<No Subject Supplied>

>I am getting very high chunk writes and almost 0 LRU writes against onstat >-F >and >high duration against onstat -m. >How do I solve this problem ? >>Through the process of trial and error we have slowly decreased our BUFFERS>>from 50000 to 16000. We also dropped LRU_MAX_WRITE to 10 and LRU_MIN_ >>WRITE to 5. As we decreased the buffer size we say the ratio of LRU/CHUNK >>change from 1/10 to about 3/2. We also decreased our checkpoint time >>from 7-8 seconds to 2 seconds. >>When we ran onstat -F at 15 second intervals between checkpoints we found that >>we never had enough dirty buffers to start the page cleaners working. By >>shrinking the number of buffers AND the percentage of them allowed to be dirty >>we forced the page cleaners to do much more work between checkpoints. I assume you are trying to reduce checkpoint times without affecting anything else. I would not reduce the number of buffers as your read/write cache percentages in onstat -p could drop considerably and so reduce OnLine performance. I would reduce LRU_MIN_DIRTY and LRU_MAX_DIRTY to increase the amount of LRU writes also to reduce checkpoint times set the number of page cleaners = number of physical disks you have and do the following: onstat -d and look at the chunks section specifically columns 2 (chunk id) and the pathnames/offsets. Ensure that ordering of your chunk id's is round robin across the disks and preferably controllers i.e. if you have 2 controllers, 4 disks (2 disks per controller) and 8 chunks then chunk 1 on controller 1, 1st disk on controller chunk 2 on controller 2, 1st disk on controller chunk 3 on controller 1, 2nd disk on controller chunk 4 on controller 2, 2nd disk on controller chunk 5 on controller 1, 1st disk on controller chunk 6 on controller 2, 1st disk on controller chunk 7 on controller 1, 2nd disk on controller chunk 8 on controller 2, 2nd disk on controller During a check point the page cleaners are assigned cleaning on a chunk by chunk basis IN INCREASING ORDER OF CHUNK ID. ***** This is not well known (I only heard about it during the recent tuning course it went on), something for Kerry's FAQ) ***** With this schema during a checkpoint each page cleaner will be cleaning a chunk on a different PHYSICAL disk and preferably different controller (above is not a good example of that but you get the idea). Since the disk I/O is evenly balanced across each controller and disk you get the least time for checkpoints. Of course this assumes all your chunks have an each amount of dirty buffers at checkpoint time Online disk layout is complex to REALLY get the best performance (- sets my head spinning :-) ) Dave.