Roman Rokytskyy wrote:

>> Perhaps it is time for a thought pause, and see if other FB Arch
>> readers have good ideas or suggestions.
>>
>
> Fine. But before we stop, some more brain food (Warning: the language of
> the article is not very appropriate in girls college, but I hope that
> our list can digest it somehow :)
>
>

A more complete description can be found at
http://www.eecs.harvard.edu/~fedorova/papers/asplos165-damron.pdf.

The argument for hybrid transactional memory (Prius-mode?) is that a
software implementation is already available but hardware assist, if
available, can make it go faster. It's a nice argument, but it begs the
question of how much slower the software implementation that other
techniques. My take on their software algorithm is that it is probably
a order of magnitude slower than user mode read/write locks (aka
SyncObject). The fact that a hardware implementation does almost
nothing to change the equation (inequality?) since:

1. In general, Sparc servers are slow and very expensive
2. Sun's declining market share is split between Sparc, AMD, and Intel
3. None of Sun's installed base has support for hardware
transactional memory

Furthermore, Sun tends to release system software under the CDDL (Common
Development and Distribution License, an MPL-derivative) that is
incompatible with the GPL, so a HyTM implementation is unlikely to show
in in Linux.

The bottom line is that hybrid transaction memory is pretty much
guaranteed to hurt performance on the tiny slice of the computing world
where it might be available. This doesn't make it very attractive.

So thanks, but no thanks.

--
Jim Starkey
President, NimbusDB, Inc.
978 526-1376