Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!uunet!mcsun!ukc!cam-cl!tdw
From: tdw@cl.cam.ac.uk (Tim Wilson)
Newsgroups: comp.periphs
Subject: Disc reliability and ECC
Summary: How much can ECCs improve HDA reliability?
Keywords: disc disk reliability MTBF ECC error
Message-ID: <1658@gannet.cl.cam.ac.uk>
Date: 16 Nov 89 20:36:31 GMT
Sender: news@cl.cam.ac.uk
Reply-To: tdw@cl.cam.ac.uk (Tim Wilson)
Organization: U of Cambridge Comp Lab, UK
Lines: 42

I am looking for information about calculated or measured hard error
rates from hard discs.  I am conducting some experiments on the
reliability of non-volatile (ie battery backed) primary memory, and
would like to compare my results with the more normal form of
non-volatile memory.

When I pick up a disc manual, I read that the hard error rate is (say)
less than 10 bits in 10 E -14.  I then read that the ECC can correct
up to (say) one burst of up to ten corrupted bits.  What I have never
seen stated is the final hard error rate *after* application of
ECC---the unrecoverable error rate as seen by the operating system.

== More detail on the experiments, in case you'd like to know ==

Non-volatile primary memory (NVM) has some appealing possibilities as
a component of file systems.  It could be used to eliminate the danger
of inconsistency resulting from write-behind caches;  it could be used
for transaction commit records, eliminating the disc latency.

NVM has been proposed as a component of file systems on several
occasions, but as far as I know, no one has tried it because of the
danger that when the system crashes, it may corrupt the NVM.
Operating systems are assumed not to exibit fail-stop characteristics.

So I am repeatedly deliberately crashing a toy op sys, and then
checking to see whether it has corrupted its NVM (preliminary results
suggest that it does happen, occasionally).  I also have some ideas on
how to try and prevent the op sys writing the NVM by accident using
the MMU (the principle of NVM protection is not new, either).

How reliable should I make the NVM subsystem?  An obvious target is to
better the disc which the NVM is buffering---hence my interest in hard
read error rate.

Thanks in advance for any info,

Tim
---
Tim Wilson |                    tdw@uk.ac.cam.cl | U of Cambridge Computer Lab,
           |      ...!uunet!mcvax!ukc!cam-cl!tdw | New Museums Site, Pembroke
Research   | tdw%cl.cam.ac.uk@nsfnet-relay.ac.uk | St, CAMBRIDGE, UK, CB2 3QG
assistant  | +44 223 334624;  Fax +44 223 334679 | (But not speaking for them)