Re: Backporting DFly patches to FreeBSD?

[Matthew Dillon <dillon@xxxxxxxxxxxxxxxxxxxx>]

:I really don't think that's the way the cookie crumbles, as you state
:it. FreeBSD has very different design goals and I believe the project
:will stick to them. Again, as some developers have said, they would like
:to see some of these changes stabilize before importing them into
:FreeBSD.
:
:Plenty of people are interested to get some of the DragonFly BSD work
:into FreeBSD. Sascha's syscons changes are especially interesting for
:the FreeBSD project as they would be quite simple to import compared to
:some of our other advancements. I don't think that interest is a lacked
:factor in this equation; I think that developer time is. While the
:project has a relatively large number of developers, the number of these
:who actively contribute and understand the parts of the kernel that are
:affected is limited to the few who are already busy with several other
:projects.
:
:I would personally like to see some well-defined scalability tests
:(perhaps in-kernel hooks would be interesting to examine performance of
:some of our more obscure modifications) before I see any more stuff
:stating that we perform better than FreeBSD on the same hardware.
:
:I'd also like to see some people work on backporting changes to FreeBSD
:and less fingers pointed towards the developers who are actively working
:on the FreeBSD project. These are people who have families, jobs and do
:already actively contribute. It's unfair to ask these people to drop
:what they're doing and import DragonFly's changes. It's also
:impractical: one reason that DragonFly exists is to explore how these
:changes actually work in practice.
:
:I could generate pages of comparisons that would make it quite clear
:that these aren't always good ideas nor reasonable expectations, but I
:think that's clear by now.
:
:Kind regards,
:
:Devon H. O'Dell

    I have to say that I disagree with this assessment of FreeBSD, and I
    disagree with the assessment as to why some of the work hasn't been
    backported.  First of all, the stability argument can't be right.
    FreeBSD developers are committing all sorts of things to FreeBSD-5/6
    that are not only unstable, but *seriously* unstable and some
    of those issues, like the scheduler, have taken a year or more to fix
    (if indeed one could call it fixed now, which I doubt).  DragonFly's
    development isn't bug free, but the regimen is a lot tighter then FreeBSD
    development these days.  Most of the mechanisms that are still 
    backportable have been stable in DragonFly for over a year.  For most
    of the rest FreeBSD has simply waited too long and the divergence has 
    become too great for any single person, even a dedicated one, to backport
    the more interesting work.  So, for example, IPI messaging should be
    backported, no question about it, and it still can be.  It's an absolute
    requirement for them to backport it and consolidate all the myrid 
    ridiculously complex IPI mechanisms they currently have, but nobody is
    doing it.  I consider that a serious management failure on FreeBSD's part.
    There is certainly interest in doing some backporting, and for those
    developers not being able to do it is nothing more then a time constraint,
    but there is also a level of hostility to any backported work and a
    level scrutiny that goes way beyond the scrutiny applied to native work.
    Generally lots of 'its not proven' excuses go flying around pretty much
    ignoring the fact that similar FreeBSD development itself is just as
    unproven.  In many respects, I think the perforce model they are using
    has resulted in even more isolation of sub projects within FreeBSD, and
    it hasn't seemed to helped in the bug department for work that finally
    gets into the CVS tree.

    Personally speaking, I think we've proven our model in all aspects 
    except performance.  What we are doing is clearly far more maintainable.
    FreeBSD has a bit of a hidden beast problem in the maintainability 
    department.  When an original author takes a vacation or stops working
    on something, the maintainability problem hits them squarely in the face,
    but its hidden as long as the original authors continue working on the
    code.  That doesn't bode well for the future.  On the otherhand, there
    have been half a dozen instances where people have come in cold and
    done bug-free or mostly bug-free (meaning fixed in a day or two)
    work on core pieces of the DragonFly code base.  Without any prior
    instruction Jeffrey Hsu was able to thread and message most of the network
    protocol stack using my LWKT messaging primitives.  Joerg was able to do
    major cleanups of the namecache code with only one or two emails
    between us.  David Xu sent me a TLS patch out of the cold that adds
    code to the core LWKT switching assembly without any instruction.
    Richard Nyberg debugged a namecache issue out of the cold.  I consider
    these and other events as undeniably proving the maintainability of 
    our codebase.

    On the performance front... I absolutely refuse to rush into removing
    the BGL.  I don't give a damn what excuses the FreeBSD people are making
    with regards to unproven performance.  They are flapping their mouths
    a lot about unproven this and unproven that, but they aren't actually
    thinking theory and that is a serious mistake.  I want to get our
    codebase using mostly MP clean algorithms BEFORE I start actually 
    turning off the big giant lock.  So e.g. the threaded network protocol
    stacks are using MP clean algorithms, but they aren't 100% MP clean
    yet and so the BGL is still turned on.  We *KNOW* that those pieces
    which are now MP clean, stable, and well tested, are not likely going
    to be the cause of any bugs when we start to deal with the remainder
    down the line.  That's what is important.  Not only that, but since
    it is a threaded subsystem we have an ability to turn off the BGL on
    a thread-by-thread basis, and even do it on the fly with sysctls,
    which is a far saner development model then FreeBSD's 'oh lets throw
    mutexes around all this junk, turn off the BGL, and pray' methodology.

    There is definitely some UP performance degredation from e.g. threading 
    the network protocol stack, as one person's recent routing tests have
    shown.  But considering the fact that we haven't actually tried to
    *optimize* the messaging yet the numbers are pretty much in-line with
    what I would expect.  More to the point, the *theory* behind getting good
    performance out of a threaded subsystem is sound, primarily the 
    ability to queue more then one piece of work before switching threads,
    and I see such a clear path for optmization and improvement (without
    having to resort to 'hacks') that I am confident that we will be able to
    soundly thrash the mutex model when all is said and done.  Certainly,
    no matter what, we will come close, and that would be a win too
    considering the vast differences in maintainability between our code 
    and FreeBSD's. 

					-Matt
					Matthew Dillon 
					<dillon@xxxxxxxxxxxxx>