[PATCH] Clear the objhash before completing restart, but delay free (v2)
orenl at cs.columbia.edu
Mon Nov 1 10:50:12 PDT 2010
On 11/01/2010 01:01 PM, Dan Smith wrote:
> OL> Does this patch solve the performance problem in freeing all
> OL> the objhash entries upfront (e.g. before letting userspace
> OL> resume) ?
> OL> If so, is there still a performance hit for doing the 'clear'
> OL> portion early before resuming the tasks ? How much does it
> OL> depend on the complexity of the hierarchy being checkpointed ?
> I've actually been unable to detect any real different in performance
> between this and the way it was before. I've run it against several
> different large checkpoints (multiple processes, with sockets, IPC,
> many files open, etc). I have another option for the functionality,
> which is to mark only specific items as needing to be free()'d early,
> but I can't discern a difference between that or this option either.
> What I can discern is that this patch is more likely to prevent
> similar issues in the future by making the behavior of the hash far
> more generally correct, IMHO.
So to be clear: is there a performance issue due to moving the
cleanup - some or all - to before allowing tasks to resume ?
(This is what I understood over IRC).
I agree that it's more correct and safe this way. But I want to
know what the implications of a (or any) fix we make.
More information about the Containers