[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: concurrent GC

yes OSF/1 is a full Mach (with horrible hacks inside to speed it up
even further, yeech) plus.. everything.  Hopefully the last kitchen
sink in OS history, certainly the most capacitous sink I've ever seen.
They are up to spec with our latest kernels, actually a bit further
along in bug fixes.  If you'll find incompats it is only going to be
on the U*x side [POSIX, for instance], not on Mach.  They have begun
shipping last January or so, BTW.

My point was that it is hopeless to think of Unix and true parallelism,
the SysV side actually blesses this deficiency and refutes threads
as unnecessary (heard from Ritchie and then over and over again).
BSD sits and waits.. for someone to donate them working code, I guess.
Which we *are* in the process of doing, BTW.
Most vendors do not have multiprocessors to sell, so they are just happy
with user-level fake parallelism, as in Lucid and Franz lisps.

There really is no choice.  And since you are forced to it, might as
well make the best out of it and put yourself on the front line, like
you did on the compiler front.  Which worked out quite well, right ?

As for the specific feature (the external pager), yes there was some
discussion two years ago as to whether it should have been included
in the first release.  Mostly Avie (NeXT) was suggesting they just
picked up what he ran on NeXTs [wonder why...], but fortunately it
was decided otherwise esp since database people strongly, very strongly
said they wanted it.  And IBM and HP listen carefully to databases ;-))
By now the OSF people have realized how much valuable it is.

As for BSD compat, assuming you really will still care about it in a year
from now, keep the code as it is, with conditionals that can bring it
back to vanilla U*x, of course losing in parallelism [there is still
around a version of CThreads implemented in a pure co-routine way
which would run on U*x, you might do the same if you wanted to.].
If you keep the sequential/parallel dependencies in the C side
as it is now you can also easily ship TWO systems to people [I
am referring here to the extra costs incurred so far in extra type
checks needed for futures: some people might want to be real sure
their sequential code does not pay for parallelism they do not use.
Rob, are you listening ;-)) ?].

Speaking of OSF/1..  I am still puzzled at what we, CS dept, will do
about it.  Howard seems to just want to "sit and wait" to see what
success it gets on its own, possibly a bit affected by this
not-born-here complex we all seem to undergo sometimes.  They (OSF)
would like to get it installed here as a preferred testing site, they
are much aware that we got where we are only thanks to "our friendly
users".  Sounds like we are headed for a very stupid chicken-egg stall.