[Devel] Re: [RFC][PATCH 2/2] CR: handle a single task with private memory maps

[Oren Laadan]

Louis Rilling wrote:
> On Fri, Aug 01, 2008 at 10:15:26AM -0400, Oren Laadan wrote:
>>
>> Louis Rilling wrote:
>>> On Thu, Jul 31, 2008 at 03:12:32PM -0400, Oren Laadan wrote:
>>>> Louis Rilling wrote:
>>>>> On Thu, Jul 31, 2008 at 12:28:57PM -0400, Oren Laadan wrote:
>>>>>> Louis Rilling wrote:
>>>>>>> On Thu, Jul 31, 2008 at 11:09:54AM -0400, Oren Laadan wrote:
>>>>>>>> Louis Rilling wrote:
>>>>>>>>> On Wed, Jul 30, 2008 at 06:20:32PM -0400, Oren Laadan wrote:
>>>>>>>>>> Serge E. Hallyn wrote:
>>>>>>>>>>> Quoting Oren Laadan (orenl at cs.columbia.edu):
>>>>>>>>>>>> +int do_checkpoint(struct cr_ctx *ctx)
>>>>>>>>>>>> +{
>>>>>>>>>>>> +	int ret;
>>>>>>>>>>>> +
>>>>>>>>>>>> +	/* FIX: need to test whether container is checkpointable */
>>>>>>>>>>>> +
>>>>>>>>>>>> +	ret = cr_write_hdr(ctx);
>>>>>>>>>>>> +	if (!ret)
>>>>>>>>>>>> +		ret = cr_write_task(ctx, current);
>>>>>>>>>>>> +	if (!ret)
>>>>>>>>>>>> +		ret = cr_write_tail(ctx);
>>>>>>>>>>>> +
>>>>>>>>>>>> +	/* on success, return (unique) checkpoint identifier */
>>>>>>>>>>>> +	if (!ret)
>>>>>>>>>>>> +		ret = ctx->crid;
>>>>>>>>>>> Does this crid have a purpose?
>>>>>>>>>> yes, at least three; both are for the future, but important to set the
>>>>>>>>>> meaning of the return value of the syscall already now. The "crid" is
>>>>>>>>>> the CR-identifier that identifies the checkpoint. Every checkpoint is
>>>>>>>>>> assigned a unique number (using an atomic counter).
>>>>>>>>>>
>>>>>>>>>> 1) if a checkpoint is taken and kept in memory (instead of to a file) then
>>>>>>>>>> this will be the identifier with which the restart (or cleanup) would refer
>>>>>>>>>> to the (in memory) checkpoint image
>>>>>>>>>>
>>>>>>>>>> 2) to reduce downtime of the checkpoint, data will be aggregated on the
>>>>>>>>>> checkpoint context, as well as referenced to (cow-ed) pages. This data can
>>>>>>>>>> persist between calls to sys_checkpoint(), and the 'crid', again, will be
>>>>>>>>>> used to identify the (in-memory-to-be-dumped-to-storage) context.
>>>>>>>>>>
>>>>>>>>>> 3) for incremental checkpoint (where a successive checkpoint will only
>>>>>>>>>> save what has changed since the previous checkpoint) there will be a need
>>>>>>>>>> to identify the previous checkpoints (to be able to know where to take
>>>>>>>>>> data from during restart). Again, a 'crid' is handy.
>>>>>>>>>>
>>>>>>>>>> [in fact, for the 3rd use, it will make sense to write that number as
>>>>>>>>>> part of the checkpoint image header]
>>>>>>>>>>
>>>>>>>>>> Note that by doing so, a process that checkpoints itself (in its own
>>>>>>>>>> context), can use code that is similar to the logic of fork():
>>>>>>>>>>
>>>>>>>>>> 	...
>>>>>>>>>> 	crid = checkpoint(...);
>>>>>>>>>> 	switch (crid) {
>>>>>>>>>> 	case -1:
>>>>>>>>>> 		perror("checkpoint failed");
>>>>>>>>>> 		break;
>>>>>>>>>> 	default:
>>>>>>>>>> 		fprintf(stderr, "checkpoint succeeded, CRID=%d\n", ret);
>>>>>>>>>> 		/* proceed with execution after checkpoint */
>>>>>>>>>> 		...
>>>>>>>>>> 		break;
>>>>>>>>>> 	case 0:
>>>>>>>>>> 		fprintf(stderr, "returned after restart\n");
>>>>>>>>>> 		/* proceed with action required following a restart */
>>>>>>>>>> 		...
>>>>>>>>>> 		break;
>>>>>>>>>> 	}
>>>>>>>>>> 	...
>>>>>>>>> If I understand correctly, this crid can live for quite a long time. So many of
>>>>>>>>> them could be generated while some container would accumulate incremental
>>>>>>>>> checkpoints on, say crid 5, and possibly crid 5 could be reused for another
>>>>>>>>> unrelated checkpoint during that time. This brings the issue of allocating crids
>>>>>>>>> reliably (using something like a pidmap for instance). