Conference ulysse::rdb_vms_competition

640.0. "Oracle shock horror! (Bad intentions)" by CHEESE::KAISER () Mon May 07 1990 23:48

I'm not sure where to put the stuff below my signature, but this conference
(which I don't follow) seems from its title as good a place as any.  The article
is something that caught my eye in Usenet comp.os.vms; it made me choke and
gasp.  Am I crazy, or are these guys really willing to trash all other activity
on a VMS SMP system just to make their own software function?

It appears that they're planning to stop all but one CPU of an SMP group just to
take a snapshot of an important piece of their database activity.  Very ugly,
very dangerous.

---Pete
--------------------------------------------------------------------------------
From: [email protected] ("Mark Porter")
Newsgroups: comp.os.vms
Subject: consistent snapshot of shared memory
Date: 4 May 90 19:51:12 GMT
Sender: [email protected]
Organization: The Internet

Organization: Oracle Corporation
Location: Belmont, California

It would be really nice if there was a	way to get a consistent
snapshot of a global section which many processes (some of which are current on
other cpus) are modifying at a furious rate. The other processes are all at
normal process priorities, ipl 0.

What I need is a routine I can run which will put all the other cpus out of
commission for some arbitrary (perhaps LONG - 1-2 minutes) length of time
WITHOUT crashing the VAX.  Then, I will go up into kernel mode, go above IPL 2,
and get my nice consistent snapshot.  Actually, I'd already be in kernel mode
from pausing all the other cpus,  but you get the idea.	  Then, I'd let them
all go, and hopefully, everything would work. hahaha.

BTW, I realize that putting the other cpus to sleep will take some number of
instructions on my cpu, thus letting them potentially do some modifications to
the global section.  I'm willing to deal with this.  Ideally, I could stop the
other cpus within 100 instructions... Thus, mechanisms like having a subprocess
execute "stop/cpu/all" probably aren't fast enough.

Has anybody written such a beast?  What mechanisms does SMP offer to do this
stuff, if any?

Thanks in advance for any ideas!
Mark.

______________________________________________________________________________
Mark Porter			Internet: [email protected]
RDBMS Kernel Group		      or: maporter%[email protected]
Oracle Corporation		   Phone: (415) 358-3421

    If Oracle does succeed in implementing this functionality, they'll have
    a hard time selling it to customers - freezing the machine (for up to
    2 minutes!), does not go over well with customers - we've all heard how
    they grumble about cluster state transitions...

    Not only that ... it will give them a good chance to bait-and-switch
    the unsatisfied customer to another platform AND another chance to
    tell the customer how bad VAXen are and how Sequent handles this
    condition so much better.  
    
    I doubt that this is a sinister plot or anything of the sort and we
    really don't know who this guy is or what he does.  Heck, the whole
    thing could be a plant !!
    
    Remember, the Oracle folks are masters at moving the blame for their
    mistakes to someone else (Digital) and must never let them get away 
    with it.

    The guy sounds like bit-twiddler on Acid.