SQL Server Performance

New Dual Core Itanium 2 plans

Discussion in 'SQL Server 2005 Performance Tuning for Hardware' started by joechang, Jul 19, 2006.

  1. joechang New Member

    For several years, with the introduction of the IBM Power4 architecture,
    followed by the Power 5, IBM ruled the top of the TPC-C,
    and often by very large margins.
    The Power line runs IBM AIX, their version of UNIX,
    IBM TPC-C results frequently used UDB2, occassionally Oracle

    In part, these results were achieved with massive 512-bit wide FSB,
    special cache-cache busses, expensive multi-chip modules,
    basically stuff that makes the platform very expensive,
    but yields superior (nearly linear) scaling into the high SMP regions

    In this time, HP was able to generate TPC-C result on Unix / Oracle
    somewhere in the range of 20% higher than on Windows/SQL
    but did not publish the Unix result because of the negative
    comparison with IBM Power

    just today or so, HP finally beat IBM at the high end,
    but look for IBM to reply

    The HP Superdome in the test had 64 Itanium 2 9050 sockets
    thats 128 physical cores + Hyper-threading - meaning 256 logical processors
    Windows will only work with 64 logical procs, so don't expect a Windows result

    Still, a 32-socket SuperDome with no HT
    or even a 16 socket with HT is still a very powerful system
    Any one interested?

    recent Oracle/Unix/Itanium results
    |Vendor|System| Processor| memory| disks| tpm-C|
    |HP| RX6600| 4 Itanium 9050 DC 1.6GHz| 192GB| 504-15K+42-15K| 359,440|
    |Fujitsu|PrimeQuest540| 16 Itanium 9050 DC 1.6GHz| 1024GB| 2144-15K| 1,238,579|
    |HP| SuperDome| 64 Itanium 9050 DC 1.6GHz| 2048GB| 4200-15K+2856-15K| 4,092,799|
    |IBM| p5 595| 32 Power5+ 2.3GHz| 2048GB| 6760-15K| 4,033,378| UDB

    Windows
    |Vendor|System| Processor| memory| disks| tpm-C|
    |HP| RX6600| 4 Itanium 9050 DC 1.6GHz| 192GB| 756-15K+8-10K| 344,928|

    ___________________________________________________________________
    Any one seriously considering the new Dual Core Itanium 2 #90xx
    is invited to discuss their thoughts here

    Lets stick with hard facts, ie, test results that have been fully scrubbed,
    there is a lot of talk not based on fact,

    Yes, i know the full platform is expensive relative to Xeon/Opteron platforms,
    but it was meant to be a high-end platform,
    so certain design decisions precluded a low cost platform
    (the 128-bit bus requires a motherboard with more layers etc)
    There are also special design requirements to meet detected & undetected soft-errors
    in multi-processor systems that is not considered in the Xeon/Opteron lines

    does anyone recall the 1998-99 issue where big Sun servers would mysteriously fail
    (whatever the Solaris equivalent of a blue/black screen is)
    the quick fix was replacing the processor module with a new module having mirrored cache.
    this is a soft-error problem, more prevalent in multi-processor systems, especially if
    the system is at high altitiude (Denver, Sante Fe, etc)

    On the matter of soft-error rate (SER) memory/cpu failures
    I would like to ask any one with a big server (>4CPU, >8 DIMMs) in a high altitude city
    comment on whether they have notice occasional unexplained blue screens or other data
    corruption

    Denver 5,260 ft
    Sante Fe 6,989 ft
    Lhasa, Tibet/China 12,087 ft
    Wenzhuan 16,730 ft
    La Paz, Bolivia 11,913
    Leadville, CO, 10430
    Cuzco, Peru 11,152
  2. gunter_zink New Member

    http://tpc.org/tpcc/results/tpcc_advanced_sort.asp?PRINTVER=false&FLTCOL1=tpcc.c_server_procs&FLTCOLOPR1=%3D&FLTCHO1=4&FLTROWOPR1=AND&FLTCOL2=tpc_dbms.db_company&FLTCOLOPR2=CONTAINS&FLTCHO2=Microsoft+++++++++++&FLTFREEFRM2=off&ADDFILTERROW=&filterRowCount=2&SRTCOL1=tpcc.c_tpmc&SRTDIR1=DESC&ADDSORTROW=&sortRowCount=1&DISPRES=100+PERCENT&include_withdrawn_results=none&include_server_cpu=ON
  3. joechang New Member

    well this is amazing, to have gunter join this forum

    the short version of the url above is:

