>> >> I've noticed this before, but today I've sat down and tested. When two
>> >> drives are in heavy use on a CATS it completely locks. For instance if
>> >> either of the two following commands are run separately, everything works
>> >> fine. If both are run at the same time the machine will completely lock.
>> 
>> >This affects diskful Sharks too, although much less so with -current
>> >compared to the 1.4 branch. It has also been reported on some i386
>> >installations.
>> 
>> My CATS will lock up under heavy use with just one drive.
>> 
>> I can reproduce it every time if I try to build the Jikes Java compiler.

>I seem to recall a similar discussion with ARMLinux on the CATS, so maybe
>there is a hardware element to the problem as well as an OS-related
>problem. Perhaps someone who remembers better than I could elaborate? I've
>also seen the problem on a RiscPC under heavy swapping which I put down to
>UVM funnies (as I would have with the Jikes example above). However, the
>tests that I did were simple extended reads from the raw device. The
>problem certainly seems to be internal IDE related. Has the problem been
>seen with SCSI on a CATS or with a Promise IDE card?

I wouldn't describe my Jikes lockup as being under heavy swapping.

I left top running on the console while I tried running the build. The
lockup happened just as the system started to swap with 3Mb out of
200Mb used. The cc1plus process had been at 99%CPU for a couple of
minutes before this point. I was building over NFS as well to reduce
local disk io.

How much memory does your dd test use for disk buffers ?

Robert