[prev in list] [next in list] [prev in thread] [next in thread] 

List:       eros-arch
Subject:    [EROS-Arch] Re: capabilities essay appreciated and OS
From:       Bill Frantz frantz () communities ! com
Date:       2000-10-17 21:58:45
[Download RAW message or body]

At 05:15 PM 10/17/00 -0400, Jonathan S. Shapiro wrote:
>2. The stabilization phase is then started. This runs asynchronously in the
>background. By careful design, writes to the checkpoint log are entirely
>sequential -- or rather, they were sequential in KeyKOS and will be again in
>the new log design. Therefore, the limiting factors on the speed of the
>stabilization phase are:
>
>    a) the sustained disk drive write throughput
>    b) interference by reads
>
>Now here we run into a problem, which is that if you drive the disk that
>hard the user cannot get anything done for lack of available I/O *read*
>bandwidth...

Another limit is the CPU time required to do I/O.  This turned out to be
significant on systems of 10 years ago, but given the way CPU performance
has improved in relation to disk bandwidth, it may no longer be a problem.
(The thought experiment is to assume a fast disk with programmed byte
transfer.)  Since the CPU used for I/O is usually at a higher effective
priority than user processes, it can starve them.

The KeyKOS logic was to prefer reads to writes.  The logic was that there
is a process waiting for a read, while nothing much cares exactly when
writes completed.  (There were exceptions for transaction journaling.)

[prev in list] [next in list] [prev in thread] [next in thread]