Very basic and generic discussion from the performance point of view. Customers still have to do their due diligence in understanding redo logs, and how they work in Oracle database, before finalizing redo log configuration for their deployments.

• size them properly
  • log writer writes to a single redo log file until either it is full or a manual log switch is requested
        Oracle supports multiplexed redo logs for availability, but this behavior of writing to a file until it is full or a log switch happens, still hold
  • if the transactions generate a lot of redo before a database commit, consider large sizes in tens of gigabytes for redo logs
  • if not sized properly, it leads to unnecessary log switches, which in turn increase checkpoint activity resulting in unnecessary slow down of the database operations
        two redo logs each with at least 5G in size might be a good start. observe the log switches, checkpoints and increase (or decrease, though there is no performance benefit) the file size accordingly
  
  
• do not mix redo logs with the rest of the database or anything else
  • in a normal functioning database, most of the time, log writer simply writes redo entries sequentially to redo logs
  • any slow down in writing the redo data to logs hurt the performance of the database
  • best not to share the disks/volumes on which redo logs are hosted, with anything else
        set of disks, volumes exclusive to redo logs, that is
  
  
• ensure that the underlying disks or I/O medium used to store the redo logs are fast, optimally configured and can sustain the amount of I/O bandwidth needed to write the redo entries to the redo logs
      if those requirements are not met, it could lead to 'log file sync' waits, which will slow down the database transactions
  
  
• redo logs on non-volatile flash storage may have performance benefits over the traditional hard disk drives
  • check this blog post out, Redo logs on F40 PCIe Cards, for related discussion (keywords: 4K block size for redo logs, block alignment)

Goal: for a given process, find out how the SGA was allocated in different locality groups on a system running Solaris operating system.

Download the shell script, sga_in_lgrp.sh. The script accepts any Oracle database process id as input, and prints out the memory allocated in each locality group.

Usage: ./sga_in_lgrp.sh <pid>

eg.,

# prstat -p 12820

   PID USERNAME  SIZE   RSS STATE   PRI NICE      TIME  CPU PROCESS/NLWP
 12820 oracle     32G   32G sleep    60  -20   0:00:16 0.0% oracle/2

# ./sga_in_lgrp.sh 12820

Number of Locality Groups (lgrp): 4
------------------------------------

lgroup 1 :   8.56 GB
lgroup 2 :   6.56 GB
lgroup 3 :   6.81 GB
lgroup 4 :  10.07 GB

Total allocated memory:  32.00 GB

For those who wants to have a quick look at the source code, here it is.

# cat sga_in_lgrp.sh

#!/bin/bash

# check the argument count
if [ $# -lt 1 ]
then
        echo "usage: ./sga_in_lgrp.sh <oracle pid>"
        exit 1
fi

# find the number of locality groups
lgrp_count=$(kstat -l lgrp | tail -1 | awk -F':' '{ print $2 }')
echo "\nNumber of Locality Groups (lgrp): $lgrp_count"
echo "------------------------------------\n"

# save the ism output using pmap
pmap -sL $1 | grep ism | sort -k5 > /tmp/tmp_pmap_$1

# calculate the total amount of memory allocated in each lgroup
for i in `seq 1 $lgrp_count`
do
        echo -n "lgroup $i : "
        grep "$i   \[" /tmp/tmp_pmap_$1 | awk '{ print $2 }' | sed 's/K//g' | 
               awk '{ sum+=$1} END {printf ("%6.2f GB\n", sum/(1024*1024))}'
done

echo
echo -n "Total allocated memory: "
awk '{ print $2 }' /tmp/tmp_pmap_$1 | sed 's/K//g' | awk '{ sum+=$1} END 
         {printf ("%6.2f GB\n\n", sum/(1024*1024))}'

rm /tmp/tmp_pmap_$1

Like many things in life, there will always be a better or simpler way to achieve this. If you find one, do not fret over this approach. Please share, if possible.

[Solaris] Cleanup all IPC resources

Run the following wrapper script with root user privileges.

for i in `ipcs -a | awk '{ print $2 }'`
do
	ipcrm -m $i 2> /dev/null
	ipcrm -q $i 2> /dev/null
	ipcrm -s $i 2> /dev/null
done

[Java, WebLogic] Find the process id (pid) of a WebLogic managed server instance

Run the following as the user who owns the process, or with root user privileges.

