer()}.
Convenience functions
-spec apply(Fun) -> term() when Fun :: fun().
Same as apply(Fun, [])
.
Clears the lock counters and then setups the instrumentation to save all
destroyed locks. After setup the function is called, passing the elements in
Args
as arguments. When the function returns the statistics are immediately
collected to the server. After the collection the instrumentation is returned to
its previous behavior. The result of the applied function is returned.
Warning
This function should only be used for micro-benchmarks; it sets
copy_save
totrue
for the duration of the call, which can quickly lead to running out of memory.
-spec apply(Module, Function, Args) -> term() when Module :: module(), Function :: atom(), Args :: [term()].
Same as apply(fun() -> erlang:apply(Module, Function, Args) end)
.
Same as pid(node(), Id, Serial)
.
Creates a process id with creation 0.
Same as port(node(), Id)
.
Creates a port id with creation 0.
Internal runtime lock counter controllers
-spec rt_clear() -> ok.
Same as rt_clear(node())
.
-spec rt_clear(Node) -> ok when Node :: node().
Clear the internal counters. Same as lcnt:clear(Node)
.
-spec rt_collect() -> [lock_counter_data()].
Same as rt_collect(node())
.
-spec rt_collect(Node) -> [lock_counter_data()] when Node :: node().
Returns a list of raw lock counter data.
-spec rt_mask() -> [category_atom()].
Same as rt_mask(node())
.
-spec rt_mask(Node) -> [category_atom()] when Node :: node(); (Categories) -> ok | {error, copy_save_enabled} when Categories :: [category_atom()].
Refer to rt_mask/2
. for a list of valid categories. All categories are enabled
by default.
Same as rt_mask(node(), Categories)
.
-spec rt_mask(Node, Categories) -> ok | {error, copy_save_enabled} when Node :: node(), Categories :: [category_atom()].
Sets the lock category mask to the given categories.
This will fail if the copy_save
option is enabled; see
lcnt:rt_opt/2
.
Valid categories are:
allocator
db
(ETS tables)debug
distribution
generic
io
process
scheduler
This list is subject to change at any time, as is the category any given lock may belong to.
-spec rt_opt(Option) -> boolean() when Option :: {Type, Value :: boolean()}, Type :: copy_save | process_locks.
Same as rt_opt(node(), {Type, Value})
.
-spec rt_opt(Node, Option) -> boolean() when Node :: node(), Option :: {Type, Value :: boolean()}, Type :: copy_save | process_locks.
Option description:
{copy_save, boolean()}
- Retains the statistics of destroyed locks.
Default:false
Warning
This option will use a lot of memory when enabled, which must be reclaimed with
lcnt:rt_clear
. Note that it makes no distinction between locks that were destroyed and locks for which counting was disabled, so enabling this option will disable changes to the lock category mask.{process_locks, boolean()}
- Profile process locks, equal to addingprocess
to the lock category mask; seelcnt:rt_mask/2
Default:true
Functions
-spec clear() -> ok.
Same as clear(node())
.
-spec clear(Node) -> ok when Node :: node().
Clears the internal lock statistics from the runtime system. This does not clear the data on the server only on runtime system. All counters for static locks are zeroed, all dynamic locks currently alive are zeroed and all saved locks now destroyed are removed. It also resets the duration timer.
-spec collect() -> ok.
Same as collect(node())
.
-spec collect(Node) -> ok when Node :: node().
Collects lock statistics from the runtime system. The function starts a server if it is not already started. It then populates the server with lock statistics. If the server held any lock statistics data before the collect then that data is lost.
-spec conflicts() -> ok.
Same as conflicts([])
.
-spec conflicts(Options) -> ok when Options :: [option()].
Prints a list of internal locks and its statistics.
For option description, see lcnt:inspect/2
.
-spec information() -> ok.
Prints lcnt server state and generic information about collected lock statistics.
-spec inspect(Lock) -> ok when Lock :: Name | {Name, Id | [Id]}, Name :: atom() | pid() | port(), Id :: atom() | integer() | pid() | port().
Same as inspect(Lock, [])
.
-spec inspect(Lock, Options) -> ok when Lock :: Name | {Name, Id | [Id]}, Name :: atom() | pid() | port(), Id :: atom() | integer() | pid() | port(), Options :: [option()].
Prints a list of internal lock counters for a specific lock.
Lock Name
and Id
for ports and processes are interchangeable with the use of
lcnt:swap_pid_keys/0
and is the reason why pid/0
and
port/0
options can be used in both Name
and Id
space. Both pids and
ports are special identifiers with stripped creation and can be recreated with
lcnt:pid/2,3
and lcnt:port/1,2
.
Option description:
{combine, boolean()}
- Combine the statistics from different instances of a lock class.
Default:true
{locations, boolean()}
- Print the statistics by source file and line numbers.
Default:false
{max_locks, MaxLocks}
- Maximum number of locks printed or no limit withnone
.
Default:20
{print, PrintOptions}
- Printing options:name
- Named lock or named set of locks (classes). The same name used for initializing the lock in the VM.id
- Internal id for set of locks, not always unique. This could be table name for ets tables (db_tab), port id for ports, integer identifiers for allocators, etc.type
- Type of lock:rw_mutex
,mutex
,spinlock
,rw_spinlock
orproclock
.entry
- In combination with{locations, true}
this option prints the lock operations source file and line number entry-points along with statistics for each entry.tries
- Number of acquisitions of this lock.colls
- Number of collisions when a thread tried to acquire this lock. This is when a trylock is EBUSY, a write try on read held rw_lock, a try read on write held rw_lock, a thread tries to lock an already locked lock. (Internal states supervises this).ratio
- The ratio between the number of collisions and the number of tries (acquisitions) in percentage.time
- Accumulated waiting time for this lock. This could be greater than actual wall clock time, it is accumulated for all threads. Trylock conflicts does not accumulate time.duration
- Percentage of accumulated waiting time of wall clock time. This percentage can be higher than 100% since accumulated time is from all threads.
Default:
[name,id,tries,colls,ratio,time,duration]
{reverse, boolean()}
- Reverses the order of sorting.
Default:false
{sort, Sort}
- Column sorting orders.
Default:time
{thresholds, Thresholds}
- Filtering thresholds. Anything values above the threshold value are passed through.
Default:[{tries, 0}, {colls, 0}, {time, 0}]
-spec load(Filename) -> ok when Filename :: file:filename().
Restores previously saved data to the server.
-spec locations() -> ok.
Same as locations([])
.
-spec locations(Options) -> ok when Options :: [option()].
Prints a list of internal lock counters by source code locations.
For option description, see lcnt:inspect/2
.
-spec save(Filename) -> ok when Filename :: file:filename().
Saves the collected data to file.
-spec start() -> {ok, Pid} | {error, {already_started, Pid}} when Pid :: pid().
Starts the lock profiler server. The server only act as a medium for the user
and performs filtering and printing of data collected by lcnt:collect/1
.
-spec stop() -> ok.
Stops the lock profiler server.
-spec swap_pid_keys() -> ok.
Swaps places on Name
and Id
space for ports and processes.