[gomd-devel] Re: Chpox support in gomd

[Gian Paolo Ghilardi]

Hi Olexander.

First of all, thanks for the quick reply.


> Hello Gian,
>
> Thank you for your interest to chpox project and for wish
> to integrate support of chpox into gomd. Good userlevel
> support should extend the capabilities of
> checkpoint/restart and usage convenience. I like your
> development plan and wish to make some suggestions.

Your comments are always welcome. Thanks!


>
> I think that the first feature should be implemented is support of
periodic
> operations (registering, checkpointing, rotation of checkpoint files). So
your
> Phase#1 should be the most rational. Probably one thread is enough
> because registering and sending checkpoint signals are rather fast.
>

Yes, you're right. Moreover I think it's safer to manage a single thread
than several ones.
As checkpointing is important, gomd must support it in the safest way.


> I also think that it is necessary to provide some kind of checkpoint files
> versions for the posibility that user can choose not the last checkpoint
but
> some earlier.

The infos of every registered process are included in a chpoxProc object.
The idea is to provide a counter to trace the number of dumps (and related
dumps filename) and autoclean the oldest dumps.

>This also should provide aditional backup copies in the
> case your last checkpoint becomes broken. At first we thought to
> integrate it with chpox, but then decided that this is not the kernel
> business and it is better to implement it in userlevel.
> Another thing I can suggest
> is automatic restart of checkpointed process after system
> reboot. Many users should like this feature.

Wow! This is a really nice idea (and maybe I know how to implement that).


>
> There are also may
> be some problems. The first is concerned with registering. If you need to
register a
> single process - it is quite easy. Old chpox versions supported only
single processes.
> Now chpox was extended
> to support of communicating processes. So it is necessary to specify
> several process that need to be checkpointed togather. Now this can
> be done by registering the parent pid and setting the flag to
> checkpoint chlidren.

Yet supported. The user can define the dump mechanism via a macro in
constants.h file (CHPOX_REGISTER_MECHANISM). At the moment this macro is
hard-coded but in future the user will be able to define the mechanism at
runtime.
Please refer to
http://savannah.nongnu.org/cgi-bin/viewcvs/gomd/gomd/daemon/constants.h?rev=
1.22&content-type=text/vnd.viewcvs-markup.


 >It is enough for communication via pipes (in
> most cases). But now we are working on support of sockets on single
> machine. It should be necessary to checkpoint togather  processes that
> has no common parent.
>So registering mechanism may be changed.
> Another problem, that checkpointing/restart may be a large security hole
> and it is necessary to remember about it.
Yes. I'll discuss with other developers about this critical issue.
Security is important.

>
> If we have new ideas we shall inform you. If you have any questions
> fell free to contact me.
>
> Regards
> Olexander Sudakov
>

Thanks a lot for your help.

Gian Paolo Ghilardi
Gomd Team



PS: yesterday the first code to support chpox was committed to CVS (a few
hours of work => code could be ugly to see). If you have a few minutes to
spend on, can you check and comment the registerProcess() function in the
file below? I'd like to know if the steps are correct. Thanks.
http://savannah.nongnu.org/cgi-bin/viewcvs/gomd/gomd/daemon/chpoxSupport.cpp
?rev=1.2&content-type=text/vnd.viewcvs-markup