At 03:57 AM 5/6/99 -0700, Mark S. Miller wrote:
>1) Cheaper Commitment.
>
>Databases often use this technique by writing a "transaction log". Once a
>transaction request has been successfully logged, the transaction can be
>committed even though a crash may destroy its result state, given that the
>transaction can be deterministically reproduced from the logged request. I
>first saw this technique applied to committing general purpose
>computation by Rob Strom: Do a full checkpoint occasionally, and cheaply
>log inputs until you take the next checkpoint.
>
>It's not clear whether this wins over KeyKOS/EROS mmu-based checkpointing,
>since that's also incremental. But is sure beats the crap out of
>non-incremental checkpointing, which is the only reasonable alternative for
>those sitting above the jvm.
[#]KeyKOS used the technique of transaction logs in conjunction with mmu-based checkpointing to achieve high transaction rates. See http://www.cis.upenn.edu/~KeyKOS/KeyTXF/KeyTXF.html