    HP Integrity rx6600 4xDual Core Itanium 2 1.6GHz, 192GB, 344,928 tpm-C, $2.24/tpm-C
    HP ProLiant DL585-G1 4xDual Core Opteron 2.6GHz, 128GB, 213,986 tpm-C, $2.13/tpm-C
    Fujitsu Primergy TX600 S3 4xDual Core Xeon 3.0GHz, 64GB, 188,761 tpm-C, $5.70/tpm-C

    bare system pricing, no memory?
    HP Integrity rx6600 4xDC It2 1.6GHz/24M$43,845 (ask for 23% discount for full system+storage)
    HP Proliant DL585 4xDC Opteron $23,000 (16% discount?)
    Fujitsu Primergy TX600 S3 4xDC Xeon 3.0GHz $25,000

    All of the above systems are for processors manufactured on the 90nm process.
    Soon there will be Xeon 65nm procs, Opteron to follow in 2007, hopefully so will Itanium
    Anyways, the design goal for Itanium was to achieve approx 50% better than the Xeon on the same manufacturing process.

    I do not know if Gunter will answer this, but i am dying to know how of the above Itanium result was due to hyper-threading
    given that the 4x1.6GHz single core Itanium scored 161K tpm-C (on Oracle, not SQL Server)
    my expectation was that a 4 DC It at 1.6GHz be in the range 270-288K tpm-C without the HT feature in Montecito
    if so, then the 344K result implies a 20-25% gain from HT

    my understanding is that the Xeon systems gain 10% in the TPC-C benchmark, 20% in SAP,
    but in actual production environments, HT is very likely to cause serious performance anomalies, if not problems
    is the HT feature in Itanium much better and safer ?
  4. JerrySommerville New Member

    We are about to move from a 4 CPU single Core Itanium Integrity server (HP) to a 16 CPU Itanium Dual Core server (HP Integrity also) and I would like to know what performance tests would you recommend , especially if they are SQL Server 2005 based tests. We want to monitor performance closely and establish some benchmarks now so we can chart performance degredation as more users are added to the application. We currently have about 400 users going to well over 800.
  5. joechang New Member

    any idea why your post is not wrapping?

    i have relatively good idea on the performance characteristics of the previous generation 16 socket HP Integrity, sx1000 chipset, and 1.6GHz Madison cores
    yours would have the new improved sx2000 chipset, and new Montecito dual core CPU with hyper-threading

    so basically you are going from 4 physical and logical cores to 16 sockets, 32 physical cores and potentially 32 logical proc (w/HT enabled)

    i do not think you need to run any particular test on your new system,
    besides, my calibration suite in not in kit form, that is, easily used by some one else.

    it is very important the you understand the performance characteristics of your application, as going above 4 cores can have positive as well negative impact,

    collect the information on my site (www.qdpma.com) before a move, do an extensive analysis before moving the production db to the new server,
    run load tests on the new server
    all this to make sure there are no negative suprises for your users and your bosses
  6. ex-oracle New Member

    I have recently attended a presentation by HP (+ microsoft + intel) where they were pushing Itanium 2 based servers for SQL Server 2005. In the UK, the pre-sales guys were mumbling about doing rx6600's for DL385 G2 money.
    i.e. big discounts to get you on message. Of course the "doing rx6600's for DL385 G2 money" can mean almost anything.
    Ex-Oracle
  7. joechang New Member

    if hp wants to sell a 4 socket Itanium 9040 (dual core 1.6GHz 9MB) for $30K, i would consider it, compared with about $20-30K for a 4 socket Xeon or Opteron

    otherwise, I am inclined to think Itanium makes more sense in > 4 socket systems
  8. jezemine New Member

    what's the connection between blue screens and high altitude? is it cosmic rays?

    www.elsasoft.org
  9. joechang New Member

    cosmic ray hits the atmosphere
    create secondary particles

    the neutral particles will penetrate to sea level and cause soft errors,
    which might cause detected or undetected problems
    depending on what memory was altered

    desktops are especially vulnerable because of complete lack of protection
    large SMP systems are also sensitive because of the sheer number of components
    more sensitive at high altitude

    that's why i am curious if any one is at high altitude with a 4-way or bigger box


Share This Page