/usr/java/bin/jps -v | grep <WLS_server_name> | awk '{ print $1 }'

I think this tip is applicable on all supported platforms.

eg.,
Finding the pid of a managed server, bi_server1.

# /usr/java/bin/jps -v | grep bi_server1 | awk '{ print $1 }'
18659

# pargs 18659 | grep weblogic.Name
argv[7]: -Dweblogic.Name=bi_server1

[Oracle Database] Make Oracle ignore hints

Set the following hidden parameter.

_optimizer_ignore_hints=TRUE

(in general, Oracle does not recommend playing with hidden parameters. Check with Oracle support when in doubt).

[Oracle Database] Data Pump Export in a RAC environment fails with ORA-31693, ORA-31617, ORA-19505, ORA-27037 errors

eg.,

ORA-31693: Table data object "<SCHEMA>"."<TABLE>":"P_1147" failed to load/unload \
        and is being skipped due to error:
ORA-31617: unable to open dump file "<FILE>" for write
ORA-19505: failed to identify file "<FILE>"
ORA-27037: unable to obtain file status
SVR4 Error: 2: No such file or directory
Additional information: 3

Workaround:

Add CLUSTER=N to the list of existing expdp options.

[Solaris, ZFS] Check the current ARC size and its breakdown

kstat -m zfs | grep size 		(any user)  - OR -
echo ::arc | mdb -k | grep size 	(root user)

echo ::memstat | mdb -k		(root user)


eg.,

# echo ::arc | mdb -k | grep size
size                      =       259391 MB
buf_size                  =         3218 MB
data_size                 =       249309 MB
other_size                =         6863 MB
l2_hdr_size               =            0 MB

# kstat -m zfs | grep size
        buf_size                        3375105344
        data_size                       261419672320
        l2_hdr_size                     0
        other_size                      7197048560
        size                            271991826224

# echo ::memstat | mdb -k
Page Summary                Pages                MB  %Tot
------------     ----------------  ----------------  ----
Kernel                   14646326            114424    4%
ZFS File Data            31948806            249600   10%
Anon                     24660113            192657    7%
Exec and libs                8912                69    0%
Page cache                 126052               984    0%
Free (cachelist)            24517               191    0%
Free (freelist)         263965754           2062232   79%
Total                   335380480           2620160

[Fusion Middleware] Disable Fusion Middleware Diagnostic Framework (DFW) Dump Sampling

The Diagnostic Framework in FMW 11g environments detect, diagnose and resolve critical errors such as uncaught exceptions, deadlocked threads and out of memory errors. It is enabled by default.

Though DFW is supposed to diagnose and fix some of the issues transparently, due to the inevitable bugs in [all kinds of] software and misconfigurations, sometimes DFW itself may become a major issue. For instance, there is a bug that reported very high system CPU time on a SPARC server where FMW 11g was running. Per the bug description, the system CPU utilization spikes every minute exactly at 00s of a minute, CPU utilization goes down within few seconds - but the pattern persists and the spiky behavior returns within a minute. Another symptom was the sudden drop in available swap space from tens of giga bytes to a few mega bytes when the CPU spike occurs. Upon close examination, it was found out that DFW in FMW is forking tens of jstack processes to collect the thread dumps from an equal number of java processes running in that FMW environment, causing the sudden spike in CPU (each process is busy gathering thread dumps at the same time) and a steep drop in swap space (each jstack process forked a jmap process. both jstack and jmap processes consume some virtual memory just like any other process). All this happened because DFW thought it found a critical issue, and it wasn't noticed or addressed by anyone including the administrators (DFW couldn't fix this particular issue on its own) - so, it kept gathering the diagnostic data continuously. In this example, DFW did the right thing but the diagnostic data collection frequency was too short - only one minute, that diminished the value of DFW and made it a liability. In such dire situations, probably it is best to disable the dump sampling feature of Diagnostic Framework tentatively while the underlying original issue is being fixed in that application environment. It can be enabled again when the critical issue was fixed, and no longer an issue.

