Re: [boost] boost lockfree queue.hpp - alternate implementation w/o compare_exchange

Subject: Re: [boost] boost lockfree queue.hpp - alternate implementation w/o compare_exchange
From: Gottlob Frege (gottlobfrege_at_[hidden])
Date: 2015-10-26 17:36:26

Have you run ThreadSanitizer on it?

On Tue, Oct 20, 2015 at 11:02 PM,  <boost_at_[hidden]> wrote:
> Is there any interest in an alternate implementation of a lockfree queue
> that doesn't use compare_exchange (assuming it is correct, or can be made to
> be correct)?
>
> I have a multi-producer / multi-consumer implementation that generally
> doesn't suffer nearly as much from contention.  On my hardware (an older
> high-end 64 bit x86 workstation) I get roughly comparable throughput with
> existing boost lockfree queue with 1 producer and 1 consumer.  With 4
> producers and 4 consumers I get about 3x throughput gain on existing boost
> lockfree queue.  However, this implementation has a weakness, it suffers in
> over subscription scenarios (threads in middle of push/pop that lose their
> time-slice); with twice as many threads as cores its performances is
> comparable to boost lockfree queue (using try_push/try_pop, otherwise
> throughput is down the toilet), and with 4x thread/core ratio the throughput
> is somewhat less than boost lockfree queue.
>
> This was tested on Windows 7 x64 with Inel Xeon X5355 x 2 (2 sockets w/ 4
> cores a piece) @ 2.66 Ghz and 32 GB RAM @ 667 MHz.
>
> The implementation can be found here:
> https://github.com/benmccart/GuarunteedMpmcQueue/blob/master/queue/queue.hpp
>
> Is the implementation correct?  Can anyone spot any race conditions, or
> conditions under which the implementation would fail to work as expected
> (other than queue is full/empty or over-subscription)?
>
> ~Ben
>
>
> _______________________________________________
> Unsubscribe & other changes:
> http://lists.boost.org/mailman/listinfo.cgi/boost