[OpenAFS] Problem with afsd -stat

[Benjamin Chardi Marco]

Dear AFS Experts,

we need your help about -stat option of the AFS daemon (afsd).

In our investigation, we have seen that with increasing value of -stat
option cache performance was actually going down.

Does anyone have an explanation of that ?

We based the disk AFS cache investigation on a read test of files from
several different AFS servers. We copied 200 000 files, sequentially
reading file names from previously randomly generated list.
The list was generated as 200 000 iterations to select 1 file
from a pool of 80 000 files (with size range: 1 KB - 20 MB).
Cache size (= 1 GB) and chuck size (= 2^18) were of course constant
during the test (together with all other options). We have seen
that during the test we transfered about 3 GB of data - so the disk
cache was full

As a performance indicator, we used ratio between:
	dcache Hits / dcache Misses
which corresponds to a ratio of:
     number of files found in the cache / number of files NOT found in the cache
The values were taken using afsmonitor program.

This ratio proved to be accurate confirming that:

 - with bigger cache, ratio is bigger
 - with bigger chunk size, we found saturation point

But - why was this ratio going down when we increased value of -stat ???

We have seen that will small -stat number (e.g. default = 300) the ratio
was actually very good.

Why is cache manager performing badly when you allocate more memory for
status of files in AFS cache ?

Can someone explain this ?

Thank you in advance for any comments.

Cheers,

		Benjamin Chardi
		CERN/IT