Oracle RAC Administration - Part 13: Cache Coherency
January 18, 2007
In our last article, we took a deeper and closer look at the OIFCFG command line utility. As a DBA, the performance of our RAC is crucial. Just like any other application, we need to make sure that our Oracle Database, whether a single node or a multi-node RAC, runs efficiently.
Although we are still into the administration part of the series, we will throw some special attention to the performance of our RAC. After all, when the installation is done, all the manuals and workbooks have been delivered and the consultants have gone home, its time to watch how the workload of the database is going to affect the performance of the RAC. It is almost always a forgotten area. Not to a professional DBA obviously, but surely to the ones who think, "Oracle 11g will be self-managing so I really dont need a professional DBA."
Some Performance metrics specific to the RAC environment
What exactly do we mean by Cache Coherency? Our Oracle RAC environment needs some added sets of metrics rather than a regular Oracle RAC installation, which I sometimes refer to as a Single-Node RAC. I call it a Single-Node RAC because someday that Oracle application will also grow and need a ticket to RACdom. A typical production DBA, responsible for uptime and upkeep of his RAC Database needs more that just some statspack runs; he will need to measure the health of his HSI (High Speed Interconnects) Network interfaces, he will have to monitor and diagnose the traffic volume across the nodes and response times. A typical high intensive OLTP environment can keep you pretty busy. To measure the traffic we will concentrate on two categories:
So what are they?
Global Cache Service
According to the manual:
It actually is more or less like your buffer cache, but here it acts globally across the nodes. This process is an integral part to the cache-fusion concepts. So what does a buffer have, data blocks obviously. Simply said, the coherency in the Global Buffer Cache is maintained by making sure that whenever an attempt to modify the database block is made, a global lock is acquired. . Now this asking instance will have both the past copy of the block (for redo purposes) as well as the current version of the block containing both committed and uncommitted transactions. Should another node come asking for that block, then it is the GCSs responsibility to do a Block Version Lookup at the node, which is currently holding the global lock to the block. The LMSn processes are crucial for a successful operation of GCS and do the block version lookup, block mode etc.
Global Enqueue Service
According to the manual:
The blocks in your RAC environment do most of the work themselves, but there is a crucial area when GES or the Global Enqueue Services come in. A seamless coordination across the nodes is crucial for RAC's operation. The GES is primarily responsible for maintaining coherency in the dictionary and library caches. The dictionary cache consists of the data dictionary master information for each node in its SGA (System Global Area) primarily for quicker lookup and access. Any DDL committed from a requesting node needs to be synced and written across all data dictionaries in all nodes of the RAC environment. The GES makes sure that the changes remain consistent across the nodes and that there are no discrepancies. Moreover, with the same directive, the locks must be created and maintained across the nodes and GES must ensure that there are no deadlocks across requesting nodes over access to the same objects. LMON, LCK and LMD processes work in tandem to make the GES operate in a smooth and seamless fashion.
Obviously, the meat part of the whole equation is, Where are my RAC views? RAC environment has additional views known as Global Views. A typical view for a Single Node installation is V$ but for RAC you have GV$ views. In addition, all these views have additional columns like INST_ID to identify nodes across the RAC environment. So a typical 4 node RAC (like that of our VMware ESX 3.0 nodes) will give you four nodes in our 4-node RAC with their own data when querying the GV$ view. Obviously, you can query individual nodes from any node. To get started try doing this:
SQL> select * from gv$sysstat where name like '%gcs %';
This will give you a result set with specific attention to GCS messages sent across the nodes. If this value is inconsistent across nodes or if huge differences are apparent then it might be time to investigate.
Finding which NIC was used for block transfer
Try to do the following to find out which NIC (private obviously) was used for Cache Fusion:
SQL> oradebug setmypid Statement processed. SQL> oradebug ipc Information written to trace file. SQL> oradebug tracefile_name /oracle/app/oracle/product/10.2.0.1/admin/ /10gr2_ora_12443.trc SQL>
The trace file will contain the details of the IPC information along with the interconnect details:
SKGXPCTX: 0xad95d70 ctx admono 0xbcfc2b9 admport: SSKGXPT 0xad95e58 flags info for network 0 socket no 8 IP 10.0.0.25 UDP 38206 sflags SSKGXPT_UP info for network 1 socket no 0 IP 0.0.0.0 UDP 0 sflags SSKGXPT_DOWN active 0 actcnt 1 context timestamp 0
It is clear that it is our node 1 using private port 10.0.0.25 via UDP.
What networks do various Oss use?
Several vendors, like Veritas (Symantec now), have their own protocols to enable a fast and efficient cache fusion operation.
In our next article, we will carefully start moving towards Workload Management. (AWR, FAN, etc.) I will try to keep the articles more fun and examples oriented. Although I must stress on the fact that it is not all examples and we must keep coming back to fundamentals.