[rc5] Why Isn't RC5 Client Fail-Soft?
mikus at bga.com
Sat Jul 26 16:16:15 EDT 1997
In the past couple of weeks, while working with things other than
RC5, I have twice managed to "crash" my system - i.e., the only
"recovery" was to re-boot. Both times, the current block that RC5
was working on was "lost" -- it had ALREADY been removed from
buff-in.rc5, but due to the crash it was NEVER recorded - neither
out on buff-out.rc5 (finished) nor back on buff-in.rc5 (checkpointed).
I imagine the way it works was designed to support multiple copies
of RC5 runing simultaneously - if the block is removed from buff-in
as soon as it is "picked up", no other copy will "pick up" that same
block. But what if the client that _has_ the block never gets the
chance to record its status?
Let me suggest using one more file - buff-work.rc5. Then "picking
up" a block would mean moving it from buff-in to buff-work, and
"putting down" a block would mean moving it from buff-work to
buff-out. And "checkpointing" would mean removing the block from
buff-work, as well as adding it again to buff-in.
With this change, those who did not wish to re-process blocks would
start RC5 normally, and it would only look at buff-in. But a new
parameter (-restart) would allow an operator, after a system crash,
to first move the blocks (if any) from buff-work to buff-in, before
starting decoding -- thereby "recovering" those blocks which were
neither finished nor checkpointed before the system went down.
To unsubscribe, send email to majordomo at llamas.net with 'unsubscribe rc5' in the body.
More information about the rc5