Flow 1.0.1
Flow project: Full implementation reference.
|
Flow module containing tools enabling multi-threaded event loops operating under the asynchronous-task proactor pattern, by providing a streamlined API around boost.asio event loops with added advanced task- and thread-scheduling features. More...
Classes | |
class | Concurrent_task_loop |
The core flow::async interface, providing an optionally multi-threaded thread pool onto which runnable Task s, optionally arranged into concurrency-avoiding Op s, can be boost.asio-posted for subsequent execution. More... | |
class | Cross_thread_task_loop |
Concrete Concurrent_task_loop that is able to efficiently schedule Task s within a given Op to execute in different threads while still properly avoiding concurrency. More... | |
class | Op_list |
Simple, immutable vector -like sequence of N opaque async::Op objects, usually corresponding to N worker threads, with typically used operations such as [index] and random Op selection. More... | |
class | Segregated_thread_task_loop |
Concrete Concurrent_task_loop that uses the legacy pin-Task s-within-1-Op -to-1-thread method of achieving required non-concurrency of Task s. More... | |
class | Single_thread_task_loop |
A Concurrent_task_loop -related adapter-style class that represents a single-thread task loop; essentially it is pleasant syntactic sugar around a Concurrent_task_loop impl that cuts out concepts that become irrelevant with only one thread involved. More... | |
class | Task_qing_thread |
Internally used building block of various concrete Concurrent_task_loop subclasses that encapsulates a thread that spawns at construction time and a dedicated-or-shared util::Task_engine (a/k/a boost.asio io_service ) run() ning in that new thread. More... | |
class | Timed_concurrent_task_loop |
Decorates a general – potentially multi-threaded – Concurrent_task_loop of any kind but with timing capabilities. More... | |
class | Timed_concurrent_task_loop_impl |
Decorator of a Concurrent_task_loop with same or greater lifetime that accumulates time elapsed in any tasks posted or scheduled onto that loop. More... | |
class | Timed_single_thread_task_loop |
Identical to Single_thread_task_loop, but all tasks posted through it are automatically timed, with the result accessible via accumulated_time(). More... | |
Typedefs | |
using | Task = Function< void()> |
Short-hand for a task that can be posted for execution by a Concurrent_task_loop or flow::util::Task_engine; it is simply something callable via () with no arguments and returning nothing. More... | |
using | Op = boost::any |
An object of this opaque type represents a collection of 1 or more async::Task, past or future, such that: if one performs C->post(J, F) and C->post(K, G) (where C is Concurrent_task_loop* , JK are the same async::Op& , or one refers to a transitive copy of the other, and FG are both Task s), then F and G will NOT execute concurrently. More... | |
using | Scheduled_task = util::Scheduled_task |
Similar to flow::async::Task but used for scheduled-in-future tasks as opposed to to-be-run-ASAP tasks. More... | |
using | Task_asio_err = Function< void(const Error_code &)> |
Short-hand for a boost.asio completion handler: The minimal type, taking only a flow::Error_code a/k/a boost::system::error_code . More... | |
using | Task_asio_err_sz = Function< void(const Error_code &, size_t)> |
Short-hand for a boost.asio completion handler: The type that takes a size_t count of things successfully transferred usually. More... | |
using | Task_engine_ptr = boost::shared_ptr< util::Task_engine > |
Short-hand for reference-counting pointer to a mutable util::Task_engine (a/k/a boost::asio::io_service ). More... | |
using | Strand_ptr = boost::shared_ptr< util::Strand > |
Short-hand for ref-counted pointer to util::Strand. More... | |
Enumerations | |
enum class | Synchronicity { S_ASYNC , S_ASYNC_AND_AWAIT_CONCURRENT_COMPLETION , S_ASYNC_AND_AWAIT_CONCURRENT_START , S_OPPORTUNISTIC_SYNC_ELSE_ASYNC } |
