Re: onstat -g ioq ??
On Wed, 28 Jan 1998, Jeremy Hawkey wrote:
> I 'm sorry to be such a peasent, but does this mean; looking at the
> activity on these gfd's I need to add another AIO vp or not. In the
> manual it says if maxlen or len is high add another aio, I think it is
> referring only to the aio being high... or does it mean the gfd's also??
>
> Hope you can help, and thank you for such a prompt reply.
You cannot tell from the onstat -g ioq run onstat -g iov to determine
whether you have too many or too few aio VPs. Here is the tail end of one
of my listings:
INFORMIX-OnLine Version 7.21.UD3 -- On-Line -- Up 4 days 03:56:19 --
1164632 Kbytes
AIO I/O vps:
class/vp s io/s totalops dskread dskwrite dskcopy wakeups io/wup
msc 0 i 0.2 62386 0 0 0 62287 1.0
aio 0 s 37.8 13616480 12469735 1144706 0 11945301 1.1
aio 1 s 35.6 12804384 11653234 1150001 0 11302437 1.1
[SNIP]
aio 134 i 0.0 2153 548 1605 0 2042 1.1
aio 135 i 0.0 2117 555 1562 0 1994 1.1
aio 136 i 0.0 1981 506 1475 0 1923 1.0
aio 137 i 0.0 1892 514 1378 0 1875 1.0
aio 138 i 0.0 1848 440 1408 0 1829 1.0
aio 139 i 0.0 1779 433 1346 0 1753 1.0
aio 140 i 0.0 1835 459 1376 0 1708 1.1
aio 141 i 0.0 1822 430 1392 0 1667 1.1
aio 142 i 0.0 1745 429 1316 0 1626 1.1
aio 143 i 0.0 1701 452 1249 0 1578 1.1
aio 144 i 0.0 1656 419 1237 0 1500 1.1
aio 145 i 0.0 1678 398 1280 0 1489 1.1
aio 146 i 0.0 1695 407 1288 0 1457 1.2
aio 147 i 0.0 1583 384 1199 0 1374 1.2
pio 0 i 0.3 94851 0 94851 0 94852 1.0
lio 0 i 6.6 2376954 0 2376954 0 2376759 1.0
You look at two things
1) The difference in total ops as you look down from aio VP 0 through aio
VP N. At some point there is a dramatic drop off if the total ops is near
zero you do not need these during normal processing but may need them to
speed peak load. If you are interested in improving peak load performance
then look at:
2) The last column (io/wp) #I/Os per wakeup. If this is >1.0 for the last
several aio VPs then they are probably being awakened during peak load and
when they finish their assigned work there is more work waiting for them
so not only are these aio VPs needed you may be able to take good
advantage of even more of them. If this number is <1.0 then sometimes
when they are awakened by the time they shake the sand out of their eyes
another aio VP (one already awake) has already performed the I/O for which
it was needed so it has gone back to sleep without doing anything. In
this case you can probably reduce the number of aio VPs.
Rule of Thumb: LRU == CLEANERS <= NUMAIOVP (NUMAIOVP adjusted as needed
using the 1.5 * #chunks rule and onstat -g iov analysis)
Art S. Kagel, kagel@bloomberg.com