Skip site navigation (1)Skip section navigation (2)
Date:      Mon, 9 Dec 1996 16:33:49 +1100
From:      Bruce Evans <bde@zeta.org.au>
To:        bde@zeta.org.au, proff@suburbia.net
Cc:        hackers@freebsd.org
Subject:   Re: strange problems with recent current
Message-ID:  <199612090533.QAA17144@godzilla.zeta.org.au>

next in thread | raw e-mail | index | archive | help
>P100/Triton mother board. 4 Triton/mb IDE drives, two scsi (ISA
>adaptec), average 120 procs, 72Mb swap used (striped over the 4
>ide drives). This configuration hasn't changed, and I don't use the
>console keyboard, infact..the console keyboard has been unplugged
>(could this have any effect?)

Not using the keyboard can only help.

All those IDE drives are bad for the non-fast-interrupt (PCI)
cy driver, at least if multi-mode is enabled.  It takes 0.5 ms to
transfer a typical 8K multi-block at 16MB/sec.  Any more than that
will cause silo overflows.  Active drives on both controllers can
certainly cause more than that, and very fast drives may be able to
hog the cpu for any number of transfers.

>> One per day is too many.
>
>cy15 is often going at 57.6k in both directions, the others are more
>sporadic typical dial-up-user traffic (at 38.4). I moved cy15 onto sio1
>(on board 16550A uart)

I would expect it to work up to 16 * 57.6k going in both directions.
In raw mode, this takes about 70% overhead on a 486/33, about 40% of which
is bus i/o overhead, 20% interrupt overhead, and 40% software overhead.
Cooked mode has a much larger software overhead (5-10 times).  A P100
should be enough faster to run some lines in cooked mode and actually
do something besides serial i/o.  However, if the PCI bus interface is
50% slower like David mentioned, then the bus i/o alone will take 60%
of a P100 instead of only 40% :-].

>Dec  9 14:58:16 suburbia /kernel: sio1: 1 more silo overflow (total 17190)
>Dec  9 14:58:21 suburbia /kernel: cy0: 2 more silo overflows (total 220)

Not good.

>This is actually quite a bit better than the cyclades (could just be the
>12 vs 15 byte byffer) but still worse than the cyclades before the
>"slow down".

sio actually only uses 2 or 3 bytes of buffering above the interrupt
threshold (the threshold is 14), while cy uses 6 bytes (the threshold is
6).  The source must be edited to change the thresholds.

>Load average is typically around 0.16

Load average doesn't show interrupt load.  Use systat to see interrupt
load for non-fast-interrupt handlers.  Nothing shows the load for
fast-interrupt handlers properly.

Bruce



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199612090533.QAA17144>