Enumeration indicating the manner in which asio_exec_ctx_post(), and various boost.asio "post" operations like it or based on it, are to actually execute the given task in relation to when the "posting" routine, itself, returns control to its caller. More... | |
Functions | |
template<typename Execution_context > | |
void | asio_exec_ctx_post (log::Logger *logger_ptr, Execution_context *exec_ctx, Synchronicity synchronicity, Task &&task) |
An extension of boost.asio's post() and dispatch() free function templates, this free function template allows the user to more easily select the synchronicity behavior as the given task is posted onto the given execution context (util::Task_engine or util::Strand at least). More... | |
unsigned int | optimal_worker_thread_count_per_pool (flow::log::Logger *logger_ptr, bool est_hw_core_sharing_helps_algo) |
Assuming a planned thread pool will be receiving ~symmetrical load, and its UX-affecting (in particular, per-op latency-affecting) operations are largely between processor and RAM: Returns the # of threads to store in that pool for efficient performance. More... | |
void | optimize_pinning_in_thread_pool (flow::log::Logger *logger_ptr, const std::vector< util::Thread * > &threads_in_pool, bool est_hw_core_sharing_helps_algo, bool est_hw_core_pinning_helps_algo, bool hw_threads_is_grouping_collated) |
Assuming the same situation as documented for optimal_worker_thread_count_per_pool(), and that indeed the pool now contains that number of running threads: Attempts to optimize thread-core-pinning behavior in that pool for efficient performance. More... | |
template<typename Handler > | |
auto | asio_handler_via_op (Concurrent_task_loop *loop, const Op &op, Handler &&handler) |
Given a boost.asio completion handler handler for a boost.asio async_*() action on some boost.asio I/O object to be initiated in the immediate near future, returns a wrapped handler with the same signature to be passed as the handler arg to that async_*() action, so that handler() will execute non-concurrently with other tasks in Op op . More... | |
template<typename Exec_ctx_ptr > | |
Exec_ctx_ptr | op_to_exec_ctx (Concurrent_task_loop *loop, const Op &op) |
Template specialization model for operation that obtains the underlying execution context, such as a util::Task_engine or util::Strand, stored in an async::Op generated by the given Concurrent_task_loop. More... | |
template<> | |
Task_engine_ptr | op_to_exec_ctx< Task_engine_ptr > (Concurrent_task_loop *loop, const Op &op) |
Template specialization for operation that obtains the underlying execution context, in this case a util::Task_engine, stored in an async::Op generated by the given Segregated_thread_task_loop. More... | |
template<> | |
Strand_ptr | op_to_exec_ctx< Strand_ptr > (Concurrent_task_loop *loop, const Op &op) |
Template specialization for operation that obtains the underlying execution context, in this case a util::Strand, stored in an async::Op generated by the given Cross_thread_task_loop. More... | |
uint16_t | cpu_idx () |
Returns the 0-based processor logical (not hardware) core index of the core executing the calling thread presently. More... | |
template<> | |
Task_engine_ptr | op_to_exec_ctx< Task_engine_ptr > (Concurrent_task_loop *loop, const Op &op) |
Template specialization for operation that obtains the underlying execution context, in this case a util::Task_engine, stored in an async::Op generated by the given Segregated_thread_task_loop. More... | |
template<> | |
Strand_ptr | op_to_exec_ctx< Strand_ptr > (Concurrent_task_loop *loop, const Op &op) |
Template specialization for operation that obtains the underlying execution context, in this case a util::Strand, stored in an async::Op generated by the given Cross_thread_task_loop. More... | |
Flow module containing tools enabling multi-threaded event loops operating under the asynchronous-task proactor pattern, by providing a streamlined API around boost.asio event loops with added advanced task- and thread-scheduling features.
There is also support for single-threaded event loops.
