Paging Space Placement Question - Aix

This is a discussion on Paging Space Placement Question - Aix ; Our AIX admins and DBAs have had some differing opinions on the layout of our systems and the use of paging space, so I thought I would post something here to see the opinions of the experts here. Here is ...

+ Reply to Thread
Results 1 to 3 of 3

Thread: Paging Space Placement Question

  1. Paging Space Placement Question

    Our AIX admins and DBAs have had some differing opinions on the layout
    of our systems and the use of paging space, so I thought I would post
    something here to see the opinions of the experts here.

    Here is the system details:
    ===================

    - P5 system with 4 processors
    - AIX 5.3
    - 8 GB of memory
    - eight internal SCSI disks set up as two VGs (rootvg X 2, datavg X 6)
    - LV mirroring to different disks will be done

    The system will host four Informix IDS 11 instances:
    - Inventory and Production (OLTP)
    - Accounting (OLTP)
    - Information Warehouse (mainly read-only lookups and reporting)
    - Testing Environment (rarely used OLTP for training new employees)

    It has always been suggested to us to isolate the O/S on rootvg and
    install our RDBMS and other packages on a separate VG, so we mirror
    rootvg between two disks, and use the remaining disks for our data
    (datavg).

    We are also advised to make paging space equal to the amount of memory
    we have, if we have more than 4GB of memory. (This amount is contested
    even within IBM!)

    Here is the scenario:
    ===============
    We want to organize the installation of the LVs in such a way as to
    maximize performance for the daily operations of the production and
    accounting instances. Our AIX admin indicates that we should spread
    our LVs across the disks to make as many spindles available and avoid
    having a hard disk become a bottleneck due to excessive activity. On
    the other hand, others are saying that we should avoid putting things
    like a paging space on those same disks, so that if the system starts
    paging, it will not interfere with the RDBMS activities.

    So the thought is to do the following:

    hdisk0 <----> hdisk4 = rootvg mirrored (hdisk4 on other controller)
    hdisk1,2,3 <----> hdisk5,6,7 = datavg mirrored (5,6,7 on other
    controller)

    Spread the two main OLTP instances and the IW instance over as many
    disks as possible, put the test database in an out of the way area
    (perhaps using up the extra space on the rootvg left from putting a
    small O/S footprint on a 72GB drive), and spread the paging space over
    all disks evenly.


    The admins then began a debate about device speeds...Traditionally,
    memory was the fastest, disk and tape access were the slowest part of
    your system. So these two questions were raised:

    1) Is it better to have one 8GB paging space on rootvg, two X 4GB, or
    even it out on all disks and have four X 2GB as we were planning?
    (Assuming that the paging spaces are mirrored)

    2) Should we isolate the two main OLTPs from each other, putting the
    production one on two disks (hdisk1, hdisk2) and the accounting on a
    separate disk (hdisk3) to prevent them from fighting for contention of
    the disks?

    Or are we simply getting too picky in our concerns about data
    placement on the disks?

    Would like to hear from any AIX memory or install gurus, or anyone who
    has had experience with Informix databases on their AIX systems.
    Thanks in advance for your thoughts on this.

    Steve

  2. Re: Paging Space Placement Question

    steven_nospam at Yahoo! Canada schrieb:
    > Our AIX admins and DBAs have had some differing opinions on the layout
    > of our systems and the use of paging space, so I thought I would post
    > something here to see the opinions of the experts here.
    >
    > Here is the system details:
    > ===================
    >
    > - P5 system with 4 processors
    > - AIX 5.3
    > - 8 GB of memory
    > - eight internal SCSI disks set up as two VGs (rootvg X 2, datavg X 6)
    > - LV mirroring to different disks will be done
    >
    > The system will host four Informix IDS 11 instances:
    > - Inventory and Production (OLTP)
    > - Accounting (OLTP)
    > - Information Warehouse (mainly read-only lookups and reporting)
    > - Testing Environment (rarely used OLTP for training new employees)
    >
    > It has always been suggested to us to isolate the O/S on rootvg and
    > install our RDBMS and other packages on a separate VG, so we mirror
    > rootvg between two disks, and use the remaining disks for our data
    > (datavg).
    >
    > We are also advised to make paging space equal to the amount of memory
    > we have, if we have more than 4GB of memory. (This amount is contested
    > even within IBM!)
    >
    > Here is the scenario:
    > ===============
    > We want to organize the installation of the LVs in such a way as to
    > maximize performance for the daily operations of the production and
    > accounting instances. Our AIX admin indicates that we should spread
    > our LVs across the disks to make as many spindles available and avoid
    > having a hard disk become a bottleneck due to excessive activity. On
    > the other hand, others are saying that we should avoid putting things
    > like a paging space on those same disks, so that if the system starts
    > paging, it will not interfere with the RDBMS activities.
    >
    > So the thought is to do the following:
    >
    > hdisk0 <----> hdisk4 = rootvg mirrored (hdisk4 on other controller)
    > hdisk1,2,3 <----> hdisk5,6,7 = datavg mirrored (5,6,7 on other
    > controller)
    >
    > Spread the two main OLTP instances and the IW instance over as many
    > disks as possible, put the test database in an out of the way area
    > (perhaps using up the extra space on the rootvg left from putting a
    > small O/S footprint on a 72GB drive), and spread the paging space over
    > all disks evenly.
    >
    >
    > The admins then began a debate about device speeds...Traditionally,
    > memory was the fastest, disk and tape access were the slowest part of
    > your system. So these two questions were raised:
    >
    > 1) Is it better to have one 8GB paging space on rootvg, two X 4GB, or
    > even it out on all disks and have four X 2GB as we were planning?
    > (Assuming that the paging spaces are mirrored)
    >
    > 2) Should we isolate the two main OLTPs from each other, putting the
    > production one on two disks (hdisk1, hdisk2) and the accounting on a
    > separate disk (hdisk3) to prevent them from fighting for contention of
    > the disks?
    >
    > Or are we simply getting too picky in our concerns about data
    > placement on the disks?


    I would create 2 G paging space as hd6 in rootvg + 6 paging00 through
    paging02 on the datavg with 2G each placing it at the inner edge. This
    will also prevent the lv's of the database using the slowest parts of
    the disks ;-)

    Don't really concern the I/O into the paging space. If your system
    starts paging relevant amounts then consider buying RAM or even more RAM
    is the very best option.
    If your paging space usage is below 0.5Gbyte this paging will be very
    static and therefore not produce significant I/O on the disks. It will
    simply make a larger address space available.

    The (very) old rule of thumb having 1.5 to 2 times amount of RAM as
    paging space is obsolete.

    --
    Uli

  3. Re: Paging Space Placement Question

    paging space ... hmm ... ok, in your scenario you do not mention how
    much memory your database requires ? in Oracle it would be the SGA and
    PGA .. in db2 it would be at least the sum of the bufferpools etc ..
    this is very important, as if you can fit all of your memory
    requirements into realmem, then you need very little paging space at
    all .. (( lru_file_repage=0 and defps=1 will take care of that for
    you )).

    Then you only have to worry about data placement for your tablespace
    containers. Generally (and I have more experience with db2) the DBAs
    ask for containers on each disk that only span 1 disk, then let db2 do
    the load balancing across those containers.

    A history lesson .. with aix 4.3 (IIRC), deffered paging space
    allocation became the default setting (defps=1), so lru only starts
    looking for candidates to page out based on the lru (least recently
    used) paging algorythm when you hit minfree on the freelist, and
    therefore you do not need 2 x realmem for paging space anymore .. this
    has been the case since 1999'ish

    HTH
    Mark Taylor


+ Reply to Thread