RE: checkpoint spikes
Mark, I have been making use of LRU Contention script that you provided (Thanks!!), to check LRU contention on one of our servers. LC (spi*100/buf) is mostly more than 5 at busier times. We have increased the number of LRUS and CLEANERS to 127 but it seems there have been little or no effect on LRU contention. What are we missing here? What else can be changed? We are running IDS 7.30.FC7 on HP-UX 11 with 10 cpus given to this instance. There are 300000 Buffers, 40 AIOVPS, 135 Chunks, 28 disks. Please advise. Thanks, Shehla > -----Original Message----- > From: Mark D. Stock [SMTP:mdstock@mydas.freeserve.co.uk] > Sent: Friday, March 03, 2000 6:40 PM > To: Arshad, Shehla > Cc: informix-list@iiug.org > Subject: Re: checkpoint spikes > > "Arshad, Shehla" wrote: > > > > Now that long checkpoints topic is in discussion I would like to get > some > > help too. We are also getting into a similar problem with long > checkpoints > > (over 13 secs sometimes). Long checkpoints in our case only happen once > in a > > while. Our LRU_MAX and LRU_MIN are at 2 and 1. Our checkpoints mostly > > complete in 0 sec as LRU writes are continously cleaning out the > buffers. > > > > Other parameters are: > > > > LRUS 28, Cleaners 28, NUMAIOVPS 24, BUFFERS 100000, CKPTINTVL 180 > > That's an extremely low checkpoint interval. What was wrong with 5 mins? > > > We are running IDS 7.30.FC7 on HP-UX 11 with 6 cpus given to this > instance > > having 150 max concurrent users. There are 76 chunks residing on 29 > disks. > > We do know that the system is a lot busier (more > updates/inserts/deletes) > > when we see a long checkpoint. Would like to know what measures can be > taken > > to eliminate the checkpoint duration spikes. > > You really need to monitor your system at the time of these spikes to > determine what is running. Some sort of batch process possibly? The > measures I usually take is to ring up the users boss and have them shot! > ;-) > > > Would addition of more LRUS and cleaners solve this problem (over 3500 > > buffers/LRU). How can I find out if the bottleneck is a disk? Also, if > more > > AIOVPs are needed? > > Yeouch! Those are certainly long queues. Also I think you could benefit > from more queues based on the number of chunks. Do you have LRU > contention? (See my other post about that.) If you do then simply keep > increasing the number of queues until it levels out. > > Also it would help if you switched on KAIO, if it is supported. > > Cheers, > -- > Mark. > > +----------------------------------------------------------+-----------+ > | Mark D. Stock mailto:mdstock@mydas.freeserve.co.uk |//////// /| > | http://www.informix.com http://www.informixhandbook.com |///// / //| > | http://www.iiug.org +-----------------------------------+//// / ///| > | |What year 2000 bug? year 2000 bug? |/// / ////| > | |year 2000 bug? year 2000 bug? year |// / /////| > | |2000 bug? year 2000 bug? year 1900 |/ ////////| > +----------------------+-----------------------------------+-----------+