In simpler terms, at its core – including when the "pool" has just one thread, which is very common – it provides a compact way of both starting thread(s) and posting/scheduling tasks and I/O to run in such thread(s). By default one must worry about each of those 2 concerns separately and decide how exactly to hook them up; which is not rocket science, but it is a ton of boiler-plate, and it is easy to make mistakes and/or omit useful startup/shutdown practices, logging, and more. This module provides, via Concurrent_task_loop and its implementations, at least that consistency/standardization. Plus, it provides certain advanced features as mentioned above.
boost::asio::io_service
), util::Strand (a/k/a boost::asio::io_service::strand
), and util::Timer. flow::async Flow module somewhat streamlines this API in such a way as to keep the user's focus on their conceptual async-task-driven algorithm as opposed to details of threads, handlers, cores, etc. The async::Op opaque type is central to this streamlined API, plus the central class Concurrent_task_loop.async
module a value-add over just boost.asio, or over just boost.asio with some thread-core-affinity utility functions on the side.std::queue<T>
is commongly a simplified std::deque<T>
or list
underneath. using flow::async::Op = typedef boost::any |
An object of this opaque type represents a collection of 1 or more async::Task, past or future, such that: if one performs C->post(J, F)
and C->post(K, G)
(where C is Concurrent_task_loop*
, JK are the same async::Op&
, or one refers to a transitive copy of the other, and FG are both Task
s), then F and G will NOT execute concurrently.
In addition, it is guaranteed that copying (via constructor or assignment) of async::Op is has performance characteristics no worse than those of shared_ptr
. I.e., it is to be thought of as light-weight.
The value Op()
is designated as a null/sentinel value and must not be passed to Concurrent_task_loop::post() or anything built on it.
That's the formal definition. We reiterate that copying these is cheap; and moreover two Op
s such that one is a copy (of a copy, of a copy, of a copy...) of another, then these are conceptually isomorphic: they represent the same op, or collection of Task
s that must never execute concurrently. Finally, tip: Don't think of an Op
as a collection of 2+ Task
s; but rather a tag or label that associates 2+ Task
s with each other. (Also, nothing prevents an async::Task being a part of 2+ Op
s simultaneously, though informally speaking it's arguably best not to make code maintainers grok such a design.)
By convention in comments we represent Op
s with the letters J, K, L.
Op
versus just a stand-alone Task
?When choosing a Concurrent_task_loop::post() (the one with Op
vs. one without; or similar choices in more advanced cases), here are some things to remember. These can be derived independently and are only included as a convenience/refresher:
Op
prevents Task
s from executing concurrently. If there is exactly 1 thread in a pool, then they couldn't anyway, so an Op
is not needed...Task
, one post()
s (or similar) exactly 1 Task
to asynchronously execute next. Since it's the last statement, and post()
and similar are explicitly thread-safe, this ensures the current and next Task
s do not execute concurrently. So an Op
is not needed...Task
s are "officially" run under the banner of a single Op
, even if some parts of the async handling of the "conceptual" operation are serial and hence don't technically require an Op
. Example: In a web server it is reasonable to create an Op
for the entire request, with all Task
s (even serially called ones) being associated with that per-request Op
; then, simply, no locking is necessary for per-request data structure(s). It's much easier to explain, "you don't need to lock 'em," vs. "you don't need to lock 'em,
unless the logic changes in such a way as to...."To be fair, those tips ignore performance; they implicitly assume using Op
pointlessly (functionally, not stylistically, so) is otherwise "free." It is not free; depending on internal details using Op
might involve a util::Strand and/or pinning stuff to a specific thread. Informally, this is seldom a big deal in practice; but in performance-sensitive projects one must remember there is a cost.
boost::any
(storing nothing heavier-weight than a shared_ptr
to satisfy that explicit guarantee above) is one way to do this. A rigid polymorphic hierarchy (virtual
) is another. Performance-wise they're similar, costing essentially a single virtual
pointer lookup per Concurrent_task_loop::post(). boost.any is much pithier, not requiring a class hierarchy at all, and otherwise it's pretty much the same in terms of how they're used internally. Some might say the polymorphic hierarchy is clearer, because it is explicit, but I feel comments alone, too, can be just as clear, and brevity is a virtue.
Definition at line 153 of file async_fwd.hpp.
using flow::async::Scheduled_task = typedef util::Scheduled_task |
Similar to flow::async::Task but used for scheduled-in-future tasks as opposed to to-be-run-ASAP tasks.
