Buffer Cache sanity check

[donC_1]

Greetings,
Our development environment utilizes ClearCase for code revisioning. We have 12 L2000 servers dedicated to the VOB segment of ClearCase. Each server has 2Gb of memory and 2X440MHz processors, and is running HP-UX 11.0 64bit.
The previous Admin I inherited these server from has dbc_max_pct set to 50 and dbc_min_pct set to 10. Now I know these are defaults values, but CC is known for it's high demand on buffer cache, and the typical usage values support this. Across all 12 servers, of the 1Gb dynamically allocated, usage very seldomly falls below about 850Mb, and is most often right at 1Gb.
Ok, so this brings me to my question/sanity check; I'm thinking that if the normal server activity utilizes close to 1Gb of memory, why not statically allocate that and remove the overhead generated by the system having to monitor/allocate/deallocate...? Does this seem like a logical thought process, or am I missing something? Thanks for all responses.

[doug mielke]

Agreed that setting buffer cache to a static value will cut down onsome of the overhead. What will you set it to? 1 Gig sounds like too much, and postings here mention HP not being able to manage large amounts of cache well.

You could try a sar -b 5 5 to see the buffer cache hit rate to help decide on a value.

[Kent Ostby]

Melanee --

HP has seen really high DBC numbers that can cause problems systems due to the time to flush the buffer cache.

This can be even more of a problem if the system runs an application that needs to poll regularly (such as MC/Serviceguard).

In general I recommend to customers to have a maximum of 400 mb as that seems to be the point where you no longer get much bang for your buck. For customers who want more then I go to my "maximum" recommendation of 750.

I never recommend over 750 MB as the flush time for the DB will slow down certain other functions on your system at that point.

Best regards,

Kent M. Ostby

[James R. Ferguson]

Hi:

A 50% setting for the 'dbc_max_pct' is high and you are probably seeing pulses by 'syncer' when it runs every 30-seconds to flush modified buffer cache. Depending on your application memory requirements, you might also be seeing some memory pressure as evidenced by highter 'vhand' activity. With 2GB of memory, I'd monitor performance with 'glance', then reduce the buffer cache size to no more than 25% and monitor again. Everyone's mileage varies and you have to determine what works best in *your* environment.

Regards!

...JRF...