Steps to disable Fusion Middleware Diagnostic Framework (DFW) Dump Sampling: (courtesy: Shashidhara Varamballi)

Method (1) Using WLST:

1. run wlst.sh
2. connect to the AdminServer
3. execute command: enableDumpSampling (enable=0, server='<server_name>')

Method (2) Manual editing of config file:

1. Edit $DOMAIN_HOME/config/fmwconfig/servers/<server_name>/dfw_config.xml
2. Change the "enabled" attribute from "true" to "false".
   eg.,
   <dumpSampling enabled="false">
3. Change the "useExternalCommands" attribute from "true" to "false".
   eg.,
   <threadDump useExternalCommands="false"/>
4. Save the changes

SEE ALSO:

Fusion Middleware Diagnostics weblog

[Solaris 11] Virtual-to-physical link (NIC) mapping

Check the output of /sbin/dladm show-phys (any user). By default, only those physical links that are available on the running system are displayed. Option -P shows the physical device and attributes of all physical links.

eg.,

$ /sbin/dladm show-phys
LINK              MEDIA                STATE      SPEED  DUPLEX    DEVICE
net0              Ethernet             up         1000   full      ixgbe0
net5              Infiniband           down       0      unknown   ibp2
net1              Ethernet             up         1000   full      ixgbe1
net6              Infiniband           down       0      unknown   ibp3
net4              Ethernet             up         10     full      usbecm2

$ /sbin/dladm show-phys -P
LINK              DEVICE       MEDIA                FLAGS
net8              ibp1         Infiniband           r----
net0              ixgbe0       Ethernet             -----
net7              ibp0         Infiniband           r----
net3              ixgbe3       Ethernet             r----
net5              ibp2         Infiniband           -----
net1              ixgbe1       Ethernet             -----
net6              ibp3         Infiniband           -----
net4              usbecm2      Ethernet             -----
net2              vsw0         Ethernet             r----

SUMMARY

Oracle Internet Directory (OID) is an LDAP v3 Directory Server that has multi-threaded, multi-process, multi-instance process architecture with Oracle database as the directory store.

BENCHMARK WORKLOAD DESCRIPTION

Five test scenarios were executed in this benchmark - each test scenario performing a different type of LDAP operation. The key metrics are throughput -- the number of operations completed per second, and latency -- the time it took in milliseconds to complete an operation.

TEST SCENARIOS & RESULTS

1. LDAP Search operation : search for and retrieve specific entries from the directory

In this test scenario, each LDAP search operation matches a single unique entry. Each Search operation results in the lookup of an entry in such a way that no client looks up the same entry twice and no two clients lookup the same entry, and all entries are looked-up randomly.

2. LDAP Add operation : add entries, their object classes, attributes and values to the directory

In this test scenario, 16 concurrent LDAP clients added 500,000 entries of object class InetOrgPerson with 21 attributes to the directory.

3. LDAP Compare operation : compare a given attribute value to the attribute value in a directory entry

In this test scenario, userpassword attribute was compared. That is, each LDAP Compare operation matches user password of a user.

4. LDAP Modify operation : add, delete or replace attributes for entries

In this test scenario, 50 concurrent LDAP clients updated a unique entry each time and a total of 50 million entries were updated. Attribute that is being modified was not indexed

5. LDAP Authentication operation : authenticates the credentials of a user

In this test scenario, 1000 concurrent LDAP clients authenticated 50 million users.

BONUS: LDAP Mixed operations Test

In this test scenario, 1000 LDAP clients were used to perform LDAP Search, Bind and Modify operations concurrently.
Operation breakdown (load distribution): Search: 65%. Bind: 30%. Modify: 5%

And finally, the:

HARDWARE CONFIGURATION

ACKNOWLEDGEMENTS

Major credit goes to our colleague, Ramaprakash Sathyanarayan

Hardly six months after announcing Siebel 8.1.1.4 benchmark results on Oracle SPARC T4 servers, we have a brand new set of Siebel 8.1.1.4 benchmark results on Oracle SPARC T5 servers. There are no updates to the Siebel benchmark kit in the last couple years - so, we continued to use the Siebel 8.1.1.4 benchmark workload to measure the performance of Siebel Financial Services Call Center and Order Management business transactions on the recently announced SPARC T5 servers.

