[SUMMARY] Why is IO so poor on my Alpha

From: Susan Rodriguez <SUSROD_at_HBSI.COM>
Date: Tue, 26 Oct 1999 12:10:04 -0700

I had a lot of good information about looking at the write speeds of my
drives, etc. I have included some of that information below. The best
single piece of information came from a Compaq whitepaper on tuning Oracle
for Digital Unix. That document is attached in Word format.

The biggest improvements came from

1 - defragmenting oracle with an oracle tool
2 - setting the vm subsystem as follows:
vm:
        vm-maxvas=8589934592
        ubc-minpercent=1
        ubc-maxpercent=10
        vm-page-free-target=2048
        vm-page-free-swap=1664
        vm-page-free-optimal=1536
        vm-page-free-min=1024
        vm-page-free-reserved=768
        vm-agressive-swap=1
        vm-mapentries=400
        gh-chunks=300

**Some of these were recommended from the list, some from a Compaq
whitepaper (attached).
3 - Setting swap mode to overcommit

I agree with Alexander Sakalleris of Greece who told me to let the EMC array
do the disk grouping rather that LSM. That will be my next step. We will
also be evauluating AdvFS for a possible performance gain.

Thanks again.

susrod_at_hbsi.com

You should not use LSM to strippe your disks,
instead you should instruct EMC to create
stripped hyper metavolumes with at least
7 hyper members on each metavolume.

I have it done so, and my Symmetrix realy kicks.
Additionaly, you should investigate whether you
have enough cache memory in your Symm.

Finaly, you should setup Oracle tablespace partitioning
in order to spread all dbf files across all your EMC disks
evenly.

(BTW, you should consider using AdvFS. It's much faster,
more tunable and highly reliable provided you have created
your advfs disks in a proper way.)

Cheers
________________________________________

                 Alexander Sakellaris
             Sr. Systems Administrator
     Operations Department / IT Division
       Network/Systems & Support Sector
        COSMOTE Mobile Telecomms S.A
                    Athens / Greece

email1: asakelaris_at_cosmote.gr


From: alan_at_nabeth.cxo.dec.com
Sent: Thursday, October 07, 1999 9:02 AM
To: Susan Rodriguez
Subject: Re: [Q] Why is I/O so poor on my Alpha?


        There are two routes to take. One is to find out what sort
        of I/O Oracle is using to create the table spaces; small
        random writes, large sequential ones or something in between?

        The other is to just test the components and see if there
        is an obvious problem. The simplest way to generate a
        simple sequential I/O load is with the dd(1) command. To
        generate a write load:

                dd if=/dev/zero of=file bs=1024k count=whatever

        The special file, /dev/zero is an infinite data source.
        For the output file (of=) you can specify a path name
        to the file system on the device or a raw device name,
        depending on the availability of the device. This example
        starts with a large I/O size, but you can make other tests
        to see what affect different sizes have. If you want to
        do read tests as well, using a size larger than physical
        memory is good since the cache may hold the whole file.

        For read tests simply use /dev/null for the output file
        and the file or device name for the input. Read tests
        can be done to raw devices even when the file system is
        mounted.

        Left to itself, RAID-5 has poor small write performance
        because it has to do at least three extra I/Os for the
        intended I/O. Some subsystems may do extra I/O to help
        keep the array consistent. Write-back caches can help
        this a lot, but only until the cache fills.

        I'd start with a simple sequential write test to the file
        system and see how long it takes (the time built-in is
        useful here). Adjust the I/O size to see what affect it
        has. UFS can do write gathering to turn small application
        writes into larger ones, but only when the data is sufficently
        sequential.

        If this write performance through the file system is relatively
        poor try a raw device and see if it stays the same. If it does,
        then the problem is either in the I/O adapter being used or
        the EMC subsystem. If the file system performance is good,
        then it must be Oracle.


-----Original Message-----
From: Blake Roberts
Sent: Friday, October 15, 1999 9:31 AM
To: Carol Duncan; Karson Cheung; Susan Rodriguez; John Losey
Subject: FW: Modifying sysconfigtab and Oracle tuning document..


Attached is a document that Compaq sent me regarding tuning the server for
Oracle. I'm going to go through this to see what fits with Oracle's
recommendations and what doesn't.

--B


-----Original Message-----
From: Lanier, John [mailto:John.Lanier_at_COMPAQ.com]
Sent: Thursday, October 14, 1999 5:11 PM
To: 'blarob_at_hbsi.com'
Cc: Lanier, John
Subject: Modifying sysconfigtab and Oracle tuning document..



Hi Blake,

Here are those documents that I mentioned in our discussion. Please call
1-800-354-9000 and ask to speak to myself or the next available specialist
should you require further assistance.

--John Lanier, Compaq Services

 <<oracletuning.doc>>

Proper way to modify the sysconfigtab:
============================

[UNIX] What is the Correct Way to Modify the Sysconfigtab File
COMPONENT : System tuning
OP/SYS : Compaq's Digital UNIX V4.0 +
QUESTION :
When making tuning parameter changes to sysconfigtab, what is the
correct way to modify the sysconfigtab file?
ANSWER :
The official way to modify the /etc/sysconfigtab file is to use the
/sbin/sysconfigdb command.

You may update attribute values or add new attributes by first putting
them into a stanza formatted text file and then use the
/sbin/sysconfigdb to merge your changes into the /etc/sysconfigtab
file.

To do this:

     #/sbin/sysconfig -q sub_system > stanza-file
(or create your own stanza formatted file)
     #vi stanza-file
(edit the stanza-file with your changes)
     #/sbin/sysconfigdb -m -f stanza-file sub-system
RESOURCES :
sysconfigdb man page
stanza man page




Received on Tue Oct 26 1999 - 19:14:02 NZDT

This archive was generated by hypermail 2.4.0 : Wed Nov 08 2023 - 11:53:40 NZDT