Re: Page Cleaners/Idle Writes

Mike Alder writes: Mike's issue is cc'ed below. Here is the way to tune to get (for most circumstances) best performance. Fiirst, ignore the bit in the manual about tuning page cleaners to disks; tune them == LRUS (8 by default). Set your LRU_MAX_DIRTY to 10-30, and LRU_MIN_DIRTY to 5-10 below that. Set your CKPTINTVL to a value that you are comfortable with, given what it will mean in terms of fast recovery time. If need be, increase your physical log size. The outcome of this is to minimize the amount of time spent doing a checkpoint. Because all transaction activity stops during checkpoints, you want them to be as brief as possible. Let the page cleaners keep the queues clean in the meantime. By having a cleaner per LRU, you stand a better chance of keeping each queue clean. Because cleaners basically pull a buffer then block on a write, they tend to accumulate little cpu. |> I have been experimenting with LRU_MIN AND MAX |> parameters and page cleaning/cleaners to try and enhance performance |> and get the most out of my page cleaners. I have 3 seperate |> spindles with 3 page cleaners and a moderate amount of constant isam |> calls. I had and still have all chunk writes and my checkpoints |> are fired to go off every 10minutes. |> |> I started out with the default values of 60/50 and decreased |> them down to 40/30 before I saw anything happen. What I saw |> happen was more page flushing and an active tbinit -s process for each |> cleaner. I continued to decrease them down with improving |> performance(except write caching 93% to 87%) in all areas |> especially dskreads and rollbacks. I stopped at 20/10 to try |> and understand what was happening with Idle writes and the fact that |> I hadn't been getting any??? |> |> I learn from the admin book that if a page cleaner determines |> that the buffer pool(MLRU ques) should be flushed it marks a page |> for flushing, flushes it, |> rechecks it against the MAX percentage , then keeps working till |> it reaches the MIN percentage of buffers that can be dirty. |> |> What this means is I still don't know what I'm seeing. I see |> "flushes" increase and occur outside of checkpoints but receive |> only chunk writes. |> To me if a page cleaner is woke up and cleans/flushes pages outside of |> checkpoints then I would think I should see some idle writes |> when I use tbstat -F. |> Yet all my writes continue to be chunk writes. |> |> I haven't had the oppurtunity to decrease further to see if I |> would actually see Idle writes and why - But for now if someone |> could clue me in on Idle writes I would be a happy camper. |> |> |> Thanx |> |> Mikea Dave Disclaimer: These opinions are not those of Informix Software, Inc. ************************************************************************** "I look back with some satisfaction on what an idiot I was when I was 25, but when I do that, I'm assuming I'm no longer an idiot." - Andy Rooney