Some very thought-provoking ideas about OS architecture.
Mon, 21 Jun 1999 11:54:21 -0400

Reminder: please reply to as well as

>I believe that you get 100% write throughput, but in many systems read
>requests are much more frequent than writes. I'd be insterested how good
>is read performance. Wouldn't your data get too fragmented? If you
>hadn't ported any applications to Eros yet, I wonder if you get good read
>performance in real enviroment.

That's a good question.  The answer is that we don't have enough EROS
applications to know for sure, but that the extent-based allocation strategy for
space appears to get you all of the locality you need for file systems.  For
applications that really care, there are mechanisms architected (but not yet
implemented) to guarantee contiguous allocation.  The disk arm behavior of EROS
differs significantly from that of Linux, so it's not clear that the locality
issues play out in exactly the same way.

>> Here's a third: Had a crash or power-out?  On reboot, the system
>> simply picks up pointers to the last checkpointed state.
>I don't think it is really so important. If system crashes, something is
>bad anyway, and you should cure the causation (fix bugs), not the
>consequence (do quick recovery).

This is a good idea, but you don't want to do it while your customers are on
hold waiting for the computer to reboot.
Jonathan S. Shapiro, Ph. D.
IBM T.J. Watson Research Center
Phone: +1 914 784 7085  (Tieline: 863)
Fax: +1 914 784 7595