Solaris vs AIX: The people problem

Solaris vs AIX: The people problem

Summary: The people problem at the heart of this whole business of trying to predict data center costs under alternative Sun and IBM scenarios is simple and disconcerting: it's the people in charge, not the technologies, that really make the difference.

SHARE:

Please be aware, as I noted last week, that experience in both Sun and IBM environments has left me less than objective on the issue - and I'm therefore relying on readers to catch any errors I may make.

That said, I think that it is only possible to fairly compare the five year costs for a data center under alternative Solaris/SPARC and AIX/Power scenarios if you have a real data center staffed by real people, and therefore a real decision, to work from. There are two levels to this: first, IBM's pricing secrecy means that you can't get real numbers without first convincing IBM's sales people that you're ready and able to write the check; and, more subtly, operational differences, particularly with respect to applications and technology management, dominate all other long term cost components.

Given that Sun's hardware and software is generally cheaper, faster, and more versatile than IBM's, you'd expect Sun data centers to be generally cheaper and more effective than the IBM centric ones, but this isn't always true. In fact it's true only if that Sun center is run by people who understand what Solaris can do and how it should be used.

You don't often find that in larger data centers. What you find instead are decision makers who got their formative experience in data processing or Wintel and, correspondingly, often make counter-productive decisions about Solaris deployments - meaning, among other things, that the company they work for is often better served if senior management quietly accepts higher costs and reduced performance as the price of operational stability.

The big reason for this, I believe, is that someone who knows how to run Unix and does a good job at it becomes invisible to upper management. No crisises means no high profile crisis meetings and correspondingly no visibility to top management - with the unhappy consequence that the better you are at the job, the less chance you'll have of building the visibility and budget management record needed to promote up and out.

There's essentially no public data on this - but management assumptions about how to run the data center are far more important to long term costs than their specific hardware or software choices.

Consider, for example, the single most obvious difference between the IBM manager's One Right Way to do things and the Unix approach: the IBM guys always want to break wholes into walled off pieces where the Unix people naturally do the opposite: ignoring technical boundaries while crossing organizational ones to unite processes, people, and technologies.

Most people think of this terms of data processing's 1960s machine partitioning and virtualization solutions, but the bigger mirror for this is in the IT organization chart where the IBM manager spends most of his time, and most of the company's IT budget, keeping people from over stepping formal boundaries while the Unix guys act to unify the IT organization by cross training everyone for almost everything.

Look at one specific example of this: the data processing manager's assumption that splitting up resources improves utilization, from the perspective of a Sun sales manager and what you see is that the customer's willingness to spend five million bucks for the right to chop a 512GB, 72 processor E25K into a bunch of V490s makes the customer absolutely right to do so - but the bottom line for the company that customer works for is waste, inefficiency, and unnecessary applications run time costs.

You can see one consequence of this in some TPC/H results. Look closely at two Sun reports on the E25K and you'll see a large SMP machine with Oracle named user licensing at $1.35 million suffering only a 6% decline in the QpPH score between the 3TB and 10TB tests - and a cluster of ten dual AMD based X4500s getting to one third of the E25K's 3TB score for less than 10% of the hardware cost.

That happens because the E25K is designed for very large problems - things like aircrew scheduling or overall optimization within a Fortune 5000 class ERP - and applying it to problems that can be easily segmented wastes most of the value its customers pay for. Look at the IBM side on this, however, and you see an opposite pricing miracle - one that reflects both IBM's expectations about customer behavior and this same reality that the benchmark has been stable while the machines grew more powerful.

Specifically, IBM's 64 Power5, 128 core, p595 result of $53 per QpPH on the 3TB tests includes only $480,000 in Oracle processor licensing - while another IBM TPC/H filing shows 32 Model p570 Power6 boxes clustered around a Catalyst switch achieving an astonishing 343,531 QpPH on the 10TB test at a list price of $21.8 million.

What's going on is that the p595 runs up the highest hardware cost in the 3TB group despite not being designed as a large SMP machine - it's designed specifically to cope with mainframe workloads and built from up to eight eight way machines in the sure and certain hope that the people who buy it will partition it into smaller machines before use. As a result it achieves maximum processor utilization eight processors at a time - and 8 x $40K x 2 x 0.75 =$480,000.

So what does all this tell us? That all applicable combinations of small/big Sun/IBM machines can be used to deliver QpPH work, that management methods premised on using big machines where small ones can do the job cost a lot more money for the same result, and that management methods based on using just the right machine for the job are cheaper only if the gear comes from Sun.

The lesson in all this is subtle: align people with technology and both the IBM and Sun approaches work, but put a Sun guy in charge of that p595 and he'll want to his Oracle licenses to cover all processors; conversely, put an IBM guy in charge of that E25K and he'll want to break it into a bunch of virtual 490s - and, either way, the company they work for will get a lot less for a lot more.

And that's the people problem at the heart of this whole business of trying to predict data center costs under alternative Sun and IBM scenarios: it's the people in charge, not the technologies, that really make the difference - and because they'll always do is what they know how to do, failure to align skills sets with technologies before you buy means your costs will be about the same no matter what technology you hand them.

See: Part 4

Topics: Data Centers, IBM, Oracle, Processors

Kick off your day with ZDNet's daily email newsletter. It's the freshest tech news and opinion, served hot. Get it.

Talkback

9 comments
Log in or register to join the discussion
  • Here we go again...

    [i]
    Please be aware, as I noted last week, that experience in both Sun and IBM environments
    has left me less than objective on the issue - and I?m therefore relying on readers to
    catch any errors I may make.
    [/i]

    Seems like there again is a lot of catching to do...

    [i]
    That said, I think that it is only possible to fairly compare the five year costs for
    a data center under alternative Solaris/SPARC and AIX/Power scenarios if you have a real
    data center staffed by real people, and therefore a real decision, to work from.
    There are two levels to this: first, IBM?s pricing secrecy means that you can?t get
    real numbers without first convincing IBM?s sales people that you?re ready and able to
    write the check; and, more subtly, operational differences, particularly with respect
    to applications and technology management, dominate all other long term cost components.
    [/i]
    Again you are not being fair. You can not get SUN prices for their High end servers,
    like the M8000, M9000 and E20 and E25, only up and including their midrange.
    For IBM pSeries servers you cannot get prices for their high end(enter price servers)
    the p57X and the p59X servers, only up and including their midrange, p50X,p51X,p52X,
    p55X,p56X.
    So basically you start out by yet another broadside of how bad IBM is, and how good
    SUN is without any documentation. Sad..

    [i]
    Given that Sun?s hardware and software is generally cheaper, faster, and more versatile
    than IBM?s, you?d expect Sun data centers to be generally cheaper and more effective
    than the IBM centric ones, but this isn?t always true. In fact it?s true only if that
    Sun center is run by people who understand what Solaris can do and how it should be used.
    [/i]

    Again without any analysis or comparison you conclude that SUN hardware and software is
    cheaper, faster and more versatile than IBM's. Not very serious.

    I've also meet a lot of skilled Solaris people, I don't count myself as one as I know
    what to compare myself against :)=

    [i]
    You don?t often find that in larger data centers. What you find instead are decision
    makers who got their formative experience in data processing or Wintel and, correspondingly,
    often make counter-productive decisions about Solaris deployments - meaning, among other things,
    that the company they work for is often better served if senior management quietly accepts
    higher costs and reduced performance as the price of operational stability.
    [/i]

    Yes I agree on that, all to often you see a company that downsizes a Unix installation to
    a Winintel installation, to save money. And what they are doing is moving money from
    the hardware account, to the account that pays for real estate, people and software licenses.
    From investment to variable expenses. But then again no Chief Information Officer wants to
    run an IT department with a small staff, a bit investment budget, and a little budget for
    variable expenses.

