I am against writing to both Masters in a dual-master setup. There are too many things that can go wrong, and they can be messy to fix -- AUTO_INCREMENTs, other duplicate keys, etc.
Hundreds of "Sleep" connections is virtually no impact on a server, so limiting to 40 is not useful. 10 or more active connections (non-Sleep) can be an issue. In that case I would look at the queries. Usually optimizing the queries is the best answer.
Note also, that every write (INSERT, UPDATE, etc) that is done on one Master must be done on all the other Masters and Slaves. So, you can't really "spread" writes around.
If you have processes that do only reads (SELECTs), then they should go to Slaves and/or the backup Master, not the live, writable, Master. This will help.
Be aware of the "critical write" problem. Example: A user posts a blog comment, then looks at his comments, but it is missing. This can happen if the write went to one machine, but the read hit another, and replication is "behind".
(My comments apply to all versions, and all APIs, not just 5.1 and PHP's mysqli.)
I stay away from mysql_pconnect (and other connection pooling mechanisms). Connection startup/teardown is very fast in MySQL. Pooled connections may have issues with @variables, transaction modes, sql_modes, etc.
I would like to suggest something radical. I got this idea from StarTrek : Deep Space 9 (Call to Arms)
A minefield was set up around a wormhole to prevent the Dominion
Forces and the Jemhadar from passing through. Each explosive mine was
armed with a replicator that allowed a mine to recreate another mine
after it explodes. Hence, the minefield stays up indefinitely.
With that DS9 analogy I bring your an interesting idea.
You set up 2 initial read slaves with MySQL Slave with innodb disabled
[mysqld]
skip-innodb
This is optional. My preference is an all-MyISAM slave to do reads because it is faster for reads than InnoDB for small datasets. Should you choose to go with InnoDB, make sure you relax the ACID compliance with this
[mysqld]
innodb_flush_log_at_trx_commit = 0
In the event of a crash, just destroy the Slave and spin up a new one
We'll call the Slaves S0 and S1
Here is something else: have this in /etc/my.cnf in S0
[mysqld]
innodb_max_dirty_pages_pct = 0;
innodb_fast_shutdown = 0
These will help S0 shutdown fast with completely flushed data.
The following is what you must script in the replicator process
When you need to generate a new slave (we will call it S2), here is what you must do
STEP 01) On S0, run service mysql stop
STEP 02) Install the same version of MySQL that S0 has into S2
STEP 03) On S0, scp /etc/my.cnf S2:/etc/.
STEP 04) On S2, you need to change the server-id
of /etc/my.cnf in S2 to a Unique Value (suggestion : use the 2nd,3rd, and 4th octet [without the dots] of the private IP of S2)
STEP 05) On S2, either remove or comment out the innodb_max_dirty_pages_pct = 0
from /etc/my.cnf
STEP 06) On S0, run rsync -av /var/lib/mysql S2:/var/lib/mysql
(Note: If you have to spin up 5 Slaves, run the 5 rsyncs at this point)
STEP 07) On S2, service mysql start
(MySQL Replication start immediately where S0 had left off)
STEP 08) On S0, run service mysql start
Once you create the replicator script, you can use it spin up MySQL on new Slaves.
Meanwhile, S1 is available for SELECT queries and continues replicating.
S0 is used to recreate a new Slave.
If you are doing this in conjunction with spinning up Amazon EC2 or some other Cloud DB Servers, check with your System Administrators on any Linux commands/API that allows you to spin up a DB Server. Then, you apply the replicator to the newly generated DB Server. Even better, you can incorporate the Db Server creation API in your Replicator Script.
CAVEAT
If you have to spin up dozens or hundreds of Slaves, all you need to do is have 10 Replicator Slave Servers (S0 - S9) and have 10 copies of the replicator script operator on different Replicator Slaves.
Best Answer
The new MySQL Router is one possible solution. Please take a look and let us know what you think!
I'd love to hear what features you most want in future versions.
http://mysqlhighavailability.com/mysql-router-on-labs-the-newest-member-of-the-mysql-family/
http://mysqlhighavailability.com/easy-load-balancing-and-high-availability-using-mysql-router/