Re: chekcpoints, BUFFER settings, and optimization

In article <348C6962.4FD1@eds.com>, Greg Moye <greg.moye@eds.com> writes >Jim - > >How many Ultra Sparc CPU's (I'm assuming 6)? >How many Informix instances (I'm assuming 1) on the box? > >Here's what I'd do: >Set Buffers to 100000 (~200 MBytes) >Set LRU_MAX_DIRTY=2 >Set LRU_MIN_DIRTY=1 > Agreed. >Then adjust buffers downward as necessary until you get the desired <2 >second checkpoints. Check Read buffer hit ratio, > 95ish %, be happy. > Agreed. >set PHYSBUFF = 1024 >set LOGBUFF = 1024 > >Monitor after a few days, and reset them so that the log pages per I/O >figure is 3/4 the (new) PHYSBUFF and LOGBUFF size (This is an SAP >guideline also, and it works pretty well...) > Agreed. I would except these to be set to 2 as you are using unbuffered logging. > >I'd also: >set RESIDENT=0 # don't tie solaris' hands unless you have to. Surely you should to stop online getting paged out. The pages are likely to be used again soon. After all the machine is a database server and there is plenty of memory on the machine. >set NUMCPUVPS=6 # you cant do the work if the processes aren't >there > >set RA_PAGES=32 # you got plenty of memory, use it. >set RA_THRESHOLD=30 # never make the CPU wait for memory > Agreed, but wouldn't read-ahread slow down checkpoints as it generates more disk I/O. >Adjust RA_ parms upward until RA-pgsused/(sum of x_RA) < 95ish % (from >onstat -p)> >as for virtual segment size, run "onstat -g seg", count up the bytes and >reset your initial size to that + 10%. Monitor and reapply as >necessary. Additional segments on Sun wont hurt you though like it will >on HP & to a lesser extent, IBM/AIX. > Agreed. But I normally set SHMADD low and monitor to get a better estimate. >BTW, 49 days is a while to be up, do you ever bounce the >machine/database? > You should not need to. Online is a true 24x7 database. Bonuce Online would effectivly flush the buffers, requiring them to be reloaded when Online starts again. >You should also consider upgrading to 7.23.UCn, n>=3 ish, I gather >Informix is really concentrating on making this release stream fast and >rock solid. > Agreed. I think 7.24 is out on some platforms. >Let me know how it turns out...with all the reads and seq scans you did >in 1 minute, this should help a *LOT*. > >Greg > >jneugebauer@ameritrade.com wrote: >> >snip...... -- David Williams