In practice it's the same thing but takes a single bool
argument with the meaning explained in util::schedule_task_from_now() doc header (spoiler alert: whether it ran as scheduled or was short-fired by user, as of this writing).
Task
s are dealt with, one may usually assume the same extends equally to a Scheduled_task
, merely at a different point in time. We omit that explicit language for brevity; it is to be assumed. Definition at line 165 of file async_fwd.hpp.
using flow::async::Strand_ptr = typedef boost::shared_ptr<util::Strand> |
Short-hand for ref-counted pointer to util::Strand.
We at times return new Strand
s (Cross_thread_task_loop::create_op()), so universally use ref-counted pointers to Strand
s to not have to worry about Strand
lifetimes too hard.
Key fact: The type Cross_thread_task_loop loads into superclass's async::Op (boost::any
) is Strand_ptr
. That is, a Strand
is the mechanism used to bundle together non-concurrent tasks in Cross_thread_task_loop.
Definition at line 212 of file async_fwd.hpp.
using flow::async::Task = typedef Function<void ()> |
Short-hand for a task that can be posted for execution by a Concurrent_task_loop or flow::util::Task_engine; it is simply something callable via ()
with no arguments and returning nothing.
By convention in comments we represent Task
s with the letters F, G, H.
Definition at line 96 of file async_fwd.hpp.
using flow::async::Task_asio_err = typedef Function<void (const Error_code&)> |
Short-hand for a boost.asio completion handler: The minimal type, taking only a flow::Error_code a/k/a boost::system::error_code
.
Definition at line 171 of file async_fwd.hpp.
using flow::async::Task_asio_err_sz = typedef Function<void (const Error_code&, size_t)> |
Short-hand for a boost.asio completion handler: The type that takes a size_t
count of things successfully transferred usually.
Definition at line 177 of file async_fwd.hpp.
using flow::async::Task_engine_ptr = typedef boost::shared_ptr<util::Task_engine> |
Short-hand for reference-counting pointer to a mutable util::Task_engine (a/k/a boost::asio::io_service
).
This is generally how classes in the Concurrent_task_loop hierarchy refer to their internally used Task_engine
s but also in advanced cases may be communicated to their user.
Why do that instead of using raw Task_engine*
? It may not be obvious as you read this now, but all kinds of pain goes away due to the possibilities of who happens to own an underlying Task_engine
– it can be shared by all threads in pool, or each can have its own Task_engine
, depending on the chosen Concurrent_task_loop subclass – and in what order those things might execute their destructors. By simply using a shared_ptr<>
everywhere, with small overhead we ensure an underlying Task_engine
does not get destroyed until everything that uses it is destroyed first, and that could be a number of things. By using shared_ptr<>
we needn't break our heads worrying about executing de-init pieces of code in just the right order just to avoid early-free. (This is said from experience.)
This is reinforced in those semi-advanced cases where a Task_engine
is passed to user via public API.
Definition at line 198 of file async_fwd.hpp.
|
strong |
Enumeration indicating the manner in which asio_exec_ctx_post(), and various boost.asio "post" operations like it or based on it, are to actually execute the given task in relation to when the "posting" routine, itself, returns control to its caller.
Basically it indicates whether the execution should be synchronous or asynchronous and how, if it all, to wait for its completion – or its initiation. The enum
members' meanings are the key things to understand; and there's some discussion in their doc headers that might be useful as a boost.asio refresher.
Enumerator | |
---|---|
S_ASYNC | Simply post the given task to execute asynchronously in some execution context – as soon as the context's scheduler deems wise but specifically not inside the posting routine itself; and return as soon as possible having thus posted it. That is: work in the manner of boost.asio In particular, suppose you're calling
Either way, |
S_ASYNC_AND_AWAIT_CONCURRENT_COMPLETION | Same as Synchronicity::S_ASYNC but the posting routine then waits as long as necessary for the given task to complete; and only then returns. That is: work in the manner of boost.asio
This mode is reminiscent of the promise/future concept and allows one to easily solve the age-old problem of "how do I ask a thread/pool to do a thing and then wait for get a result?". One might do this by manually using a promise/future pair; or even mutex/condition variable pair; but by using this such boiler-plate is reduced (along with fewer bugs).
