Conference ulysse::rdb_vms_competition

try looking at existing notes in this conference that discuss Oracle at
some length....
Hope it helps.
nigel

    Hi,
    
    Sorry I have no pointers to Oracle on DECsystems either but my customer
    is also interested. They are currently testing Oracle on Pyramid and
    Sequent, we might try to test one of our SMP Ultrix machines too
    (Decsystem 8540). If I have any results that are worth mentioning, I
    will post them here.
    
    I'm curious wether Oracle will have problems with SMP on RISC as well
    as on VAX. Btw, how do they do that on Sequent as that machine has
    multiple processors too.
    
    Henny

    ORACLE has worked with us to do some preliminary testing of their
    products on our SMP RISC machines.  Our implementation platform
    left a lot to be desired (58xx are not exactly scaleable screamers).
    We will be following up with more SMP work in their labs.
    
    Mark
    

    A little clarification from anyone in the know might help clear up a
    lot of confusion over Oracle and multiprocessing and Sequent and
    DECsystems.  We have three basic combinations of Oracle and hardware
    platforms under discussion:
    
    1. Oracle running on VAXes running VMS, possibly with SMP and/or
    Vaxclusters
    
    2. Oracle running on Sequent Symmetry systems
    
    3. Oracle running on DECsystems running Ultrix, possibly (with 4.0)
    with SMP.
    
    
    Now we pretty much understand the problems with #1, either with respect
    to VAXclusters and Oracle's problems with DLM, or SMP and oracle's
    performance characteristics caused by being single-threaded on updates.
    
    2. What about Sequent?  I presume that they are running in an SMP-like
    mode, that they have not actually parallelized Oracle to run different
    parts of the Oracle code on different processors.  But do they still
    have the same problem with being single-threaded on updates?  if so,
    they how could they possibly claim such linear performance increases by
    adding processors to a Symmetry system?  True, a Symmetry can have more
    processors that we can, but if it is still single threaded on updates,
    it's going to level off pretty quickly in terms of performance.  OR,
    have they "fixed" Oracle so that it is no longer single-threaded on
    updates?
    
    3. So how does this all relate to Oracle on DECsystems with Ultrix
    4.0's SMP?  If the answer to #2 is that they "fixed" Oracle to
    bemulti-threaded on updates, did they do it only for the Symmetry
    engine, or did they do it for anything running UNIX?  would we benefit
    from any of that?
    
     Does anyone out there have any good info  on this, or pointers (I have
    enough riding on this to do some of the research) to finding the real
    answers?
    
    -reed 

    Note that Ultrix has supported MP systems even prior to V4.0.
    The application shouldn't care whether it's SMP or ASMP, or whether
    the maximum number of processors supported is two or more.  So,
    how well does Oracle run on Ultrix on a VAX 6320 or DECsystem 5820?
    
    

Bill, I don't know how you can make the ASMP doesn't matter comment since we
are dealing with processes doing I/O.  The whole point of the argument against
ORACLE is that they single thread all writes through one process.  Therefore,
if that process requires more than 1 processors worth of CPU, you have a bottleneck.
Which brings us to Reed's question.  The single-threaded nature of the 
database writing process is a THEORETICAL bottleneck, not necessarily one
that you would see on any given workload.  ORACLE can claim linear performance
improvement because it probably is linear for the workload they ran (TP1?).

Now, to get into the tradeoffs:

ORACLE V5 doesn't work right on SMP, so you need V6.
ORACLE V6 doesn't handle VAXclusters, so you need V5
You can't make full use of both VAXclusters and SMP combined.  Sorry.

ORACLE V6 should scale as well on DEC SMP systems, 58xx included, as on
Sequent given the following conditions:

1) There is no hardware bottleneck.  This is a hard one for me to answer.
I sure don't know enough about Sequent's design to comment on hardware
bottlenecks.  On our hardware, there are certain bottlenecks that can occur.
For example, the NMI Bridge that connected the 3rd and 4th processors on
an 88xx system.  On XMI systems, it is possible for workloads with relatively
low cache hit ratios to saturate the XMI as you add lots of fast processors.
Although I have no direct evidence to back it up, it would not surprise me in
the least if ORACLE saturated the XMI for the 3rd and/or 4th processors in a 
58xx system.

2) The quality of the SMP software implementations may vary.  Are there operations
that are not fully symmetric?  How big are the critical regions in the operating
system?  Etc.  The single ORACLE writer process may be an irrelevent part
of the discussion if the reality is that even multiple processes doing
I/O end up doing too much single threading through the O/S or 
context switching to one processor (ala ASMP).

Hal

    Re: .3  Mark, is Oracle going to keep their mouth shut about our
    DECsystem SMP performance or should we expect to hear them bad mouth
    it like they do VAX SMP machines ?
    
    Re: .4  They can (and do) claim anything they want when it comes to
    performance numbers !!
    
    I can tell you, based on what my Oracle customer told me, that my
    Oracle customer's batch benchmark (5 jobs) ran about 40% faster on the 
    6440 than it did on the big Sequent machine and the 6420 beat the small
    Sequent machine by a similar margin.  Given what I know about the
    Oracle V6 architecture and the fact that the VMS Oracle engine is NOT
    decomposed, it seem very unlikely that Oracle is doing anything special
    in the *x implementation of V6 detached processes.
    
    -- gerry

    Do we know what changed in V6 to make it "work on SMP" ?? Have been
    trying to think of what they could have done that would work better on
    SMP but not on VAXclusters.
    
    re .6: Can we get more info on the nature of your customer's benchmark?
    Would they be referencable by any chance (should we take this
    discussion off line?)
    
    thanks,-reed

    Try reading 324.1, 349.2, 349.4, and 424.8 for more information on
    Oracle, VAX SMP, and VAXclusters.
    
    -- gerry

    Gerry:  The ORACLE tests pointed out various areas that needed work.
    They have chosen to not make these tests known, and are not giving
    customers any indication of performance due to the problems.  We now
    have new SMP processors, and near-production quality Ultrix 4.0, and
    will rerun tests which should show signficant improvement.  Many of
    our potential commercial customers want to see commercial benchmark
    data (like debit/credit or TPC-A) from our RISC machines, but we
    keep throwing specmarks, aim, megaflops, ..stones, etc. which they
    don't want to hear about, at them.  I am working to convince the
    right folks that our RISC systems are missing the largest segment
    of the market (commercial), and we better get good commercial data
    inot the market asap.  If you know of any other RISC based commercial
    applications we can use to benchmark, please let me know.
    
    Mark