Committee: ISO/IEC JTC1 SC22 WG21 SG1 Concurrency
Document Number: P0261R3
Date: 2017-03-13
Authors: Lawrence Crowl
Reply To: Lawrence Crowl, Lawrence@Crowl.org
Audience: SG1 Concurrency
We propose four class templates implementing distributed atomic counters.
Introduction
Solution
General Counter
Counter Broker
Counter Array
Implementation
Questions, Answers and Alternatives
Wording
?.? Distributed counters [distctr]
?.?.1 General [distctr.general]
?.?.1.1 Synchronization [distctr.sync]
?.?.1.2 Value Computation [distctr.value]
?.?.2 Synopsis [distctr.syn]
?.?.4 Bumper concept [distctr.bumper]
?.?.5 Class template general_counter [distctr.genctr]
?.?.6 Class template counter_broker [distctr.broker]
?.?.7 Class general_counter_array [distctr.generalarray]
?.?.8 Class counter_broker_array [distctr.brokerarray]
Revision History
References
Counters are ubiquitous in computing. For multi-threaded programs, atomic integers can provide safe concurrent counting.
In many cases, counters are much more commonly incremented than read. For example, long-running multithreaded server programs may maintain many counters to aid in program diagnosis. Counters are often read only in response to a network query or program termination.
Unfortunately, the cost incrementing an atomic integer is significantly greater than reading one, even when reads are rare relative to increments. As the program scales, the cost of incrementing rarely-read atomic integers can limit program performance.
We need a mechanism that minimizes the cost of incrementing a counter, even if it has increased costs to read the counter.
Cilk adder-reducers [Cilk-Reducer] are one such mechanism. Unfortunately, the reduction is control-dependent and hence not suitable for external inspection of the count.
We propose a counter that has a complexity similar to atomic integers in the simple case, but enables programmers to easily distribute that counter across multiple threads when scalability becomes a concern.
This proposal restricts itself to precise counters. Statistical counters, where the read count may be slightly different from the count operations, can produce better performance [Dice-Lev-Moir-2013].
We propose two coordinated counter types that collectively provide for distributed counting. The essential component of the design is a trade-off between the cost of incrementing the counter, the cost of loading (reading) the counter value, and the "lag" between an increment and its appearance in a "read" value. That is, the read of a counter value may not reflect the most recent increments. However, no count will be lost.
These counters are parameterized by the base integer type
that maintains the count.
Avoid situations that overflow the integer,
as that may have uninterpretable or undefined behavior.
This constraint implies that counters must be sized to their use.
(For dynamic mitigation, see the exchange operation below.)
The base of the design is the concept of a counter bumper. This concept only provides for increasing or decreasing a counter. None of its operations provide any means to read the count.
The first type programmers will encounter is the general_counter.
This counter implements the bumper concept
and provides two additional operations,
load and exchange.
The load operation returns the current count.
Note that 'current' is a fuzzy notion when other threads
are concurrently incrementing the counter.
The exchange operation
replaces the current count with the given value
and returns the old value.
This operation has particular utility in long-running programs
in occasionally draining the counter
to prevent integer overflow.
general_counter<int> red_count;
void count_red( Bag bag ) {
for ( auto i: bag )
if ( i.is_red() )
++red_count;
}
int how_many_red() {
return red_count.load();
}
When contention becomes a problem,
programmers distribute the counting
with a counter_broker objects.
These will typically be thread-local objects,
but could be CPU-local objects or even stack-based objects.
general_counter<int> red_count;
thread_local counter_broker<int> thread_red( red_count );
void count_red( Bag bag )
for ( auto i: bag )
if ( i.is_red() )
++thread_red;
}
The association of counter brokers to general counter variables is an essential part of the distribution of counters, and not easily moved or copied. Therefore, we provide no copy or move operations for these types.
Managing the association between general counters and their attached brokers
can require a non-trivial amount of space and time.
Therefore, we provide a means to handle many counters with one name
with the general_counter_array.
The size of the counter arrays is fixed at construction time.
Counter arrays have the same structure as single-value counters,
with the following exceptions.
Increment a counter by first indexing into its array
and then applying one of the bumper operations.
E.g. ++ctr_ary[i].
The load and exchange operations take an additional index parameter.
Olivier Giroux suggested an implementation in which the set of attached brokers was tracked by an intrusive list through the brokers themselves, rather than through a separate set data structure. This approach has some advantages.