|
S_ASYNC_AND_AWAIT_CONCURRENT_START | Same as Synchronicity::S_ASYNC but the posting routine then waits as long as necessary for the given task to just about to begin executing concurrently (so that any subsequent This is most similar to
|
S_OPPORTUNISTIC_SYNC_ELSE_ASYNC | Execute the given task synchronously, if the scheduler determines that the calling thread is in its thread pool and is the best thread for the task; otherwise act identically to Synchronicity::S_ASYNC. That is: work in the manner of boost.asio
|
Definition at line 222 of file async_fwd.hpp.
void flow::async::asio_exec_ctx_post | ( | log::Logger * | logger_ptr, |
Execution_context * | exec_ctx, | ||
Synchronicity | synchronicity, | ||
Task && | task | ||
) |
An extension of boost.asio's post()
and dispatch()
free function templates, this free function template allows the user to more easily select the synchronicity behavior as the given task is posted onto the given execution context (util::Task_engine or util::Strand at least).
It also adds TRACE logging including that book-ending the task's execution (aiding debugging, etc.). The synchronicity
argument controls the specific way in which task
is posted onto *exec_ctx
; see Synchronicity doc header.
This call causes task
to execute in a thread controlled by *exec_ctx
. The latter, at this time, must be either util::Task_engine or util::Strand (which itself is born of a Task_engine
). It is likely that it will work equally well for other entities satisfying the boost.asio ExecutionContext
concept (see boost.asio docs), but this is untested and not thought through formally, so officially such uses cause undefined behavior as of this writing.
Semantics can be found in Synchronicity doc headers which are required reading before using this function. However, briefly and informally, the utility of this function is as follows:
post()
works on a Task_engine
or Strand
already and equals mode Synchronicity::S_ASYNC; with this function you'll get some added debug logging as well.dispatch()
and Synchronicity::S_OPPORTUNISTIC_SYNC_ELSE_ASYNC are similarly related: same thing but more logging.post()
with a promise/future pair, wherein the caller performs unique_future.wait()
after the post()
, while the task always sets promise.set_value()
just before returning (and, again, more logging).promise.set_value()
executes just before executing task()
. Hence any tasks queued before task()
will first execute (same as previous bullet); and then asio_exec_ctx_post() will return, just as task()
begins executing as opposed to wait for its completion. This is useful to prevent E.stop()
after our return (where E
is the Task_engine
that either is or produced *exec_ctx
) will be too late to prevent task()
from executing and completing.In all cases, one gets more logging and arguably a bit of syntactic sugar, but S_ASYNC_AND_AWAIT_CONCURRENT_COMPLETION
and S_ASYNC_AND_AWAIT_CONCURRENT_START
in particular eliminate quite a bit of tedious and hairy code and explanations.
Lastly, if *exec_ctx
is currently not running, then the semantics described in Synchronicity doc header still apply but are deferred until it does run. In particular in mode Synchronicity::S_ASYNC_AND_AWAIT_CONCURRENT_COMPLETION and Synchronicity::S_ASYNC_AND_AWAIT_CONCURRENT_START this function shall not return until *exec_ctx
at least does begin running in at least 1 thread; while the other two modes reduce to post()
, which returns immediately, leaving task()
to run once *exec_ctx
starts.
Execution_context | util::Task_engine or util::Strand. See note above regarding other possibilities. |
logger_ptr | Logger to use in this function. |
exec_ctx | The execution context controlling the thread pool onto which to load the task for ASAP execution. |
synchronicity | Controls the precise behavior. See above. |
task | The task – taking no arguments and returning no value – to load onto *exec_ctx . |
Definition at line 31 of file util.hpp.
References cpu_idx(), FLOW_LOG_SET_CONTEXT, FLOW_LOG_TRACE, S_ASYNC, S_ASYNC_AND_AWAIT_CONCURRENT_COMPLETION, S_ASYNC_AND_AWAIT_CONCURRENT_START, S_OPPORTUNISTIC_SYNC_ELSE_ASYNC, and flow::log::Logger::should_log().
