Tuning aha1540 driver (ISC 2.2)

Heiko Blume src at scuzzy.mbx.sub.org
Sat Aug 25 22:03:04 AEST 1990


larry at focsys.uucp (Larry Williamson) writes:


>A few weeks back (late june, I think) there was a thread here on
>tuning the scsi driver for ISC Unix 2.2. I saved most of those
>articles since we were planning on installing 1542A's on some machines
>here.
>My tests seem to indicate that the default values for BUSON, BUSOFF
>and DMASPEED are the fastest. 

they are not. however this (partly) depends on the scsi devices you have.
you *will* experience higher throughput with disks that do read-ahead, caching
etc. provided you have taken care of various other things (read on).
it also depends on the other cards you have installed. by increasing a
cards buson time (and/or decreasing busoff) you are stealing bus cycles from
the other cards. this can cause problems with floppy controllers etc (you'll
get 'controller timeout error' etc.). you should ONLY do this if your bus
has enough total cycles, i.e. is running fast enough to have time for other
card's DMA. if your bus runs at 6/8MHz the adaptec tends to monopolize the bus
during heavy disk usage if you set DMASPEED and BUSON too high (and BUSOFF
too low). especially funny if you use ethernet cards...
you should configure your motherboards bus speed to 10 or 12 MHz, and, of
course, you must not set the adaptec's DMA speed to anything higher or equal
to the motherboards bus speed.
also note that no board until now groks 10 MB/s DMA (you can set the old
aha-1540 to that :-). therefor the max settable speed of the 1540A is 
8 MB/s.

>I've installed the Sync option jumper on the controller. 

that jumper doesn't matter - it's setting is overridden by ISC's scsi
driver.

>With BUSON 5, BUSOFF 9 and DMASPEED 0; the system seems to be
>performing just fine. 

sure, these are the 'works 100%' parameters. not the 'works 100% and
performs 100%' paramters.

>I'm not entirely sure how to modify the DMASPEED parameter. Values of
>0 and 1 seem to work, but 2 causes the system to enter the kernel
>debugger at boot time. 

since you didn't say what your motherboards bus speed is, i guess you
got a bus error in the kernel ?! that is, the 6.7 MB/s rate was too much.

>I've not bothered trying a value of 3. How do
>these values relate to the DMA speeds of 5, 5.7, 6.7 and 8 Mhz that
>the controller is supposed to be able to support? Is it the obvious
>0->5, 1->5.7, 2->6.7, 3->8?

correct.

>The test bed system I'm using is an AMI MB with 4 Meg ram, 64K cache
>and 20 Mhz clock.

note that motherboards with c&t chipset have(had?) problems with DMA,
for instance my previous board (also ami bios) panic'ed the kernel
during high volume disk transfers when i set the board to 24Mhz.

>I'm bothered that if I modify these values in /etc/conf/cf.d/[ms]tune,
>I get compiler warnings about these values being redefined. So I've
>been modifying them in /usr/include/sys. I don't like this. What is
>the normally accepted practice? Ignore the warnings? I hope not.

perhaps here's the reason why you didn't see any performance improvement.
which settings are actually used?? i don't know if stune comes before
aha1540.h. a workaround would be to set the parameters in stune and
put #ifndefs around the ones in aha1540.h *if* stune comes first.

>During these tests I've left the DMA speed selection header on the
>board open. This selects the default of 5 Mhz. I'm going to try
>other selections on the jumpers. Is there a reason why this is not a
>good idea?

these jumper settings are also overridden by the software.

>Any advise you can offer would be greatly appreciated. Thanks.

you're welcome. don't give up, it's worth it! however, you might have
to do some serious calculations if your machine has, say, a ethernet card
a CPU graphics card, a multiport card etc.
especially if you use the machine as a nfs server you may encounter
starvation problems if the network card's DMA doesn't get the share it needs.

good luck!
-- 
Heiko Blume c/o Diakite   blume at scuzzy.mbx.sub.org    FAX   (+49 30) 882 50 65
Kottbusser Damm 28        blume at netmbx.UUCP           VOICE (+49 30) 691 88 93
D-1000 Berlin 61          blume at netmbx.de             TELEX 184174 intro d
scuzzy Any ACU,e 19200 6919520 ogin:--ogin: nuucp ssword: nuucp



More information about the Comp.unix.i386 mailing list