Re: SMP/MPP and PQO

Hi Lisa

        SMP - In SMP many CPUs share the same memory. Typical example would be E450, 4 processors, 4GB memory. Only L2 cache in the CPUs need to be co-ordinated when a memory buffer is changed or intended to change.
        MPP - In MPP each CPU has its own memory and its own disks that they control (typically). An example would be nCUBE or IBM RS6000, where each processor runs its own OS and messaging is used heavily to coordinate the access to the resources.
        NUMA - Nun -uniform memory access - Group of CPUs share the same memory and the messaging is used between the group of the CPUs to coordinate the access to the resource. An example would be Sequent NUMA-Q, where there are multiple quads and each quad has 4 cpus and 2GB memory (typically) and IQ-link coordinates the access to the resources between the quads. HP has ccNUMA for this.

        Now, I don't think, E10k domain would qualify for a MPP, since there is no coordination or access restrictions between the domains ( at least that I know of). Each domain can manipulate its own disk / memory /devices without any messaging between the domains, in a non-clustered environment.
       
        PQ architecture splits the FTS or long running operations on non-partitioned tables, in to multiple chunks based upon the rowid ranges. It is not a simple split though. First, work is divided in to degree of parallelism. For example, if the QC process determines that it has to scan 100,000 blocks and parallelism is determined as 4 then 25000 would be the split. But not all 25000 block is handed over to the slave processes. QC uses 9/13 rule. QC hands over 9/13 of the work (i.e. 25000 * 9/13=17307) blocks equally to the slave processes. When the slave processes finishes up the 9/13 work, then they get 3/13 of the work (25000 *3/13=5769) blocks and 1/13 of the work when the second set of 3/13s are exhausted. This is done to evenly distribute the load. For example, there could data skew such that one process could finish up its 9/13 work much faster than other processes and hence those processes can take up the remain!
!
ing work.

        The suitability of PQO parallelism should be defined by IO sub system in conjunction with CPU power. If you have many spindles in  2 cpu server my take on it would be to use parallelism of 2 to4 depending upon the cpu clock speed, disk speed, memory etc..
       

Thanks
Riyaj "Re-yas" Shamsudeen
Certified Oracle DBA
i2 technologies   www.i2.com

"Koivu, Lisa" <lisa.koivu_at_efairfield.com> Sent by: root_at_fatcity.com
06/12/01 09:17 AM
Please respond to ORACLE-L

       
        To:        Multiple recipients of list ORACLE-L <ORACLE-L_at_fatcity.com>
        cc:        
        Subject:        SMP/MPP and PQO

Hello everyone,
I was reading up on the differences between SMP, MPP and how they may affect PQO (Parallel Query). My understanding is that MPP is a host with defined domains (like an e10k with virtual machines on it).  SMP is a standalone host with no domains and multiple processors.  I am not considering clustering here. It seemed to me the only requirement that you really need to run PQO is to have available resources to power it.  For example, a little 2-cpu box that is pinned a majority of the time is only going to suffer if PQO is turned on.  However, if we had a 16-cpu box with abundant resources, turning on PQO would help fts and large index scans in a dw-type environment.  (At least this is what I saw in the past).   Also, I was taught that PQO should not be used when a table/index is not partitioned.  However, upon reading the doco, it states that the slaves split up work by blocks (or was it extents?).  Seems to me this could cause more problems than it's worth (i/o contention?) and partitioning, if done carefully, would be the smarter way to go.  Would the slaves really be smart enough to divy up work intelligently on a non-partitioned object?  My initial thought is NO.   In addition, on metalink they even went so far as to state it is OK to use PQO on a 2-processor NT machine.  Seems to me the statement that 'PQO provides no benefit on a SMP machine' is not warranted, unless Oracle Support was just pacifying the customer who wanted to see PQO work. Maybe my idea of SMP is too simple.  If I am off my rocker can someone please set me straight? Thanks
Lisa Koivu
Oracle Database Administrator
954-935-4117
The information in the electronic mail message is Cendant confidential and may be legally privileged, it is intended solely for the addressee(s) access to this internet electronic mail message by anyone else is unauthorized. If you are not the intended recipient, any disclosure, copying, distribution or any action taken or omitted to be taken in reliance on it is prohibited and may be unlawful. The sender believes that this E-mail and any attachments were free of any virus, worm, Trojan horse, and/or malicious code when sent. This message and its attachments could have been infected during transmission. By reading the message and opening any attachments, the recipient accepts full responsibility for taking protective and remedial action about viruses and other defects. Cendant Corporation or Affiliates are not liable for any loss or damage arising in any way from this message or its attachments. Received on Tue Jun 12 2001 - 12:19:17 CDT