Benchmark Details

The latest Siebel 8.1.1.4 benchmark was executed on a mix of SPARC T5-2, SPARC T4-2 and SPARC T4-1 servers. The benchmark test simulated the actions of a large corporation with 40,000 concurrent active users. To date, this is the highest user count we achieved in a Siebel benchmark.

User Load Breakdown & Achieved Throughput

Average Transaction Response Times for both Financial Services Call Center and Order Management transactions were under one second.

Software & Hardware Specification

Additional Notes:

• Siebel Gateway Server was configured to run on one of the application server nodes
• Four Siebel application servers were configured in the Siebel Enterprise to handle 40,000 concurrent users
  • - Each SPARC T5-2 was configured to run two Siebel application server instances
  • - Each of the Siebel application server instances on SPARC T5-2 servers were separated using Solaris virtualization technology, Zones
  • - 40,000 concurrent user sessions were load balanced across all four Siebel application server instances
• Siebel database was hosted on a Sun Storage F5100 Flash Array consisting 80 x 24 GB flash modules (FMODs)
  • - Siebel 8.1.1.4 benchmark workload is not I/O intensive and does not require flash storage for better I/O performance
• Fourteen iPlanet Web Server virtual servers were configured with Siebel Web Server Extension (SWSE) plug-in to handle 40,000 concurrent user load
  • - All fourteen iPlanet Web Server instances forwarded HTTP requests from Siebel clients to all four Siebel application server instances in a round robin fashion
• Oracle Application Test Suite (OATS) was stable and held up amazingly well over the entire duration of the test run.
  • - The test ran for more than five hours including a three hour ramp up state
  • - While we are at it, do not forget to check the Oracle Application Testing Suite (OATS): Few Tips & Tricks page
• The benchmark test results were validated and thoroughly audited by the Siebel benchmark and PSR teams
  • - Nothing new here. All Sun published Siebel benchmarks including the SPARC T4 one were properly audited before releasing those to the outside world

Resource Utilization

Finally, how does this benchmark stack up against other published benchmarks? Short answer is "very well". Head over to the Oracle Siebel Benchmark White Papers webpage to do the comparison yourself.

[Credit to our hard working colleagues in SAE, Siebel PSR, benchmark and Oracle Platform Integration (OPI) teams. Special thanks to Sumti Jairath and Venkat Krishnaswamy for the last minute fire drill]

Copy of this blog post is also available at:
Siebel 8.1.1.4 Benchmark on SPARC T5

To be clear, this post is about a white paper that's been out there for more than two months. Access it through the following url.

  Best Practices for Deploying Oracle Solaris Zones with Oracle Database 11g on SPARC SuperCluster

The focus of the paper is on databases and zones. On SuperCluster, customers have the choice of running their databases in logical domains that are dedicated to running Oracle Database 11g R2. With exclusive access to Exadata Storage Servers, those domains are aptly called "Database" domains. If the requirement mandates, it is possible to create and use all logical domains as "database domains" or "application domains" or a mix of those. Since the focus is on databases, the paper talks only about the database domains and how zones can be created, configured and used within each database domain for fine grained control over multiple databases consolidated in a SuperCluster environment.

When multiple databases are being consolidated (including RAC databases) in database logical domains, zones are one of the options that fulfill requirements such as the fault, operation, network, security and resource isolation, multiple RAC instances in a single logical domain, separate identity and independent manageability for database instances.

The best practices cover the following topics. Some of those are applicable to standalone, non-engineered environments as well.

