[Info-vax] DIRCACHE hit rate.

[DJ]

On Jan 20, 9:08 am, Jan-Erik Söderholm <jan-erik.soderh... at telia.com>
wrote:
> Hi.
> I have an 7.3-2 Alpha system where I think that
> the hit rate on the DIR-cache is way to low.
>
> DS20 with 9GB disks in BA356 shelfs.
>
> Here is what it typicaly looks like :
>
>                         CUR     AVE   MIN      MAX
>   Dir Data (Hit %)     3.00    3.20  0.00    30.00
>       (Attempt Rate) 312.00  548.90  0.00  2460.00
>
> The ACP_DIRCACHE is at the moment at 4000 (blocks)
> but I have raised it to 10000 without any measurable
> improvment.
>
> Yes, there are a few DIR's that are "large", but not
> *that* large, maybe 5-10.000 files and a few DIR files
> of 1-2000 blocks. Another thing is that all files
> are timestamped in the filename, so there are
> normaly only ;1 files.
>
> My batch jobs create a number of temp files during
> processing and I've got a feeling that this is
> slowing them down a bit.
>
> Now, what I'd like to ask, is if anyone knows if
> this part of VMS has had any major improvments in
> the 8.x versions ? We are currently thinking of
> upgrading anyway, and it would be nice to know
> if these hit rates would improve simply by upgrading.
>
> Best Regards,
> Jan-Erik.

Dear Jan-Erik Söderholm:

I have work extensively with disk cache tuning over the past 25 years.
The last system that I was on had VMS 8.3 on GS80 that turned out to
be a email system. From this I have found out you can’t look at just
one of the ACP_*cache parameters, you have to look at four of them
because they work together. These four cache parameters affect your
backup speeds, shadow set disk merges, access to data, mail systems,
and etc. BTW, a very good book on the subject is VMS File System
Internals by Kirby McCoy, pages 171-172.

There are three occurrences when disk caching is important and they
are: normal load on the disks, heavy loads on the disk, and backups.
The backup can take the most resources when you are tuning these
parameters. Although, on an email system the disk caches had to very
large to accommodate the disk writes and reads. So, if you don’t have
these parameters tuned where the hit ratios are 90% or better on
average during normal load your system will suffer significant
performance hit, because as the disk activity increases so does the
flushing out of the caches which causes the hit ratios to dropped.

The four ACP_*cache are ACP_DINDXCACHE, ACP_DIRCACHE, ACP_HDRCACHE,
and ACP_MAPCACHE. How are these four sysgen parameters related?

ACP_DINXCACHE – it is a pool of memory that stores part of the
indexf.sys file.
ACP_HDRCACHE – it is a pool of memory that stores file headers.
ACP_DIRCACHE – it is a pool of memory that stores the directory block.
ACP_MAPCACHE – it is a pool of memory that stores the disk bitmap.

Your hit rates should be in the following order of highest to lowest
hit rates: index, header, directory, and bit map, although I have had
header higher than index. If you increase one of these parameters, it
might not make a difference. Therefore, I have in the past keep index
and header file cache the same size followed by directory cache and
bit map cache. What should these values be? In my opinion, high enough
to get your cache hit ratios to be 95% or greater on average. I notice
that you set the header cache to 10,000 blocks and saw no difference.
I am not surprise. Header Cache directly affects file opens and
closes. However, if you have file searches and then file opens, this
would affect the index file and header file cache. I believe Image
Backups will start with Index file, file headers, directories, and bit
map in that order. The bit map cache becomes important with shadow
sets.

In the past, I have set the following values for these parameters:

ACP_DINXCACHE 	– 16384
ACP_HDRCACHE 	–  16384
ACP_DIRCACHE 	–  8192
ACP_MAPCACHE 	–  1024

Increase these parameters; you then must increase the paged pool size.
If you don’t, the cache sizes will drop down to 2 blocks per disk and
then the system will crash. I saw it happen on a VAX/VMS 6.2 system.
How much should we increase the page pool?

1.	Add up the four ACP parameters value – 16384+16384+8192+1024 =
41,984 blocks or 512 bytes of memory.
2.	Multiply 41,984 by 512 = 21,495,808 bytes
3.	Take the above value in bytes and multiply it by 1.5 times =
32,243,712 bytes
4.	This is the value that the page pool should be set to.

Don’t look at the system right away because the cache hit ratios will
be low. Give it a couple hours before you start monitoring the cache
hit ratios. Have the cache ratios improved?

I hope this help you out! It has in the past, for instance you might
find a “DIR” on a several directories will take a few seconds rather
than minutes.

Regards,
Daryl Jones