changing dbc_max_pct daily

[Chris Cruz_1]

We are trying to improve the performance of our new rx7620.

We found quite a performance boost when we increased the buffer cache (dbc_max_pct) to 50%, but this is not a realistic setting during the business day when 200+ users are on the system.

Is it reasonable, or have you heard of anyone changing this parameter daily? Raising the "dbc_max_pct" at night to 50% and lowering it in the morning to 30%.

If we did this, would we expect errors, memory garbage... any problems.

[James R. Ferguson · Accepted solution]

Hi Chris:

Well, if it works for you, so be it.

No, there should not be any problems nor errors. When the change is made dynamically, I'd expect the syncer daemon do do a bit of extra work to bring the buffer cache down, but that's about all.

Regards!

...JRF...

[Victor BERRIDGE]

Hi chris,

I used to in older days when resouces and performance were difficult to get, it was not in the morning but in the evening when you could say the system ( 822 - 855)passed from transactional to batch mode...
What I dont understand in your case is you seem to work it out the other way round...
Therefore I had 2 kernels, one for daytime with plenty users on for night where most resources were affected to very few users but they had more of everything...


All the best
Victor

[Geoff Wild]

That depends - how much ram is in your system?

Usually - you want DBC_MAX_PCT to be a percent of ram that equals between 450 and 800 MB.

So, if you have 8GB of ram, 10% would be enough...

If you have 2GB, then 40%

With 11iV2 - you can change this on the fly...

Rgds...Geoff

[A. Clay Stephenson]

I'm a little surprised that you need to do this. HP-UX at 11.11 and up has been very good at ramping backdown the buffer cache as the system loads and then bumping it back up automatically. You really shouldn't need to change a dynamic value "on the fly" because it is, well, dynamic. If 50% works well for you then leave it there. I would reduce dbc_min_pct to something very low (4%?) to give the system better dynamic range.

Your 30% value even seems high to me but since you didn't bother to indicate the amount of physical memory, swapinfo output, and page-out rates, it's difficult to be very helpful.

At first glance, dynamically tuning a dynamic tunable seems a bit misguided.