Referenced by flow::net_flow::Node::close_abruptly(), flow::net_flow::Node::connect_with_metadata(), flow::net_flow::Node::event_set_close(), flow::net_flow::Node::event_set_create(), flow::net_flow::Node::listen(), and flow::net_flow::Node::sock_info().
auto flow::async::asio_handler_via_op | ( | Concurrent_task_loop * | loop, |
const Op & | op, | ||
Handler && | handler | ||
) |
Given a boost.asio completion handler handler
for a boost.asio async_*()
action on some boost.asio I/O object to be initiated in the immediate near future, returns a wrapped handler with the same signature to be passed as the handler arg to that async_*()
action, so that handler()
will execute non-concurrently with other tasks in Op op
.
This is analogous to boost.asio's bind_executor(Strand, Handler)
(which replaces boost.asio's now-deprecated util::Strand::wrap(Handler)
).
The mechanics of using this are explained in Concurrent_task_loop doc header. Using this in any other fashion leads to undefined behavior.
Handler | boost.asio handlers are, essentially, all void -returning but can take various arg sets. E.g., util::Timer (a/k/a boost::asio::basic_waitable_timer ) expects a handler that takes only an Error_code ; while boost::asio::ip:tcp::socket::read_some() expects one to take bytes-received size_t and an Error_code . This template supports all handlers via auto magic. |
loop | Active loop that spawned Op op . |
op | See 3-arg Concurrent_task_loop::post(). |
handler | Completion handler for the boost.asio async_*() operation to be initiated soon. It may be move d and saved. |
handler()
but also satisfying the constraints of Op op
. Definition at line 691 of file concurrent_task_loop.hpp.
References flow::async::Concurrent_task_loop::post(), and S_OPPORTUNISTIC_SYNC_ELSE_ASYNC.
uint16_t flow::async::cpu_idx | ( | ) |
Returns the 0-based processor logical (not hardware) core index of the core executing the calling thread presently.
Since the value of this can change from statement to statement in the same thread, probably this is not to be used except for logging/reporting.
Definition at line 30 of file async.cpp.
Referenced by asio_exec_ctx_post(), flow::async::Cross_thread_task_loop::schedule_from_now(), and flow::async::Segregated_thread_task_loop::schedule_from_now_impl().
Exec_ctx_ptr op_to_exec_ctx | ( | Concurrent_task_loop * | loop, |
const Op & | op | ||
) |
Template specialization model for operation that obtains the underlying execution context, such as a util::Task_engine or util::Strand, stored in an async::Op generated by the given Concurrent_task_loop.
Each subclass (impl) of Concurrent_task_loop shall provide a specialization of this template with Exec_ctx_ptr
template param being the appropriate boost.asio-compatible execution context type for that loop type's Op create_op()
.
The mechanics of using this are explained in Concurrent_task_loop doc header. Beyond that please see the particular specialization's doc header.
Exec_ctx_ptr | A pointer type (raw or smart) pointing to an execution context type satisfying boost.asio's "execution context" concept. As of this writing the known values would be pointers to util::Task_engine and util::Strand, but really it depends on the particular subclass of Concurrent_task_loop for the *loop arg. See its doc header near the particular Concurrent_task_loop subclass. |
loop | Loop object that, one way or another, generated and returned op . |
op | async::Op from *loop from which to extract the execution context object on which you'd like to perform custom boost.asio work. |
Strand_ptr op_to_exec_ctx< Strand_ptr > | ( | Concurrent_task_loop * | loop, |
const Op & | op | ||
) |
Template specialization for operation that obtains the underlying execution context, in this case a util::Strand, stored in an async::Op generated by the given Cross_thread_task_loop.
boost.asio tip: The returned util::Strand may be useful not only as an argument to bind_executor()
(formerly Strand::wrap()
, now deprecated) but can also be passed in lieu of a util::Task_engine into boost.asio-enabled I/O object constructors (util::Timer, boost::asio::ip::tcp::socket
, etc.). The latter use uses the Strand
as an "execution context."
