Boost-announce

Download

boost-announce@lists.stage.boost.cppalliance.org

October 2015

1 participants
1 discussions

Boost.Fiber mini-review report
by Nat Goodspeed 10 Oct '15

10 Oct '15

The mini-review of Boost.Fiber by Oliver Kowalke ran from Friday September 4th through Sunday September 13th. There was lively discussion, but only two actual reviews: Agustín K-ballo Bergé: not yet (list of issues below) Niall Douglas: yes, plus suggestions (below)[0] In addition I received two "review withheld" notes: Vicente J. Botet Escriba wanted me to address the previous review point by point. That seems like a reasonable request, and I apologize, but I too have time pressures. Giovanni Piero Deretta made suggestions (synopsized below) with the statement that he will make a recommendation when he sees the library author's progress along those lines. While in theory these results might seem like a hung jury, the very paucity of formal reviews supports a cautious approach. The verdict is: not yet, with specific points as noted below. No reviewer explicitly designated specific points as being essential. (Niall Douglas explicitly designated his suggestions as post-adoption enhancements.) I have made an attempt to rank the recommendations received in descending order of importance to the community. My criteria are: Robustness: does lack of this request pose significant risk to a consuming program? API stability: could this request be considered an implementation detail, or would the change break early adopters? Demand: how many people requested the same thing? If you feel that I have misrepresented the importance of one of your points (or omitted something entirely), please respond. Ultimately, of course, the importance of any given point to the community at large will be reflected in the next mini-review. ----------------------------------------------------------------- Requests: Cross-thread fiber migration: Evidently the community strongly desires the ability to transfer a suspended fiber from one thread to another.[1] Of course there must be controls. Migration might be explicit; but if implicit, there must be a way to designate which fibers may and may not be migrated. An RAII class that temporarily disables migration would also be desirable -- for instance, to guarantee TLS consistency for some specific period. Cross-thread migration support implies a requirement to document the thread safety requirements of the customization API. Standard required timing specifications: meet (30.2.4 Timing specifications [thread.req.timing]). Adapt one or more existing test suites for thread facilities to validate the correctness of the Fiber library's synchronization constructs: e.g. from libc++, libstdc++, Boost.Thread, HPX. std::chrono support: ensure that wait_until() can accept any supported time_point specialization. (The library author asserts that this has already been fixed on the 'develop' branch.) C++11 compatibility: Agustín K-ballo Bergé volunteered to help with this point post-adoption. Paul Fultz offered fit::capture to emulate C++14 move capture. Document that this_fiber::yield() (et al.) are valid from main(), or in general from a thread function, as well as from explicitly-launched fibers. Make shared_state lock requirements explicit: for example, make mark_ready_and_notify_() accept a std::unique_lock<>& (or &&) to force its caller to provide one. Avoid holding mutex during condition_variable::notify_*() calls. (The library author asserts that this has already been fixed on the 'develop' branch.) packaged_task must not decay-copy its task arguments. Align fibers::async() signature, result description and invoke_helper implementation with std::async(). Avoid returning std::move(...). Wait list implementation: try an intrusive linked list of fiber_context objects. Avoid sleeping thread when no fibers are ready -- block on external event. Better handling of waiting tasks (decentralized wait queues?). Optimization to resume awakened fiber on the thread that awakens it. Template redundancy: eliminate duplication between specializations for T, T& and void. The future, shared_future, shared_state, promise and task_object templates were specifically cited. Leverage predicate-based condition::wait() overload internally. Nested schedulers: make schedulers be schedulable entities themselves. Improve robustness of `promise' handling in Asynchronous Callbacks documentation. Avoid overconfident remarks about it. Fix Note overstating the requirements on condition_variable::wait(). The discussion also surfaced the desirability of Grand Unified Future Theory, but it was generally agreed that this is out of scope. Put differently, it would be unreasonable to put the onus on Fiber to solve that very broad cross-library problem. Many thanks to all participants, and thanks especially to Oliver Kowalke for all his work on this library! Nat Goodspeed Boost.Fiber Review Manager ________________________________ [0] Niall Douglas snuck his review recommendation into the following mail message: http://lists.boost.org/Archives/boost/2015/09/225228.php (second to final paragraph). His suggestions were in private mail. [1] Giovanni Deretta suggests that it might be possible to allocate -- without actually running -- an actual kernel thread per fiber, simply to give each fiber its own TLS area.

1 0