This post is equally applicable to Rac databases deployed, both on oracle database machines as well as traditional server/storage architectures, running RHEL5 or OEL5 (OEL6 has a new feature called transparent hugepages).

The Best practices for database consolidation on exadata database machine, White Paper outlines a lot of good guidelines. However the best practices seems to be getting occasionally overlooked, in some cases resulting in Node evictions/restarts in the cluster.

High amounts of memory and cpu consumption can result in some real bad things happening (Like for eg: Node evictions). So it is important to configure the operating system and the databases you deploy on the machine optimally, to use the available CPU and memory resources.

Let us first review what those available memory and cpu resources are on an Oracle Exadata Database Machine, on each database node.

[table “1” not found /]

Anytime a workload is deployed on a database server, that exceeds the, operating systems ability to efficiently use the above mentioned, available cpu and memory resources, bad things can happen. (I know i am over simplifying in the prior comment, but you get the idea).

Managing Memory Resources.

Setting up HugePages

First we need to make sure that HugePages on each node, is setup correctly, for the databases you have deployed on the node.
Tanel Poder has explained the reasons for setting up HugePages on Exadata systems.
When the database machines are installed, the HugePages is configured to accomodate the one database that is pre-installed. So if you change the SGA setting for that database or deploy more databases you should now adjust the HugePages setting accordingly. It goes without saying that if you have already configured the hugepages to accomodate all the databases you have, and then you removed some databases from the node, or resized the SGA’s to be smaller, you should again resize your hugepages to free up the excess hugepages.

Once all the database instances are configured and started up you can run the script, hugepages_settings.sh from Mos Note 401749.1, which calculates the number of hugepages you need to setup. I usually recommend adding 10% to the value that hugepages_settings.sh suggests.

With 11.2.0.2 and 11.2.0.3 you should also be setting USE_LARGE_PAGES, preferably to ONLY, following the instructions in Mos Note 1392497.1.

You can then follow the instructions in MOS note 361323.1, under the section “Kernel version 2.6”, to set the value. You need to restart the databases so the SGA gets allocated with the hugepages.

You should also follow instructions from MOS note 1546861.1, and set the value of vm.min_free_kbytes = 524288, in /etc/sysctl.conf, to avoid page allocation failure messages when kernel memory is depleted.

Managing Memory Usage

Ideally for critical implementations, your SGA+PGA+Individual server processes, memory allocations should not exceed 75% of the physical memory on the database node.

The Exadata consolidation white paper above suggests that

For critical Hardware Pools, we recommend an even more conservative approach by not exceeding 75% physical memory per database node.
OLTP applications:
SUM of databases (SGA_TARGET +PGA_AGGREGATE_TARGET) + 4 MB * (Maximum PROCESSES) < Physical Memory per Database Node
DW/BI applications:
SUM of databases (SGA_TARGET + 3 * PGA_AGGREGATE_TARGET) < Physical Memory per Database Node

You can monitor the pga usage a few different ways.

1) SELECT s.inst_id, s.sid, s.serial#, p.spid, s.username, s.program, p.pga_used_mem FROM gv$session s JOIN gv$process p ON p.addr = s.paddr AND p.inst_id = s.inst_id WHERE s.type != ‘BACKGROUND’ and s.program not like ‘%(P%’ and p.pga_used_mem > <APPLICATION_MEMORY_THRESHOLD>
order by s.inst_id, s.sid, s.serial#;

The value you use for APPLICATION_MEMORY_THRESHOLD is dependent on your application and howmuch pga it needs to efficiently run your queries (Performing the sorts in memory using the space for PGA is usually faster than using the TEMP tablespace). However the white paper suggests using 1Gb for OLTP applications and 10Gb for DSS applications.

2) You can monitor the sysmetric “Total PGA Allocated” from dba_hist_sysmetric_summary, to see how much PGA is being allocated.

3) You can use the “PGA Memory Advisory” and “SGA Target Advisory” sections of the awr report, to guage, the impacts of increasing or decreasing the SGA and PGA Allocations.

4) You can monitor the memory usage at the o/s level using the top (Mem: free),free or vmstat (bo (page-out),bi (page-in)) commands.

Managing O/S Processes (And as a consequence CPU usage).

It is important to control the number of processes that are actively executing on each node. Each of these processes consume cpu and memory resources.

The following guidelines are important to follow.
– Use Application connection pooling to appropriately limit the number of sessions connected to the oracle database.

This is important for two reasons.
– Right-Size the resource consumption on each database node.
– Minimize the impact of connection storms.

– Configure the parallel query servers appropriately.
– For all the database’s deployed on a node the total setting for PARALLEL_MAX_SERVERS should be less than or equal to the following

X2-2 or X3-2, <= 240
X2-8 or X3-8, <= 1280

I often see implimentations where parallel_max_servers = 240 (or a high value) in each instance on the machine, which results in not so good consequences. I have occassionaly seen systems where there is sustained 100% cpu usage with load averages of > 300, resulting in node reboots.

– Use Instance Caging and Database Resource Manager to manage cpu allocation on the server.

It would be a good idea to always review the Best practices for database consolidation on exadata database machine, White Paper and follow the guidelines, to have a well performing and  stable database environment.

It would be a good idea to review the performance management screens from Enterprise Manager 12c every day to have a good understanding of the database and server workloads, and/or analyze the oswatcher data to understand the server workload profile.