Hello Gary et al.
 
On Thu, 16 May 1996 15:00:29 GMT, Gary Clarke at Compaq wrote:
|Machine: 922-200, OS: MPE/XL B.40.00
|
|Over the past few weeks we have begun to experience severe 'hangs' on
|one of our production systems. Investigation has shown that dbputs to
|a detail set may be the culprit. Normally, updates are performed
|immediately but now the delay can be from 1 minute to 10 minutes.
|During this time glancexl shows the process to be using a lot of disk
|i/o (but there is no unusual physical disk activity). Other processess
|attempting to write to this dataset are IMPeded (there is item level
|locking on all processes).
 
The problem has probably been building for a while and is also probably
getting steadily worse.  Right?
 
|Glance also show potential memory bottleneck (more memory is on order)
|but this does not appear to be a factor in this problem.
 
The memory symptom is due to the high number of I/Os happening.  It's a
second order effect which will be alleviated by solving the actual problem.
Adding more memory will have a minimal beneficial impact on the I/O
problem.
 
|The detail set contains 2.9 million records (cap: 3.6 mill). There are
|two search items, the first is to an automatic containing 5.0 million
|records (cap: 5.6 mill), the second is to a manual containing 8,519
|records (cap: 11411).
 
I would wager that the problem is a numeric key type in one or both of the
master sets, probably the auto-master.  A second guess is that you have a
path with sorted chains in the detail set which included one or more long
chains which you are addind entries which do not fall at the end of the
sort order.
 
|Capacity changes have been performed fairly recently, but running
|'Sample Master Capacity' in DBGeneral on and around our current
|capacity for a 10% sampling did not identify anything significant.
|
|No configuration changes have been made to this machine recently. The
|database is shadowed via Netbase, and also written to from another
|machine via NB equations.
 
I presume that the other machine has exactly the same data in the
corresponding database.  Does it have the same set capacities?  Does it have
the same performance issues?
 
|Effort is underway to archive some more of this data but will not be
|happening immediately. Has anybody experienced had any similar
|experiences, or have any insight to offer?
|
|Also, are there any tools that will simulate the actions required when
|an image intrinsic is executed? I would be interested to see how many
|(bitmaps of) records in the masters are read through to find an empty
|slot when the primary calculated address is already filled.
 
There used to be a program which would run on classic (MPE/V) CISC systems
(but not on MPE/XL RISC systems) which would print a histogram of the number
of records used on a block by block basis in a dataset.  I don't remember
the name for sure, but I think it may have been DBSCAN.  I never had a
comparable program for MPE/iX (formerly MPE/XL) and never got around to
writing one.
 
Let me (or us) know the answers to these questions and we can make some
suggestions or at least ask better questions.  :)
 
Good luck!
--
Jeff Woods
[log in to unmask] at Unison Software
[log in to unmask]   at home
[PGP key available]
Native American proverb:
We did not inherit the Earth from our parents, nor they from theirs;
We are borrowing the Earth from our children, just as our parents did from us.