Note Concurrent_task_loop::task_engine() is spiritually related to this function; but while that one gives one a util::Task_engine, which corresponds to the entire thread pool, this one returns an execution context specifically assigned to a multi-step async op op
.
loop | Loop object that, one way or another, generated and returned op . Behavior is undefined if the concrete pointed-to type is not Cross_thread_task_loop. (assertion may trip). |
op | async::Op from *loop from which to extract the execution context on which you'd like to perform custom boost.asio work. Behavior is undefined if it is not from *loop (assertion may trip). |
E
such that loop->task_engine() == &E
. Definition at line 420 of file x_thread_task_loop.cpp.
Referenced by flow::async::Cross_thread_task_loop::post(), and flow::async::Cross_thread_task_loop::schedule_from_now().
Strand_ptr flow::async::op_to_exec_ctx< Strand_ptr > | ( | Concurrent_task_loop * | loop, |
const Op & | op | ||
) |
Template specialization for operation that obtains the underlying execution context, in this case a util::Strand, stored in an async::Op generated by the given Cross_thread_task_loop.
boost.asio tip: The returned util::Strand may be useful not only as an argument to bind_executor()
(formerly Strand::wrap()
, now deprecated) but can also be passed in lieu of a util::Task_engine into boost.asio-enabled I/O object constructors (util::Timer, boost::asio::ip::tcp::socket
, etc.). The latter use uses the Strand
as an "execution context."
Note Concurrent_task_loop::task_engine() is spiritually related to this function; but while that one gives one a util::Task_engine, which corresponds to the entire thread pool, this one returns an execution context specifically assigned to a multi-step async op op
.
loop | Loop object that, one way or another, generated and returned op . Behavior is undefined if the concrete pointed-to type is not Cross_thread_task_loop. (assertion may trip). |
op | async::Op from *loop from which to extract the execution context on which you'd like to perform custom boost.asio work. Behavior is undefined if it is not from *loop (assertion may trip). |
E
such that loop->task_engine() == &E
. Definition at line 420 of file x_thread_task_loop.cpp.
Referenced by flow::async::Cross_thread_task_loop::post(), and flow::async::Cross_thread_task_loop::schedule_from_now().
Task_engine_ptr op_to_exec_ctx< Task_engine_ptr > | ( | Concurrent_task_loop * | loop, |
const Op & | op | ||
) |
Template specialization for operation that obtains the underlying execution context, in this case a util::Task_engine, stored in an async::Op generated by the given Segregated_thread_task_loop.
While *loop
is running, the Task_engine is running in exactly 1 thread.
Note Concurrent_task_loop::task_engine() is spiritually related to this function; but while that one gives one a random thread's util::Task_engine, this one returns the specific thread's assigned to a multi-step async op op
.
loop | Loop object that, one way or another, generated and returned op . Behavior is undefined if the concrete pointed-to type is not Segregated_thread_task_loop. (assertion may trip). |
op | async::Op from *loop from which to extract the execution context on which you'd like to perform custom boost.asio work. Behavior is undefined if it is not from *loop (assertion may trip). |
E
used by *loop
. Definition at line 419 of file segregated_thread_task_loop.cpp.
Referenced by flow::async::Segregated_thread_task_loop::post(), and flow::async::Segregated_thread_task_loop::schedule_from_now().
Task_engine_ptr flow::async::op_to_exec_ctx< Task_engine_ptr > | ( | Concurrent_task_loop * | loop, |
const Op & | op | ||
) |
Template specialization for operation that obtains the underlying execution context, in this case a util::Task_engine, stored in an async::Op generated by the given Segregated_thread_task_loop.
While *loop
is running, the Task_engine is running in exactly 1 thread.
Note Concurrent_task_loop::task_engine() is spiritually related to this function; but while that one gives one a random thread's util::Task_engine, this one returns the specific thread's assigned to a multi-step async op op
.
loop | Loop object that, one way or another, generated and returned op . Behavior is undefined if the concrete pointed-to type is not Segregated_thread_task_loop. (assertion may trip). |
op | async::Op from *loop from which to extract the execution context on which you'd like to perform custom boost.asio work. Behavior is undefined if it is not from *loop (assertion may trip). |
E
used by *loop
. Definition at line 419 of file segregated_thread_task_loop.cpp.
