Thursday 19th was the last day of DOAG Confernce 2009. Compared with the second day, the program was much more dense: For my (DBA) profile there have been several hours, where more that one talk would have been appropriate.
This list is, what I collected during the day.
Oracle Grid Infrastructure:
- DNS is mandatory for SCAN addresses, you need 3 IPs
- CTSS (Cluster Time Synchronisation Service) should not be mixed with NTP
- let’s try to install “software only” for Grid Infrastructure ans activate cluster afterwards (useful for cloning nodes)
- a diskgroup with redundancy NORMAL containing voting disks has to have three disks at least
- sshd parameter LoginGraceTime has to be set to “0” (unlimited) for OGI install
- CSSD bypasses ASM when reading the voting disks. So crashing one ASM instance does NOT mean that the cluster does reboot! Mr. Burleson tells something different, but he’s wrong. Just test it. At worst, the node where the crashing ASM resides, might reboot.
- in 11gR2 (it means using Oracle Grid Infrastructure), the Voting Disk is backed up automatically, as OCR is all time since 10g.
- with Voting Disk in ASM, it’s still possible to have one Voting Disk on NFS. The assistant can’t do it at the moment, but you can create a “QUORUM FAILGROUP” with CREATE DISKGROUP, the syntax is in the docs.
- an available patchset gives us SLA-Management (ressource planning) for ServerPools in OGI. The next patch level 184.108.40.206 will include that feature in the whole
- use “yum install oracle-validated” with Oracle Enterprise Linux
- “set scan off;” stops SQL*plus interpreting “&” variables
- let’s follow ISO 144576, it’s said to be the one single practically useful IT norm. 🙂
- I want to review Quest Benchmark Factory
- Oracle Real Application Testing feature: You can apply filters only at capture time, NOT at apply time
- RAT: the overhead for recording will be between 5% and 20%, it’s mostly disk IO
- let’s use PENDING STATISTICS
- and extended column statistics (combine columns for statistics gathering, concept like a multi coulumn index)
- SQL Plan Management is an EE feature
- and it uses SQL profiles internally
- View: DBA_SQL_PLAN_BASELINES
- SQL Plan Management uses SYSAUX
- Where is a useful border to the improved Bind Peeking (Adaptive Cursor Sharing)?
- SQL Plan “Management” sounds like only a patch to avoid root canal therapy, instead of working with the CBO and statistics to weed out performance killers
Hammerora Oracle Benchmarking:
- Sonds interesting, the concept is TCP-C and TCP-H so far
- The hammerora tool is free and open source! Written by Steve Shaw, who’s currently working for Intel.
- Amdahls Rule of Thumb: If a computer has to perform one instruction per second, it has to be provided with 1 byte of RAM and a throughput capability of 1 byte per second
- all TPC-C benchmarks are done with GC_FILES_TO_LOCK to special values
CBO in general
- obviously, COST is CPU+IO+TEMP
- IO cost depends on history, if no system stats are available. For example, if full table scans have been 20 blocks per row in the past, you will see cost of 20 for each row you select now.
- This cost is multiplied with the multiblock read count
Okay, this was quite a lot and I had impressive days in Nuremberg. Hope you can use the stuff I provided in the last days, to start your own search engine tasking and wish., you find out lots of news.