    [i]
    The big reason for this, I believe, is that someone who knows how to run Unix and does a
    good job at it becomes invisible to upper management. No crisises means no high profile
    crisis meetings and correspondingly no visibility to top management - with the unhappy
    consequence that the better you are at the job, the less chance you?ll have of building
    the visibility and budget management record needed to promote up and out.
    [/i]

    Jup. Don't disagree on that. Not many in top management know about the success full system
    that haven't been down in years and the people who are responsible for it. But the crappy
    system that have been in trouble many times, they sure know that one.
    Bad publicity is better than no publicity, and you do look better if you solve problems,
    rather than if no problems occur.

    [i]
    There?s essentially no public data on this - but management assumptions about how to run
    the data center are far more important to long term costs than their specific hardware
    or software choices.
    [/i]

    Yes, I agree.

    [i]
    Consider, for example, the single most obvious difference between the IBM manager?s
    One Right Way to do things and the Unix approach: the IBM guys always want to break
    wholes into walled off pieces where the Unix people naturally do the opposite: ignoring
    technical boundaries while crossing organizational ones to unite processes, people,
    and technologies.
    [/i]

    What are you talking about ? Should we all form Unix militias ?

    [i]
    Most people think of this terms of data processing?s 1960s machine partitioning and
    virtualization solutions, but the bigger mirror for this is in the IT organization
    chart where the IBM manager spends most of his time, and most of the company?s IT budget,
    keeping people from over stepping formal boundaries while the Unix guys act to unify
    the IT organization by cross training everyone for almost everything.
    [/i]

    Again what are you talking about ? Looks like you are trying to draw up a picture
    where it's UNIX (which equals SUN in your word) vs IBM. Well it might have looked
    like that 30 years ago from your perspective. But hey ...

    [i]
    Look at one specific example of this: the data processing manager?s assumption that
    splitting up resources improves utilization, from the perspective of a Sun sales
    manager and what you see is that the customer?s willingness to spend five million
    bucks for the right to chop a 512GB, 72 processor E25K into a bunch of V490s makes
    the customer absolutely right to do so - but the bottom line for the company that
    customer works for is waste, inefficiency, and unnecessary applications run time costs.
    [/i]

    Well you think that using Domains on a E25K is a waste, well ok. But it has been one
    of the major selling points since the E10K.

    [i]
    You can see one consequence of this in some TPC/H results. Look closely at two Sun
    reports on the E25K and you?ll see a large SMP machine with Oracle named user
    licensing at $1.35 million suffering only a 6% decline in the QpPH score between
    the 3TB and 10TB tests - and a cluster of ten dual AMD based X4500s getting to
    one third of the E25K?s 3TB score for less than 10% of the hardware cost.
    [/i]

    Jup the TPC/H benchmark has a relative small memory footprint compared to the number
    of processors you need, and it scales in a way that makes it suitable for running
    on clusters.

    [i]
    That happens because the E25K is designed for very large problems - things like
    aircrew scheduling or overall optimization within a Fortune 5000 class ERP -
    and applying it to problems that can be easily segmented wastes most of the value
    its customers pay for. Look at the IBM side on this, however, and you see an
    opposite pricing miracle - one that reflects both IBM?s expectations about
    customer behaviour and this same reality that the benchmark has been stable
    while the machines grew more powerful.
    [/i]

    Besides the marketing praises of an E25K, and your statement that IBM makes
    pricing miracles aka they almost cheat with their prices...

