Re: [Caml-list] Threads and "transaction isolation" in OCaml

[Yaron Minsky <yminsky@janestreet.com>]

I'd like to second Markus' request here.  We definitely take advantage
of this behavior in the compiler, and I hadn't realized that it didn't
hold in byte-code.  I do think this is a rather valuable hack, and it
would be good to make the bounds of its applicability more explicit.

y

On Fri, Aug 16, 2013 at 12:07 PM, Markus Mottl <markus.mottl@gmail.com> wrote:
> On Fri, Aug 16, 2013 at 4:46 AM, Török Edwin <edwin+ml-ocaml@etorok.net> wrote:
>> I wrote a little test program and this is not true for bytecode -vmthread, but seems to work for bytecode -thread and native -thread:
>
> Thanks for this example.  It's not surprising that this trick does not
> work with VM-threads, but I have never seen anybody use them for
> anything so I'm not bothered by this.
>
> But, surprisingly to me, I did manage to get a failure with byte code
> + POSIX threads when running this example on Linux (Centos 6.4) via
> VMware Fusion (Mac OS X).  I admittedly have never thoroughly tested
> these "transactions" with any sort of byte code, since I practically
> exclusively run native code.  Still, I would have expected the runtime
> to work identically here.  Interestingly, I couldn't replicate the
> error on a VPS running the same OS and compiler, but using Xen (via
> Linode) instead of VMware Fusion for virtualization.
>
> Note that some open source libraries, e.g. the Nano_mutex module in
> Jane Street Core, depend on this trick so it seems important to figure
> out how safe and portable it really is.  It would be a huge pity to do
> without, since it can offer quite dramatic performance improvements
> for certain applications.
>
> Does anybody know a reason why byte code + POSIX threads might fail
> with this trick?
>
>> I'd consider this very fragile as you'd need to know the implementation details of every function you call (or only calling your own safe functions). I was surprised to find out that even Array.fold_left creates a temporary ref for example.
>
> It would certainly be unwise to call any function whose implementation
> is not fixed and fully understood wrt. transaction isolation.  This
> trick is obviously low-level and its use should be clearly documented
> by the developer.  Most of the time it will only consist of a few
> assignments or pattern-matches anyway.
>
> Regards,
> Markus
>
> --
> Markus Mottl        http://www.ocaml.info        markus.mottl@gmail.com
>
> --
> Caml-list mailing list.  Subscription management and archives:
> https://sympa.inria.fr/sympa/arc/caml-list
> Beginner's list: http://groups.yahoo.com/group/ocaml_beginners
> Bug reports: http://caml.inria.fr/bin/caml-bugs