[botan-devel] Threaded Filters/Operation Parallelisation

Jack Lloyd lloyd at randombit.net
Thu Jan 3 12:00:22 EST 2013


Thanks Joel!

Have you done any measurements on what kind of speedup can be obtained
on a multicore processor? I would imagine that the overhead of thread
interactions exceeds the gains for filters below some cost/byte. I'm
not sure if it is cycles/byte though; cache line access/byte may be a
more important question in many multicore cases.

Something else that would be very useful would be a demo application
that demonstrates usage and can let users understand what kind of
speedup they might see on their machines from using it, do you have
some test code that you could adapt to an example?

Thanks,
  Jack

On Wed, Jan 02, 2013 at 11:01:12PM +0000, Joel Low wrote:
> Thanks for the review, Jack.
> 
> I've included the updated patch. Let me know if that works.
> 
> Regards,
> Joel
> 
> -----Original Message-----
> From: botan-devel [mailto:botan-devel-bounces at randombit.net] On Behalf Of
> Jack Lloyd
> Sent: Wednesday, 2 January 2013 11:49 PM
> To: Botan development list
> Subject: Re: [botan-devel] Threaded Filters/Operation Parallelisation
> 
> Hi Joel,
> 
> I've only had time to very briefly review the patch but it looks good.
> 
> A few minor comments:
> 
> - Can you use an m_ prefix on all class members? I know a lot of the
>   code doesn't use a prefix or suffix at all, but I am trying to
>   ensure that all new or heavily touched code adopts m_ (rather than
>   suffix _ as you use in Semaphore or no prefix).
> 
> - Semaphore seems generally useful, move it to src/util/semaphore.{cpp,h}
> 
> - Add your name to the copyright headers of each modified file. (And
>   make sure all new sources have one)
> 
> - For some reason, despite all your work on Windows support and other
>   patches, your name is not already included in the list of copyright
>   owners in the license, so please also update license.rst
> 
> Can you make those changes and send the updated patch as an attachment to
> the list (so it can be archived by mailman)?
> 
> Thanks,
>   Jack
> 
> On Wed, Jan 02, 2013 at 06:51:10AM +0000, Joel Low wrote:
> > Hello all,
> > 
> > So I've finally managed to sit down and write this patch. The design 
> > seems quite different from what I had in mind, since this round I 
> > wanted it to be a drop-in replacement for Fork.
> > 
> > There's a bit of code taken from a blog post to emulate a semaphore: 
> > The link is there, and if someone knows of a better class (or better 
> > still, one within Botan itself), let me know and I'll use that class
> instead.
> > 
> > Let me know what you think.
> > 
> > Patch: http://pastebin.com/kgF0b7Yp (1 month visibility)
> > 
> > Regards,
> > Joel
> > 
> > -----Original Message-----
> > From: botan-devel-bounces at randombit.net 
> > [mailto:botan-devel-bounces at randombit.net] On Behalf Of Joel Low
> > Sent: Tuesday, 24 April 2012 5:07 PM
> > To: botan-devel at randombit.net
> > Subject: [botan-devel] Threaded Filters/Operation Parallelisation
> > 
> > Hello all,
> > 
> > Recently I've been playing with the idea of having a threaded Fork 
> > filter to be used together with Pipe: processing for each subsequent 
> > filter downstream of the threaded Fork filter is done in a separate 
> > thread of its own. This could potentially bring performance benefits 
> > (in theory) especially with a move on to having an increasing number of
> computing cores per CPU.
> > 
> > So I've emailed Jack separate to the list to get some of his opinions. 
> > The main points he raised were that:
> > 
> >  - The approach used for Fork would be most promising in terms of 
> > working with the current design and not forcing a full rewrite of the
> filter system.
> >  - He proposed defining a new Filter subclass Threaded_Filter which 
> > itself takes a Filter* as an argument which will spawn a thread and 
> > uses two message queues for I/O with the filter it manages.
> >  - When write() is called on the Threaded_Filter, it pushes it to the 
> > input queue, which the worker thread pulls off and write()s to the 
> > underlying filter.
> >  - With this approach the application can control concurrency very 
> > finely (perhaps too finely), since the application can specify which 
> > filters run on the main thread, and what runs on a secondary thread.
> > 
> > I was leaning to a slightly different approach: with the knowledge 
> > that Fork operations parallelise the easiest, I should adapt the Fork 
> > class to let each downstream filter run in a separate thread. This 
> > shares much commonality in its design with Jack's idea (especially in 
> > terms of thread sync and friends), provides a little less control over 
> > what runs in a separate thread, but comes with a slightly easier
> implementation.
> > 
> > Jack stated that I probably should email the devel list and solicit 
> > ideas, since everyone may have different expectations. I'll be 
> > implementing this in my spare time so I'd like to accept as many ideas 
> > and combine them before acting on it. I think that the pipe/filter 
> > design Botan has is intuitive, and I'd like to keep that as much as 
> > possible, without compromising on potential performance gains. 
> > Intuitively, both seem to run contrary to each other, but I think we 
> > can work something out here to an API that is both powerful yet easily
> mastered.
> > 
> > Any thoughts?
> > 
> > Regards,
> > Joel
> > 
> 
> 
> 
> > _______________________________________________
> > botan-devel mailing list
> > botan-devel at randombit.net
> > http://lists.randombit.net/mailman/listinfo/botan-devel
> 
> _______________________________________________
> botan-devel mailing list
> botan-devel at randombit.net
> http://lists.randombit.net/mailman/listinfo/botan-devel





> _______________________________________________
> botan-devel mailing list
> botan-devel at randombit.net
> http://lists.randombit.net/mailman/listinfo/botan-devel




More information about the botan-devel mailing list