Buffer Cache usage in a cluster

[Alexander Santamaria Ji]

We have a cluster with MC/Serviceguard and 2
rp7410 servers attached to a SAN with 1 x VA7100, if a package of Data Storage Protector is asigned to a server then the
Buffer cache usage changes from 0.5 Gb to 2GB, then if the package is moved to the other node then buffer cache remains using 2GB.

How can i control this usage?

[Geoff Wild · Accepted solution]

Change the dbc_max_pct kernel parameter.

The default is 50% - for SAP / Oracle servers we set it to 10% or 15%.

RGds...Geoff

[Patrick Wallek]

Once your buffer cache usage increases, it will only decrease if you start running into memory pressure. Otherwise it will keep increasing until it reaches the value governed by dbc_max_pct.

[A. Clay Stephenson]

That is typical of dynamic buffer cache. I suspect different settings of dbc_max_pct and dbc_min_pct on the two boxes. One option is to use a fixed buffer cache on both boxes by setting bufpages to a non-zero value. Bugpafes is set in units of 4k pages (unless some has altered the ralted paramters - very unlikely).

[Bill Hassell]

Just a recommendation: 2Gb is way too large for a typical database system. It could be useful for read-mostly situations but tends to increase system overhead for 50:50 read:write ratios as so many of the write records must be written to the disk at each sync cycle. Keep the buffer cache below 500-700 megs max. You can then give the extra RAM to the DBA who quickly use it in shared memory.

[melvyn burnard]

I agree with Bill's recommendations.
I have seen a system end up doing a SG TOC due to the system being too busy to flush a very large buffer cache, and not being able to service the cmcld daemon.

[Alexander Santamaria Ji]

Bill, Melvyn;

I agree with you that 2GB is too large because our systems have 5GB of memory each one, i am going to limit the buffer cache size at 750MB max.

Thanks for your recommendations (and also for Geoff, Patrick and Clay)

Best Regards

Alexander Santamaria