Hi all!
A few time ago I passed by some performance issues on AIX working with instances with different configuration (proc/mem). The root cause was basically the inefficient configuration of networking for interconnect (UDP).
As you know, the UDP is a non-response (for that reason with less metadata and faster) protocol. By the default, every server have a pool to send udp (and tcp) messages and another to recieve.
In my situation, once there was an ‘inferior’ instance, the pools were automatically set smaller in this one, and it was causing a high interconnection block sending statistics from other instances. In deed, it was lots of resends caused by overflows in this smaller instance…
There is one one to evaluete how much loss are you having for UDP in your AIX server:
netstat -s | grep 'socket buffer overflows'
If you are having considerable number of overflows, it’s recommended to reavaluate the sized of your udp_recvspace. And, of course, maintain the calculation of pools.
Oracle recommends, at least:
tcp_recvspace = 65536 tcp_sendspace = 65536 udp_sendspace = ((DB_BLOCK_SIZE * DB_FILE_MULTIBLOCK_READ_COUNT) + 4 KB) but no lower than 65536 udp_recvspace = 655360 (Minimum recommended value is 10x udp_sendspace, parameter value must be less than sb_max) rfc1323 = 1 sb_max = 4194304 ipqmaxlen = 512
This and others details about configuring RAC on AIX ban be found in note: RAC and Oracle Clusterware Best Practices and Starter Kit (AIX) (Doc ID 811293.1)
I’d recommend you take a look too.
Have a nice day!
Matheus.