No dynamic allocation is required.
The constructor of the general counter can be constexpr, which avoids initialization order problems for global objects.
The head of the list is null until such time as a broker attaches itself, which substantially reduces the cost of using a general counter in the case where there are no brokers.
The benefit of a constexpr initializer is high enough so that we choose to adopt the constexpr initializer. Unfortunately, counter arrays require allocation for their array of bumpers, and so cannot have constexpr initializers.
An implementation of an earlier version of the design
(with the separate broker set data structure)
is available from the Google Concurrency Library
https://github.com/alasdairmackintosh/google-concurrency-library
at
.../blob/master/include/counter.h.
This paper's general_counter
maps to counter::strong_duplex
and
this paper's counter_broker
maps to counter::strong_broker.
No. The design point for the paper is when reads are rare releative to increments. Reference counting has closer to a one-to-one ratio.
general_counter just a wrapper around an atomic integral?
No.
The general_counter maintains the list of associated brokers
and uses that list to poll the brokers
for load and exchange operations.
counter_broker?
The type brokers the increments for the counter.
I welcome specific suggestions with rationale.
counter_brokers a needless complication?
No. The essential performance gain comes from distribution, which requires variables in different parts of the system [Vollmann-2015].
counter_brokers be better?
There is no language mechanism for automatically creating the distributed variables. Mechanisms do exist on some systems for hand-coding a distribution. The standard should no rely on narrowly implemented facilities.
Even so, these hand-coded mechanisms generally do not have the performance
of language-defined thread_local variables.
The cost of finding 'your' broker
may quickly become the performance bottleneck.
Access to variables is best integrated into the compilation system.
The current design enables end-user programmers to exploit any system-provided distribution mechanism.
When high-performance distribution mechanisms are more widely implemented and integrated into the compilation system, automatic distribution becomes more viable.
thread_local brokers.
Wouldn't CPU-local brokers be better?
Yes. However, not all platforms provide efficient access to a CPU-local variable. If your platform provides CPU-local variables, you can use brokers with them.
Yes.
This technique was used
in the weak_duplex and weak_broker types
in earlier versions of the paper.
The downside is that one cannot have the exchange operation
because the exchange would be a concurrent write to the broker.
SG1 chose to remove the weaker version
to simplify the interface.
Restartable sequences are only now have limited availability.
More importantly, they require exclusive CPU access,
or at least the ability to detect potential non-exclusive acces,
which effectively prohibits the exchange operation.
That is, restartable sequences only work with the weaker broker.
There is nothing in this proposal that precludes adding the weaker version later. Given that SG1 has a preference for a simple initial facility, adding the weaker version should be a separate paper.
Yes. Higher-performance facilities for use with automatic variables were in the earlier version of the paper. They used a 'push' model of returning counts to the main counter to eliminate atomic variable cost entirely. These facilities were removed by SG1 to reduce the complexity of the interface.
The distributed counter definition is as follows.
Add a new section.
Add a new section.
This section provides mechanisms for distributed counters. These types optimize modifying the count over reading the count. That is, reading is assumed to be relatively rare compared to modifying. These mechanisms ease the production of scalable concurrent programs.
The
general_counterprovides a single counter. When contention on that counter becomes undesireable, one can distribute the counting and representation by associating one or morecounter_brokers with thegeneral_counter. Together, they povide a single abstract counter.Every operation on a
general_counterits associatedcounter_brokeris an operation on the abstract counter.Likewise,
general_counter_arrayandcounter_broker_arrayprovides an indexed set of abstract counters. Operations on both with the same index are operations on a single abstract counter.Each abstract counter has a set M consisting of the modify operations (
+=,-=, andexchange) and a set R consisting of the read operations (loadandexchange). Theexchangeoperations exist in both sets.
Add a new section.
All read and modify operations are atomic. [Note: They will not cause a data race. —end note]
The classes described in [distctr] impose no memory ordering between increment and decrement operations and any other operations. So, the set of modify operations might not be totally ordered. [Note: Detecting a count value that indicates that an event has occured does not mean that any other memory updates associated with that event are visible to the current thread. Acting on any particular count requires additional synchronization. —end note]
Add a new section.
The set of observed modify operations Oi of a read operation Ri of an abstract counter is the union of the following sets:
- The set Pi contains the union of Oj such that Rj happens before Ri.
- The set Bi contains those operations in M that happen before Ri.