    [i]
    Specifically, IBM?s 64 Power5, 128 core, p595 result of $53 per QpPH on the 3TB
    tests includes only $480,000 in Oracle processor licensing - while another
    IBM TPC/H filing shows 32 Model p570 Power6 boxes clustered around a Catalyst
    switch achieving an astonishing 343,531 QpPH on the 10TB test at a list
    price of $21.8 million.
    [/i]

    Yes, your point ? The way that Oracle licensing is constructed gives an
    advantage to POWER5/6 over SPARC with regards to cost of licenses, due to the high
    performance of the Chips.
    The 32 times 4 way p570 with DB2 EEE is IMHO way to expensive in software cost, if
    you actually read the full disclosure you'd notice that even the 4 way boxes are
    partitioned in DB2 so it runs as a 256 node IMHO a totally mad benchmark :)
    But besides from the fact that you can run both BIG SMP workloads and distributed
    network connected workload which is managed from a single DB2 management interface,
    (The network is the computer you know :)) what is your point ?

    [i]
    What?s going on is that the p595 runs up the highest hardware cost in the 3TB
    group despite not being designed as a large SMP machine - it?s designed
    specifically to cope with mainframe workloads and built from up to eight
    eight way machines in the sure and certain hope that the people who buy
    it will partition it into smaller machines before use. As a result it achieves
    maximum processor utilization eight processors at a time -
    and 8 x $40K x 2 x 0.75 =$480,000.
    [/i]

    Ok here comes the point, an almost clever one. First a statement that a p595 is
    not a pure SMP machine. But it is designed to run lots of smaller workloads
    which are partitioned into 16 way chunks.

    And a that the p595 is the most expensive server in the 3TB category.
    Jup 17KUSD more expensive than the E25K@1.8GHz or staggering 0.5% more if you want it
    in percent. Furthermore the E25K submission is from April 2007 and the p595
    is from September 2005. So do compare current systems to current..

    And then there is a list price calculation that shows that the oracle enterprise
    edition list price for an 16 way partition is 480K$ the same as what is listed
    as the price for the whole p595 in the benchmark,
    hence implying with your Miracle pricing statement, that the actual cost of
    an IBM solution will be many times greater than what's listed in the benchmark.

    The real deal is that the E25K and the p595 in the 3TB TPC/H benchmarks both
    use oracle and they both list a price of 10K pr Chip. They both take the
    number of cores and multiply by 0.75. Hence the 108 for the 144 and the 48 for
    the 64 way. It's very simple, Oracle have a strange price calculation that lets
    them charge more for Solaris customers to get the same work done, than for AIX
    customers.
    That's not IBM pricing miracles, it's Oracle exploding their Solaris customers.
    And perhaps you should bark at Oracle not the AIX people out here in the world.

    Another almost smart move is the way you described the E25K as a SMP machine
    and the p595 as a box made to be carved up.

    It shows that you don't understand micropartitioning and that U don't understand
    the way a p595 and a E25K is constructed. Because the truth is the exact opposite.

    The E25K is a board based machine, hence you'll get good very good performance as
    long as you don't do to much cross board communication. The speed of the Fire backplane
    is afair 166MHz. The inter MCM speed of the p595 is half the CPU speed, hence 1.15Ghz
    of the POWER5+ p595, and the intra MCM speeds is at full CPU speed hence 2.3Ghz.

    That's why you don't see any TPM/C, SAP 3 tier or Oracle APP benchmarks with a high-end E25,
    it's just better at running workloads that fit into a board, which btw. most workloads do.

    And before you start to talk about SPECJBB, then compare the 1 JVM to the 72 JVM
    result for the E25K.

    1149100 specJBB's for one JVM
    2105264 specJBB's for 72 JVM's
    same number of processors etc.

    And that's for a workload that does no IO, fits in memory and where SUN controls
    the whole software stack, and thus should have all the advantages they possible
    can get.
    And it still shows that running as a pure full system image with one app running
    equal across all boards will only get you 54.5% compared to a single os image with
    a partioned application, now that isn't impressive.

    Links:
    http://www.spec.org/osg/jbb2005/results/res2007q2/jbb2005-20070507-00295.html
    http://www.spec.org/osg/jbb2005/results/res2007q2/jbb2005-20070508-00296.html

    The p570 that you mentioned, also has a benchmark with one 16 core image and 1
    244361 specJBB's for one JVM
    224200 specJBB's for eight JVM's

    it has no where the same interconnect as the p595, but still gets a 92% in a
    single app image compared to a partitioned one.