Solaris Zones

• CPU, memory and disk space allocation
• Zone Root on Sun ZFS Storage Appliance
• Network configuration
• Use of DISM
• Use of ZFS filesystem
• SuperCluster specific zone deployment tool, ssc_exavm
• ssctuner utility

Oracle Database

• Exadata Storage Grid (Disk Group) Configuration
• Disk Group Isolation
  • Shared Storage approach
  • Dedicated Storage Server approach
• Resizing Grid Disks

Oracle RAC Configuration
Securing the Databases, and

Example Database Consolidation Scenarios

• Consolidation example using Half-Rack SuperCluster
• Consolidation example using Full-Rack SuperCluster

Acknowledgements

A large group of experts reviewed the material and provided quality feedback. Hence they deserve credit for their work and time. Listed below are some of those reviewers (sincere apologies if I missed listing any major contributors).

Kesari Mandyam, Binoy Sukumaran, Gowri Suserla, Allan Packer, Jennifer Glore, Hazel Alabado, Tom Daly, Krishnan Shankar, Gurubalan T, Rich long, Prasad Bagal, Lawrence To, Rene Kundersma, Raymond Dutcher, David Brean, Jeremy Ward, Suzi McDougall, Ken Kutzer, Larry Mctintosh, Roger Bitar, Mikel Manitius

Just like the Siebel 8.1.x/SPARC T4 benchmark post, this one too was overdue for at least four months. In any case, I hope the Oracle BI customers already knew about the OBIEE 11g/SPARC T4 benchmark effort. In here I will try to provide few additional / interesting details that aren't covered in the following Oracle PR that was posted on oracle.com on 09/30/2012.

    SPARC T4 Server Delivers Outstanding Performance on Oracle Business Intelligence Enterprise Edition 11g

System Under Test

The entire BI middleware stack including the WebLogic 11g Server, OBI Server, OBI Presentation Server and Java Host was installed and configured on a single SPARC T4-4 server consisting four 8-Core 3.0 GHz SPARC T4 processors (total #cores: 32) and 128 GB physical memory. Oracle Solaris 10 8/11 is the operating system.

BI users were authenticated against Oracle Internet Directory (OID) in this benchmark - hence OID software which was part of Oracle Identity Management 11.1.1.6.0 was also installed and configured on the system under test (SUT). Oracle BI Server's Query Cache was turned on, and as a result, most of the query results were cached in OBIS layer, that resulted in minimal database activity making it ideal to have the Oracle 11g R2 database server with the OBIEE database running on the same box as well.

Oracle BI database was hosted on a Sun ZFS Storage 7120 Appliance. The BI Web Catalog was under a ZFS/zpool on a couple of SSDs.

Test Scenario

In this benchmark, 25000 concurrent users assumed five different business user roles -- Marketing Executive, Sales Representative, Sales Manager, Sales Vice-president, and Service Manager. The load was distributed equally among those five business user roles. Each of those different BI users accessed five different pre-built dashboards with each dashboard having an average of five reports - a mix of charts, tables and pivot tables - and returning 50-500 rows of aggregated data. The benchmark test scenario included drilling down into multiple levels from a table or chart within a dashboard. There is a 60 second think time between requests, per user.

BI Setup & Test Results

OBIEE 11g 11.1.1.6.0 was deployed on SUT in a vertical scale-out fashion. Two Oracle BI Presentation Server processes, one Oracle BI Server process, one Java Host process and two instances of WebLogic Managed Servers handled 25,000 concurrent user sessions smoothly. This configuration resulted in a sub-second overall average transaction response time (average of averages over a duration of 120 minutes or 2 hours). On average, 450 business transactions were executed per second, which triggered 750 SQL executions per second.

It took only 52% of CPU on average (~5% system CPU and rest in user land) to do all this work to achieve the throughput outlined above. Since 25,000 unique test/BI users hammered different dashboards consistently, not so surprisingly bulk of the CPU was spent in Oracle BI Presentation Server layer, which took a whopping 29%. BI Server consumed about 10-11% and the rest was shared by Java Host, OID, WebLogic Managed Server instances and the Oracle database.

So, what is the key take away from this whole exercise?

SPARC T4 rocks Oracle BI world. OBIEE 11g/SPARC T4 is an ideal combination that may work well for majority of OBIEE deployments on Solaris platform. Or in marketing jargon - The excellent vertical and horizontal scalability of the SPARC T4 server gives customer the option to scale up as well as scale out growth, to support large BI EE installations, with minimal hardware investment.

Evaluate and decide for yourself.