- The set Ui, is an arbitrary subset of those operations in M that neither happen after nor happen before Ri and are not Ri. [Note: The
exchangeoperation does not observe its own modification. —end note]The abstract value Vi of a read operation Ri is the sum of:
- the initial abstract value I,
- the argument to each
+=operation Mi that is in Oi,- the negative of the argument to each
-=operation Mi that is in Oi, and- the argument to each
exchangeoperation Mi minus its abstract value Vi such that the operation Mi is in Oi.The actual value is the abstract value when the abstract value never exceeds the range of the template parameter type. Otherwise, the actual value is implementation-defined.
Add a new section.
template< typename Integral > class general_counter; template< typename Integral > class counter_broker; template< typename Integral > class general_counter_array; template< typename Integral > class counter_broker_array;
Add a new section.
The bumper concept provides a common interface for increasing or decreasing the value of a counter. It consists of the following operations. [Note: These operations specifically return
voidto avoid leaking information that would defeat the purpose of optimizing for increments over loads. —end note]
void operator +=( integer )
void operator -=( integer )
- Effects:
Atomically add/subtract a value to/from the counter.
void operator ++()
void operator ++(int)
void operator --()
void operator --(int)
- Effects:
As if
counter+=1orcounter-=1, respectively.
general_counter [distctr.genctr]Add a new section.
namespace std { template< typename Integral > class general_counter { general_counter( const general_counter& ) = delete; general_counter& operator=( const general_counter& ) = delete; public: constexpr general_counter( Integral ); constexpr general_counter(); ~general_counter(); void operator +=( Integral ); void operator -=( Integral ); void operator ++(); void operator ++(int); void operator --(); void operator --(int); Integral load(); Integral exchange( Integral to ); }The
Integraltemplate type parameter shall be an integral type for whichstd::atomichas a specialization (29.5 [atomics.types.generic]).The
general_counterclass template implements the bumper concept ([distctr.bumper]).
constexpr general_counter( Integral )
constexpr general_counter()
- Effects:
Initialize the
general_counterwith the given value, or zero if no value is given.
~general_counter()
- Requirements:
All attached
counter_brokerobjects shall have been previously destroyed.- Effects:
Destroys the
general_counter.
Integer load()
- Returns:
Atomically returns the value of the
general_counter. [Note: The value will be an approximation if any threads may concurrently update thegeneral_counter. —end note]
Integer exchange( Integer )
- Effects:
Atomically replaces the value of the
general_counterwith the value of the parameter.- Returns:
The replaced value.
counter_broker [distctr.broker]Add a new section.
template< typename Integral > class counter_broker { counter_broker( const counter_broker& ) = delete; counter_broker& operator=( const counter_broker& ) = delete; public: counter_broker( general_counter<Integral>& ); ~counter_broker(); void operator +=( Integral ); void operator -=( Integral ); void operator ++(); void operator ++(int); void operator --(); void operator --(int); };The
Integraltemplate type parameter shall be an integral type for whichstd::atomichas a specialization (29.5 [atomics.types.generic]).The
counter_brokerclass template implements the bumper concept ([distctr.bumper]).
counter_broker( general_counter<Integral>& )
- Effects:
Initialize the broker, atomically associating it with the given
general_counter.
~counter_broker()
- Effects:
Atomically moves any part of the count value retained within the
counter_brokerto its associatedgeneral_counter. Atomically disassociates thecounter_brokerfrom thegeneral_counter. Destroys thecounter_broker.
general_counter_array [distctr.generalarray]Add a new section.
template< typename Integral > class general_counter_array { general_counter_array( const general_counter_array& ) = delete; general_counter_array& operator=( const general_counter_array& ) = delete; public: typedef size_t size_type; general_counter_array( size_type size ); some_bumper_type& operator[]( size_type idx ); size_type size(); Integral load( size_type idx ); Integral exchange( size_type idx, Integral value ); };The
Integraltemplate type parameter shall be an integral type for whichstd::atomichas a specialization (29.5 [atomics.types.generic]).
constexpr general_counter_array( size_type );
- Effects:
Initialize the
general_counter_arraywith the given number of counters, each initialized to zero.
~general_counter_array()
- Requirements:
All attached
counter_broker_arrayobjects shall have been previously destroyed.- Effects:
Destroys the
general_counter_array.
size_type size()
- Returns:
The value passed to the constructor.
some_bumper_type& operator[]( size_type index )
- Precondition:
0 ≤
index<size()- Returns:
A reference to an object, identified by the
index, implementing the bumper concept [distctr.bumper].