    Don't get me wrong the E25K is a really great machine, the p595 is just better,
    when it comes to running a single system image vs distributed workload.

    And you do use micropartitioning on a p595, no matter if you run only one 64 core
    image or 2 x 32 core or... it is always partitioned.

    What you don't get is because you can carve up a p595 into micropartitions, it
    will have to have good single system performance.
    It's actually quite simple, to figure out isn't it ?

    It's very Carl Rove like of you to attack the strongest points of your 'opponent'.

    [i]
    So what does all this tell us? That all applicable combinations of
    small/big Sun/IBM machines can be used to deliver QpPH work, that management
    methods premised on using big machines where small ones can do the job cost
    a lot more money for the same result, and that management methods based on
    using just the right machine for the job are cheaper only if the gear comes from Sun.
    [/i]

    *BOGGLE* Damn what conclusion, SUN marketing couldn't have written it better.
    Nothing in of what you have written supports that conclusion.

    [i]
    The lesson in all this is subtle: align people with technology and both the
    IBM and Sun approaches work, but put a Sun guy in charge of that p595 and
    he?ll want to his Oracle licenses to cover all processors; conversely, put
    an IBM guy in charge of that E25K and he?ll want to break it into a bunch
    of virtual 490s - and, either way, the company they work for will
    get a lot less for a lot more.
    [/i]

    You forget that the licenses for the p595 and E25K both covered all the
    cores in the machine, in the TPC/H benchmark, and that they used the
    same Oracle price and reference.
    Comparing that price to the WEB price of Oracle 10G Enterprise ed. for 16 cores,
    and then conclude that cause it's the same as the one used in the benchmark
    then IBM does pricing miracles. You are being unethical.

    What makes the p595 good at consolidating workloads is micropartitioning,
    you don't have to carve it up, it runs almost as fast if you run it in a
    single system image, so it's actually the other way around.
    Try to ask a Solaris admin and an AIX admin to host a 1M TPM/C 0LTP Oracle
    workload on his/her E25K or p595 machine.
    Your Solaris admin will start talk about Oracle RAC or buying a 48 core E69K,
    your AIX guy will just give you a partition with 16 Cores and 512GB memory,


    [i]
    And that?s the people problem at the heart of this whole business of trying
    to predict data center costs under alternative Sun and IBM scenarios: it?s
    the people in charge, not the technologies, that really make the difference
    - and because they?ll always do is what they know how to do, failure to
    align skills sets with technologies before you buy means your costs
    will be about the same no matter what technology you hand them
    [/i]

    It _is_ the technology, you just don't want to put the time and effort it takes
    to understand other vendors than SUN's technology, sure it's also the people,
    and their ability to exploit the technology.

    And your answer to my answer will be.. you are all wrong....

    // Jesper
    JesperFrimann
    • Prices

      We made a comparison between the current M5000 and p570 (POWER6).

      The p570 was almost 50% more expensive than the M5000. Service & Support from IBM was also about 20% more expensive. This included all discounts we could get (both sides knew about each other).

      Because our MS-Windows-type-manager was afraid of the Solaris Container concept ), we went down the IBM road...

      He was looking for a "one application-one LPAR solution", just like in the good old windows world.
      Burana
      • Expensive comparison

        [i]
        We made a comparison between the current M5000 and p570 POWER6).

        The p570 was almost 50% more expensive than the M5000. Service & Support from IBM was also about 20% more expensive. This included all discounts we could get (both sides knew about each other).
        [/i]

        Yes but the p570 is also a fair bit faster than the M5000.
        There really isn't much performance data available on the M5000,
        but on it's big brother the M8000, but with 2.4GHz , not the
        2.1 GHz processors there is.

