Re: System tuning
On Thu, 21 Mar 1996, Dave Kosenko wrote:
> > of the outputs from the onstat command. Most recently, we increased our
> > BUFFERS value, and the output from onstat -g glo changed from having all
> > the time go to AIO vp's to CPU vp's. Can anyone decipher this occurance?
> > Does that mean that we WERE blocked on AIO, but now are blocked on CPU?
>
> It means that with fewer buffers to hold data, you had to read more of the data
> from disk. AIO vps do the job of reading/writing data from/to disk. With the
> increased number of buffers, you are simply doing LESS I/O.
This is a good thing! It's nice to have SOMETHING positive happening.
Interestingly, our number of "longspins" - whatever that might be - seems
to have increased when we made our last change as well :(
> I can't say whether or not you are "blocked" on CPU - does vmstat/sar show any
> idle time? If not, then you are out of cpu. If there is idle/wait time, then
> something else is the bottleneck.
Periodically our idle dips to 0.
Here is the report from onstat -p
RSAM Version 7.10.UC2 -- On-Line -- Up 1 days 05:10:21 -- 53720 Kbytes
Profile
dskreads pagreads bufreads %cached dskwrits pagwrits bufwrits %cached
48795 131828 76540974 99.94 82552 186815 2560196 96.78
isamtot open start read write rewrite delete commit rollbk
39610206 1661035 3989669 28796484 906462 52514 4036 12672 7
ovtbls ovlock ovuserthread ovbuff usercpu syscpu numckpts flushes
0 0 0 0 24557.51 3778.06 208 682
bufwaits lokwaits lockreqs deadlks dltouts ckpwaits compress seqscans
8501 81 97569075 3 0 272 53004 35714
ixda-RA idx-RA da-RA RA-pgsused lchreqs lchwaits
13661 96 11130 24824 671793347 175750
I can include the values that we have in ONCONFIG, or any of the onstat
outputs. if that would help anyone make some sense of this.
This is on a system that has between 20-35 informix users, 128Mb RAM,
2x88100 CPU's running DG/UX5.4R3.10 The database is on 5x1Gb RAID5
disks. Does it seem immediately obvious to anyone as to what we might be
missing here? Our system DOES NOT run as fast as it did with 5.x Our
users aren't terribly happy these days :(