Integer load( size_type index )
- Precondition:
0 ≤
index<size()- Returns:
the value of the bumper with the given
index. [Note: The value will be an approximation if any threads may concurrently update the counter. —end note]
Integer exchange( size_type index, Integer value )
- Precondition:
0 ≤
index<size()- Effects:
Atomically replaces the value of the bumper at the given
indexwith the givenvalue.- Returns:
The replaced value.
counter_broker_array [distctr.brokerarray]Add a new section.
template< typename Integral > class counter_broker_array { counter_broker_array( const counter_broker_array& ) = delete; counter_broker_array& operator=( const counter_broker_array& ) = delete; public: typedef size_t size_type; counter_broker_array( general_counter_array<Integral>& ); some_bumper_type& operator[]( size_type idx ); size_type size(); };The
Integraltemplate type parameter shall be an integral type for whichstd::atomichas a specialization (29.5 [atomics.types.generic]).
constexpr counter_broker_array( general_counter_array< Integral >& );
- Effects:
Initialize the broker array. Atomically associate the
counter_broker_arraywith thegeneral_counter_array.
~counter_broker_array()
- Effects:
Atomically moves each count value retained within the broker to its associated position in the counter array. Atomically disassociates the broker array from the counter array. Destroys the broker array.
size_type size()
- Returns:
The value passed to the constructor.
some_bumper_type& operator[]( size_type index )
- Precondition:
0 ≤
index<size()- Returns:
A reference to an object, identified by the
index, implementing the bumper concept [distctr.bumper].
This paper revises P0261R2 - 2017-02-05. Its changes are as follows.
Add introductory text to [distctr.general].
Define modify and read operations in [distctr.general].
Add a Synchronization section [distctr.sync] in [distctr.general]. Define read operations that happen after other read operations to include all those increments seen by the other. Define the increment and decrement operations to impose no memory order.
Add a Value Computation section [distctr.value] in [distctr.general]. Define the abstract value of the counter for any read operation as the sum of observed modify operations.
Add a questions, answers and alternatives section to address design choices.
Add the descriptions of the size function.
Add prerequisites using size where appropriate.
P0261R2 revised P0261R1 - 2016-10-13. Its changes are as follows.
Merge all synchronization clauses into a single new section. Require cache coherence on counts.
Clarify that moving the count from the broker does so an atomic element at at time, not atomically in bulk.
Remove "There is no default value" as it follows from the declaration and is potentially confusing.
Move the note about increments returning void
to the top of the section so that it applies to all operations.
P0261R1 revised P0261R0 - 2016-02-14. Its changes are as follows.
Clarify that general_counter::load is atomic.
Use the type-specific phrase "counter_broker object"
rather than the general phrase "counter broker"
in the proposed wording.
Do the same with the array forms.
Correct parameter and description of effective broker constructors.
Add a size_type to the counter array classes.
Clarify initialization of counter array values.
Clarify the adoption of constexpr general_counter constructors.
Expand on the problem of counter overflow.
Update the document header.
Add an abstract.
Add subheadings to Solution.
Do general editorial cleanup.
P0261R0 revised N3706 - 2013-09-01. Its changes are as follows:
Add more context to the introduction and rework subsequent discussion.
Reduce the number of types to four,
a general_counter
equivalent to the earlier strong_duplex,
a counter_broker
equivalent to the earlier strong_broker,
a general_counter_array
equivalent to the earlier strong_duplex_array, and
a counter_broker_array
equivalent to the earlier strong_broker_array.
Remove the atomicity specification. Only full and relaxed atomicity is now supported.
As a consequence of the above two points, drastically reduce the solution presentation.
Update the implementation discussion with a new technique and new pointers to the code base.
Add proposed wording.
N3706 revised N3355 = 12-0045 - 2012-01-14. Its changes are as follows:
Add more context to the introduction.
Change named inc and dec functions to operators.
Place all declarations within namespace counter.
The namespace serves to avoid redundancy in names.
Rename types for increased clarity.
Change separate serial and atomic counter types with a single type templatized on the atomicity.
Add issues of memory order and contention to the atomicity parameter.
Add counter arrays.
Add open issues section.
Add implementation section.
Add revision history section.
Add references section.