        32 core/16 Chip/16 Socket M8000 @2.4 GHz does
        36570 SAPS
        440207 SPECJBB2005

        16 core/8 Chip/8 Socket p570 @4.7 GHz does
        40070 SAPS
        691975 SPECJBB2005

        and a good estimate on a M5000 would be
        ( M8000 number /1.8(for scaling) * 2.4/2.15Ghz)

        16 core/8 Chip/8 Socket M5000 @2.15 GHz could do.
        18200 SAPS
        219085 SPECJBB2005.

        Well comparing the p570 to a M5000 is.. well a bit expensive,
        unless you expect to save a lot of money on software licenses,
        aka Oracle. Eg a db2 license is much cheaper on a quad core box.

        the
        16 core/4 Chip/4 Socket p560Q @1.5Ghz sounds more like a match
        18300 SAPS
        226291 SPECJBB2005

        And the p560Q is also a cheap box.
        8 core 1.5GHz 16 GB RAM, 2 boot disks 4U space -> 38882 USD
        compared to the M5000
        8 core 2.1GHz 16 GB RAM, 2 boot disks 10U space -> 84000 USD

        Links
        http://www-03.ibm.com/systems/p/hardware/midrange/560q/browse.html
        and
        http://shop.sun.com/is-bin/INTERSHOP.enfinity/WFS/Sun_NorthAmerica-Sun_Store_US-Site/en_US/-/USD/ViewStandardCatalog-Browse?CategoryName=HID-240460404&CategoryDomainName=Sun_NorthAmerica-Sun_Store_US-SunCatalog



        [i]
        Because our MS-Windows-type-manager was afraid of the Solaris Container concept ), we went down the IBM road...

        He was looking for a "one application-one LPAR solution", just like in the good old windows world.
        [/i]

        Well partitions do give you a better isolation, but that is only
        valid if you need it to address issues like:
        Software stack issues
        Different landscapes like test development education etc.
        Security issues, back end apps, web servers hence different
        security zones.

        I've never had with anything to do with a customers that ran
        with production system in containers.

        But containers, fit in nicely between running things inside one
        OS instance and LPARS. But they don't address all issues IMHO.

        But you can use workload partitions on AIX 6.1, which is similar
        to containers.

        // Jesper
        JesperFrimann
        • 4.7GHz

          Oh, they do deliver the 4.7GHz Power6 by now?

          Of course we compared the performance, we even put 6 CPUs into the M5000
          instead of 4 CPUs for the p570. The price difference was still around 30% more
          expensive. And I doubt that POWER6 is 50% faster than SPARC64 VI.

          You really can't compare the p560Q to the M5000 (RAS features).

          Oracle licenses were BTW a non-issue (had enough named user licenses).

          In another company I worked, we were running the core-banking application in a
          zone without problems. We did this with the second release of Solaris. Everything
          went smoothly...
          Burana
          • Only seen one yet..

            [i]
            Oh, they do deliver the 4.7GHz Power6 by now?
            [/i]
            Well did a solution with one a few month ago, no prob getting
            that one. But hey that's not volume.

            [i]
            Of course we compared the performance, we even put 6 CPUs into
            the M5000 instead of 4 CPUs for the p570. The price difference
            was still around 30% more expensive. And I doubt that POWER6 is
            50% faster than SPARC64 VI.
            [/i]

            You've gotta be kidding it not 1.5 times faster.. it more like
            3 times, depending on your workload. Sure if it's single threaded, VMT prob isn't as effective as SMT, the difference is
            smaller 21.6 vs 10.8 on Specint2006 for 4.7Ghz POWER6 (p570) vs
            SPARC64 VI 2.15GHz(M4000)
            But at 8 Cores the difference grows to 240 vs 81.6, not that
            Specint have anything to do with business apps.
            But looking at serious benchmarks like SAP std app, power6 looks
            frightening fast. And that's still running 5.3 and not using
            features like the Decimal Floating Point Unit, which practically
            no software supports yet.
            I also talked to the guy who build the solution I designed on
            the p570 I mentioned before, and he was impressed.