Referenced by flow::async::Segregated_thread_task_loop::post(), and flow::async::Segregated_thread_task_loop::schedule_from_now().
unsigned int flow::async::optimal_worker_thread_count_per_pool | ( | flow::log::Logger * | logger_ptr, |
bool | est_hw_core_sharing_helps_algo | ||
) |
Assuming a planned thread pool will be receiving ~symmetrical load, and its UX-affecting (in particular, per-op latency-affecting) operations are largely between processor and RAM: Returns the # of threads to store in that pool for efficient performance.
This will be used, by doc header contract, by all (as of this writing) Concurrent_task_loop subclasses if so specified via n_threads_or_zero == 0
. So in that context one needn't call this directly. However, it may be useful directly when one is operating a thread pool but without a Concurrent_task_loop.
logger_ptr | Logger to use in this function. |
est_hw_core_sharing_helps_algo | Set this to true if you estimate the intended use for this thread pool is such that 2+ identically loaded pool threads sharing 1 physical core would handle the load (in total over those 2+ threads) better than just 1 thread using that same core would. Set it to false otherwise. Note that, generally, this should be assumed false , unless there is significant cache locality between those 2+ threads, meaning they tend to work on the same cacheably-small area in memory at ~the same time. For example, parallel matrix multiplication algorithms can thus benefit and would set it to true ; but that is the not the case by default; one would have to prove it, or design the algorithm with that in mind. |
Definition at line 35 of file concurrent_task_loop.cpp.
References FLOW_LOG_INFO, and FLOW_LOG_SET_CONTEXT.
Referenced by optimize_pinning_in_thread_pool().
void flow::async::optimize_pinning_in_thread_pool | ( | flow::log::Logger * | logger_ptr, |
const std::vector< util::Thread * > & | threads_in_pool, | ||
bool | est_hw_core_sharing_helps_algo, | ||
bool | est_hw_core_pinning_helps_algo, | ||
bool | hw_threads_is_grouping_collated | ||
) |
Assuming the same situation as documented for optimal_worker_thread_count_per_pool(), and that indeed the pool now contains that number of running threads: Attempts to optimize thread-core-pinning behavior in that pool for efficient performance.
hw_threads_is_grouping_collated
if desired. See namespace
flow::async doc header.logger_ptr | Logger to use in this function. |
threads_in_pool | These raw threads, which must number optimal_worker_thread_count_per_pool(same-relevant-args) , comprise the pool in question. They must be already spawned (e.g., have had some caller's code execute OK). |
est_hw_core_sharing_helps_algo | See optimal_worker_thread_count_per_pool(). |
est_hw_core_pinning_helps_algo | Set this to true if you have reason to believe that pinning each of the pool's threads to N (N >= 1) logical cores would improve performance in some way. Set it to false otherwise. As of this writing I don't know why it would be true specifically; but it can be researched; and I know in practice some applications do (in fact) do it, so it's not necessarily worthless, at least. |
hw_threads_is_grouping_collated | When the number of physical cores does not equal # of logical cores (hardware threads) – otherwise this arg is ignored – this determines the pattern in which each set of 2+ core-sharing hardware threads is arranged vs. the other sets. When false , it's like ABCDABCD, meaning logical cores 0,4 share core, 1,5 share different core, 2,6 yet another, etc. When true , it's like AABBCCDD instead. It seems true is either rare or non-existent, but I do not know for sure. |
Definition at line 78 of file concurrent_task_loop.cpp.
References FLOW_ERROR_SYS_ERROR_LOG_WARNING, FLOW_LOG_INFO, FLOW_LOG_SET_CONTEXT, FLOW_LOG_TRACE, optimal_worker_thread_count_per_pool(), and flow::util::ostream_op_string().
Referenced by flow::async::Segregated_thread_task_loop::start(), and flow::async::Cross_thread_task_loop::start().