On Wed, 14 Jun 2000 01:34:35 -0700, qmangi
> i need help on one of my systems that seems to be slowing
>down. it is a k-class with hpux 10.20.
> the first noticable figures that goes up is the load from
>/usr/bin/top. however, the cpu idle is at a very high 60% so it
>seems that there is no cpu bottleneck.
> i use glance for further investigation and found out that
>eventhough there are around 612 page faults rate, there is no
>page in or page out. reactivated/deactive proc. how do page
>faults get satisfied?
Sounds to me that the page faults are being satisfied from memory
buffer cache. If I recall correctly, 10% of the total memory of the
machine is allocated for cache by default.
> i do get disk bottleneck, but i don't know what else to do, it
>seems that that was how the program was written i.e. read/write
I have built a variety of machines that live in this type of
environment. It seems the key to making them work well is having
plenty of memory and the best you can afford in disk technology.
Although HP does not recommend it, I allocate most all of the unused
memory on a machine to buffer cache (why waste it?). The key here is
to know what your peak usage is otherwise. HP suggests that there are
some issues with a large buffer cache; I suggest the additional losses
involved with the os dealing with a large buffer are still far less
than processing waiting for disk.
Another item I make sure that I have plenty of is ninode. Watch the
usage during peak periods and expand accordingly. Don't necessarily
go by the high water mark in Glance though where performing a backup
will cause it to generally hit 100% usage during that evolution.
Another thing to watch for is carefully allocating/matching disk
block/fragment size to the needs of the environment and the database
engine in particular. Knowing the ins and outs of how the db engine
works is the key here. I have found the default 1K fragment size of
VxFS file systems is generally too small. Sequential files are best
handled with the fragment size the same as the block size. A good
example is a spool output file. There is certainly no benefit to
having multiple memory page faults when one will do the job.
> the network guys says there is a network congestion, but i can
>see my network queue at 0 and collision rate a little under 1%.
> all these data don't seem to match, can anyone out there please
>help me on how i can really understand this.
>thanks in advance
>* Sent from RemarQ http://www.remarq.com The Internet's Discussion Network *
>The fastest and easiest way to search and participate in Usenet - Free!