            [i]
            You really can't compare the p560Q to the M5000 (RAS features).
            [/i]

            Oh you can't ?
            You do know that a p560 is the same box as a p570, just with
            feature restrictions. Which btw. is a pain and that You could
            have my help in flaming IBM for that, at any time.

            I mean even an old bugger like POWER2 can go for years on mars ;)

            [i]
            Oracle licenses were BTW a non-issue (had enough named user licenses).
            [/i]
            That is always nice :)
            [i]
            In another company I worked, we were running the core-banking
            application in a zone without problems. We did this with the
            second release of Solaris. Everything went smoothly...
            [/i]
            Now that is a good thing.

            // Jesper should be working.
            // Jesper
            JesperFrimann
          • Performance

            3x faster?

            I'm not really sure about that one. There was a comparison between Ultrasparc IV
            + and Power5+ for a specific business application (long running batch job/couple
            of hours), for another bank I know. The difference was a couple of minutes,
            between the systems. No real differentiation between those system for this kind of
            job... :-)

            What I've heard about POWER6 is a performance boost of about 1.5x compared to
            POWER5+ for single threaded applications (quite reliable sources). I guess this
            difference will be about the same between UltraSPARC IV+ and SPARC64 VI.

            Don't get me wrong about the M-Series... I'm not a huge fan of these systems.
            They are too little, too late, and my guess is Sun is aware of it. But Sun had to give
            us something between UltraSPARC IV+ and ROCK (canceling USV was IMHO no
            mistake).

            Even if POWER might be faster at the moment, IBM has more hurdles...

            On the lower end, IBM really has nothing to deliver for running AIX, such as the
            Niagara1/2 systems. Add this to the fact that AIX is not available for X86, this
            could be the beginning of a _very_ slow dying of AIX.

            Currently (Open)Solaris just has the greater momentum. The only problem Sun has
            with Solaris X86 is to make the customers believe they really stand behind it (this
            time...).

            Here Sun has a lot in common with IBM. IBM tells you they stand behind Linux, but
            they really want to sell you AIX :-)
            Burana
          • Heh..

            [i]
            I'm not really sure about that one. There was a comparison
            between Ultrasparc IV + and Power5+ for a specific business
            application (long running batch job/couple of hours), for another
            bank I know. The difference was a couple of minutes, between the
            systems. No real differentiation between those system for this kind
            of job... :-)
            [/i]
            Jup, sometimes the math don't add up. Often it's cause there is
            a bottleneck somewhere in the solution stack that you cannot
            move, or the app just isn't
            Was in a project many years ago where the customer moved away
            from Alpha to another platform. And they had a database that
            kind of acted as a Message broker, and it ran on a 4 core box
            I had sized it to run on a dual core partition. When we moved
            the app to the 2 core partition, it ran with half the speed,
            of what it did before. We then added 2 more cores to the partition, rebooted etc. and ran a test. Now it ran 3 times
            faster than it did on the 4 core alpha server. The whole
            application complex was written to run on a 4 core box.

            But from what the Unix dudes I design solutions for say the
            SPARC64 VI boxes are more or less equal to the POWER5 boxes
            in performance, not POWER5+.
            Being an architect I don't get to do much hands on anymore,
            but I do get to see a lot of official sizing tables for
            different workloads (not the marketing stuff).

            [i]
            What I've heard about POWER6 is a performance boost of about
            1.5x compared to POWER5+ for single threaded applications
            (quite reliable sources). I guess this difference will be about
            the same between UltraSPARC IV+ and SPARC64 VI.
            [/i]

            Yes, sound about right, with regards to power6, but SMT should
            really have improved. 1 thread on 1 core gives you a specint of
            21.6 where as 4 threads hits 60.9. That is almost a factor of
            3 by doubling the cores and turning SMT on.
            POWER4-6 have been about throughput computing but with excellent
            single threaded performance.
            With regards to SPARC64 VI vs IV+, I think that it depends on
            the application, and on some workloads it's only the difference
            on the clock that makes the 64 goes faster.
            I think SUN was a bit disappointed with Fujitsu.

