Want to hire me? For availability and scheduling please email info@arcepm.com

02 October 2014

Oralce OpenWorld, Day 5

The craziness continues, part the fifth

Hybrid Essbase:  Evolution or Revolution

Here I am, at the first session, watching Tim German and Dan Pressman present on Hybrid Essbase.  
Tim and I presented on this subject at ODTUG’s Kscope14 session.

Hybrid Essbase was released this year with Essbase 11.1.2.3.500 and really is a revolution in Essbase.  ASO on top of BSO (technically, I believe that it is really ASO-like but is so close in functionality it’s functionally equivalent) obviates the need for ASO reporting cubes, slow BSO calcs, BSO size, etc. really is going to change what we do with Essbase.  I am very, very, very excited about this tool.

Exalytics,  Essbase, and internals

Above are Kumar and Steve talking about some Really Cool Stuff.

You will see some of the Kumar Secret Essbase Sauce below as it relates to Exalytics.  Yes, I did ask if it was okay to blog this.

And remember that this is all safe harbored, so no promises, no dates, no nothing.  All of the below may happen but who knows.  Don’t hold them to any of this, and don’t make any plans (other than pondering it).
  • New X4-4
    • Chip is faster than anything you can buy as a commodity customer
    • X4-4, 2TB RAM, 60 cores max
    • The fewer the cores, the faster the clock
    • T5-8, 4TB RAM, Sparc
  • Oracle database
    • Run the Oracle database in RAM
    • 100x faster queries:  real-time analytics
    • Instantaneous query results using the database in memory in lieu of TimesTen
    • Cheap way of trying in memory db 12c
      • OBIEE 11.1.1.7 certification for 12c
      • Summary advisor
      • Aggregate persistence
      • Why use the db – skills, true relational query, good push to OBIEE, can be faster for simple reporting compared to Essbase
      • Can still be a normal data source
      • But, Essbase is still going to be faster when it comes to analytic processes
      • Think of it as an in-memory data mart
      • TimesTen still around
  • 2x increased transaction time, although Exalytics is not the target of transactional processing
  • HFM
    • Coming with PS4 (11.1.2.4)
    • Certified for Exalytics, both on the iron and VMs
    • No Linux HFM except on Exalytics
  • Essbase enhancements
    • Pure in-memory engine as calculation will no longer wait for I/O (background write)
      • Finish calc, allow user to use it, write happens in background later
      • Thread management and thread based memory allocation
      • Fundamental improvement, impacts resource consumption, stability, and performance
      • All aggregate views will stay in memory – once read from disk, it will never return to disk
        • This is only going to happen in Exalytics because they can know how much RAM the box has vs. commodity hardware with any amount of RAM
        • Where a particular task goes, from a CPU and memory perspective will be handled from a hardware and software level
    • Leverage X4-4 capabilities by improving scalability
    • No really firm release dates on all of the above – some in PSU4, others later, sometime
    • Lots of work with the hardware engineering
    • Performance
      • X4 improvements in BSO MDX grows as threads increase
        • Read-only
        • Mixed load (like Planning in queries and calcs happening simultaneously)
          • Bigger improvement in this scenario than read-only although that too has an improvement
          • X3 tends to bog down as this increases, X4 doesn’t thrash
      • T5
        • Similar work with Sparc engineering to improve performance on the Exalytics platform
        • Performance improvements almost 50% in .504 release
    • Why Exalytics?
      • Uptime
      • Volume
      • Iterations by users
      • ???
      • Pretty impressive improvements but of course variable results depending on application – YMMV
    • X vs. T
      • X – Destroyer
        • Faster with less load, but less capacity
      • T – Battleship
        • Slower with equivalent load, but faster when capacity level requirements are higher
    • Development model
      • Exalytics first
        • First developed in Exalytics, but released later to commodity
        • Faster for Oracle to write and test based on known hardware target
        • E.g., in-place block writes
      • Restricted
        • Performance potential (more cores, for instance) greater on Exalytics, so less restrictions on Exalytics, more on commodity
        • E.g., FixParallel
      • Only
        • This is Oracle’s USP
        • In-memory aggregate views because they know the amount of memory and have no idea what commodity has
    • Challenges
      • Symmetric multi-processing (SMP) guarantees uniform latency for CPU, but overall RAM is limited
      • Multi socket boxes needed for large amounts of RAM, but memory latency is not uniform
      • Complexity surfaces in software, not hardware
      • As databases get bigger, harder to keep in RAM
      • Essbase was not written for huge databases and RAM, but instead designed for smaller databases
        • As memory went over socket RAM, the speed in the socket is faster than out of the socket
        • Non Uniform Memory Access (last word right?)
        • THIS IS WHY RUNS GIVE DIFFERENT RUN TIMES
      • Two approaches
        • Use local socket ram in critical parts of the software by using thread affinity
        • Use padding to avoid false sharing – align important memory structures to cache lines
        • This is all only on Exalytics – commodity hardware does not get the above approaches
      • File sharing
        • Go google for this – this is an additional problem that Oracle handles on Exalytics
      • High CPU core challenges
        • Many users, fast and small workloads
        • Small users or single user with large highly parallelizable workload
        • Large and legacy code that runs sequentially but requires fasterCPU
        • Approaches (all part of 11.1.2.3.500)
          • Semaphores, mutual exclusion, and synchronization do not help
          • Any locking is bad and leads to poor CPU utilization
          • Lockless (for a block, for instance) algorithms based on Intel hardare instructions (compare and swap) were designed and implemented – chip specific instructions
          • Shared data structures, but cannot use typical semaphores – use lockless algorithms to reduce contention

CON7520  --  The Future of Oracle Business Intelligence and Oracle Essbase Integration

You see Mitch Campbell speaking, with the by-now-surely-well-known Steve Liebermensch and Gabby Rubin up on the dias.  Essbase is moving in new directions, it’s up to us if we will invest the time and training we’ll need to follow.

CON5532  --  Which Reporting Tool Should I Use?

Sob, this is the last session of OpenWorld.  It has been quite a whirlwind with meetings, meetups, old and new friends, and just general activity.  I am both sorry that it is over and oh so glad all at the same time.

Glenn Schwartzberg, aka My Man In California aka the-older-brother-I-wish-I-had-but-the-sentiment-is-not-reciprocated is speaking on selecting a reporting tool for EPM.  As always, good stuff, even if he doesn’t think so.  :)

What’s next?

And that’s the end.  Whew, I have blogged, tweeted, talked, texted, and just in general social mediaed myself into a tizzy.  I am ready for a break.

But I’ll be back in a little bit with (hopefully) some interesting technical stuff.  The learning and fun never stop – it would be a bit boring if it did.
 
 Be seeing you.


1 comment:

srx said...

Cameron this a great job you did and still doing, really nice coverage! But please buy a real camera (http://www.lesnumeriques.com/appareil-photo-numerique/gm1-p17408/test.html).

Also fully agree regarding what you said in previous paper : a PHD is required for OBIEE.