Conference ulysse::rdb_vms_competition

632.0. "My observations of an Oracle Benchmark" by ZUDEV1::OLLODART (Expatriate) Mon Apr 30 1990 17:30

    I don't know whether this is the right place to discuss this, so if
    it isn't, Mr. Moderator, just delete the note.
    
    I had the "pleasure" of being the technical leader for a benchmark that
    Oracle did for a customer of ours on our machine. I am of course pro
    RDB and RALLY, so I watched the entire implementation of the benchmark
    like a hawk, in addition to getting myself an account to test out their
    tools. 
    
    I actually learned Oracle pretty fast, because we spent a whole week
    tuning the system.
    
    Some of my comments:
    
    1. Oracle version six will never be able to run in a customer because
    (please correct me if I am wrong), the DBWR process from Oracle that
    does all the updates to the DBMS doesn't have a clue what the DBWR is
    doing on another machine. An instance in Oracle (i.e. Create database
    statement) must be defined per machine and two machines can't go onto
    the instance at the same time. 
    
    2. In addition, you can't start more DBWR processes, so you will never in
    your life have SMP with writing, because any dimwit knows one process
    can only be scheduled to one cpu at any one time. 
    
    3.The reading on the other hand is done by the user process, so SMP is
    supported here.
    
    4. Version 6 is extremely CPU intensive. I have seen not only on this
    occassion, but at other V6 sites, that a small amount of users (15 on
    a VAX 6410 and 5 on 8350) have choked the CPU to no end. I could not 
    believe my eyes when a signal SQLforms form started 15 times with a
    Keystroke file choked the cpu on our 6410. NO IO's and NO memory. What
    is that thing doing ?. All the data was in the buffers, I can't imagine
    why all this processing is going on. I had to wait 2 minutes just to
    finish the login procedure when logging in. 
    
    (Oracle rep says "That is how we sell Oracle, cheapest resource, the
    cpu, isn't that great !) "GREAT, Really great, I just love those
    balanced systems. In fact, after a week of tuning we still couldn't get
    the system balanced.)
    
    5. I found Sqlforms to be really primitive. Version 6 finally came with
    Plsql, and ADL type SQL. I also found that programing joins between 
    relations was tedius and not elegant.
    
    6. RDMS$DEBUG_FLAGS in oracle is called the explain command, which 
    does not start the query. You have to :
    
    1. Create a table with about 10 attributes
    
    2. Fix up the sql statement with the explain command
    run the sql statement
    
    3. write a wierd looking sql statement to retrieve the information
    in a readable fashion.
    
    No cost or cardinality analysis, no BLR, etc. In addition, I don't
    think you can use it with any of Oracle layered products. Some help
    for transaction analysis. 
    
    Optimizer is not statistics based, rather rule based, so I don't know
    how good it works.
    
    
    7. Partitioning of tables. I had the Oracle rep sweating on this one.
    You can partition tables alright, just if you know how much data you've
    got.
    
    It works in that if you have 100 megabytes of data (you suppose to know
    how much data your going to have for the next year !) then you can
    split it up into 10 tablespaces @ 10 megabytes !
    
    No partitioning via attributes, rather pure physical distribution whereby
    each file is filed sequentially one by one. With RDB you can map via
    the field name and at least have insindental io scattering using hash
    indexes. And don't think you can just simply add a couple more table
    spaces when it gets short,  you have to define them new and copy the
    data from the old ones into the new ones. 
    
    8. Rollback seqments. You have to define the size and table space for
    your rollback segments. This is the RUJ of RDB, except you better know
    how big they are going to be for ALL types of transactions, because 
    if its too small, the system crashes. Nothing like being dynamic ehhh ?
    
    9. No two-phase commit and no transaction monitor like ACMS. I don't
    have a clue as to why they call it a Transaction processing option.
    There is nothing TP about Oracle at all. The closest they came is 
    optimizing their sql parser to work more efficiently by grouping
    statements and having some statments call low-level routines. I didn't
    find it that wonderful after we waited 30 minutes to retrieve 1280 rows
    from a join of three tables, Tabel 1 287'000 rows, Tab2, 14 rows, tab3
    5281 rows. AND Oracle was using indexes, except I don't think they have
    a zig-zag strategy.
    
    
    I have got all the scripts and the VPA databases from the benchmark if
    you are interested in having further information. 
    
    I don't know how they sell Oracle so well, technically speaking, it is
    not that great.
    
    Regards,
    
    Peter o.

    Not being an Rdb or Oracle person, I would like to see an Rdb
    "counterpoint" to this topic.  I am working with a customer on other
    items (MAILbus related) but am always hearing Rdb and "yecchh" used in
    the same sentence.  T'would be nice to have some ammo to fire back at
    them.
    
    Don Bevis