            [i]
            Don't get me wrong about the M-Series... I'm not a huge fan of
            these systems. They are too little, too late, and my guess is
            Sun is aware of it. But Sun had to give us something between
            UltraSPARC IV+ and ROCK (canceling USV was IMHO no mistake)

            Even if POWER might be faster at the moment, IBM has more hurdles...
            [/i]

            Well, I don't agree they are gaining marked share up to and in
            the middle of a product replacement. Think they gained 6 points
            in pSeries sales in Q3. And that is with a lineup where many
            of the boxes are looking kind of tired in my eyes. But then if
            you compare the boxes to what the competition can come up with.
            I mean it's not like anybody is leapfrogging power5+.
            And power isn't just faster.. it's much much faster. And I just
            can't stop loving micropartitioning,

            [i]
            On the lower end, IBM really has nothing to deliver for running
            AIX, such as the Niagara1/2 systems. Add this to the fact that
            AIX is not available for X86, this could be the beginning of a
            _very_ slow dying of AIX.
            [/i]

            Well I wouldn't call niagara boxes low end, they are expensive
            in the versions you see benchmarked. And the smaller cheaper
            versions are still much more expensive than the pSeries low end
            boxes.
            For example
            a T5120 with 4@1.2GHz cores and 4 GB RAM is 13.995USD.
            a p505Q with 4@1.65GHz cores and 4 GB RAM is 6.551USD.
            now the T5120 has 2x146GB disks the p505 has 2x73GB that is
            an extra 608$to the p505
            the T5120 has 1 year next business day support the p570 3 year
            next business day support. Upgrade to 3 years is an extra
            4.392 USD. for the T5120.
            Now... for two low end systems, as you yourself defined them
            with same performance, and equal support, space and options.
            T5120 is 24387 USD.
            p505Q is 7159 USD.

            You can verify my numbers on IBM and SUN's web pages.

            And you claim that pSeries have nothing in the low end ?
            Niagara based servers are expensive and not all workloads will
            run smoothly on them.

            But that said, I think they a fresh breath of air, and from what
            I've heard they are great as web servers and for workloads that
            require many light threads. So a really great product, but not
            one that well suited for all workloads.

            As for AIX dying, we are all dying slowly :)= U forget that AIX
            sales are still growing.

            [i]
            Here Sun has a lot in common with IBM. IBM tells you they stand
            behind Linux, but they really want to sell you AIX
            [/i]
            Well... SUN did, on purpose or not I don't know, give SCO the
            financial capital to sue IBM for putting UNIX code into Linux.
            Sure microsoft also put money behind SCO, but it started with
            SUN.
            http://www.eweek.com/article2/0,1759,1492533,00.asp

            There is a big difference between trying to sabotage the spread
            of Linux and then rather wanting to sell AIX than Linux.

            // Jesper
            JesperFrimann
    • umm..

      1) I have some basic price comparisons coming tomorrow - but only on the current cmt vs power6 stuff.

      The M series machines aren't made by Sun - that's why their pricing isn't on the store.sun.com site. When they were introduced even SUn's sales people didn't know what they cost.

      2) I had a great deal of trouble writing this blog-isode. In fact, I reworked it more often than any other so far - in large part because the business of what works and what doesn't is ill defined and the corelation to the tpc/h results is subtle (one try at this ran over 3000 words). i.e. I understand why you don't understand and admit that's my fault - but your counter argument doesn't even begin to make sense in large part because you don't get what you're countering.

      I'll be taking another run at this sometime - and maybe then we can both be clearer.
      murph_z
  • One Great Takeaway

    This is single best observation in the essay. You can work with a crack team or hot technology or both--none of it will matter if management lacks the right stuff:


    "There???s essentially no public data on this - but management assumptions about how to run the data center are far more important to long term costs than their specific hardware or software choices."
    jcawley