lpsm!
H. Peter Anvin
hpa at zytor.com
Thu Jun 1 22:15:33 PDT 2006
Jacques Mony wrote:
> How do you deal with the filesystem commits? We would need a way to
> make sure the checkpoint is totally written to disk before we consider
> it is. Is there a user space solution to this?
The commits are handled by forking the process (thereby creating a
snapshot), then writing the log, and using fsync and fdatasync() to
enforce coherency.
Both blocking and nonblocking checkpointing is available; blocking
checkpointing obviously stops everything until the checkpoint is
complete; nonblocking checkpointing allows execution to continue, but
you can *always* revert to a checkpoint; for nonblocking checkpoints
this may be the one before the latest instead of the latest if the
checkpoint wasn't complete by the time the process was terminated.
-hpa
More information about the LPSM
mailing list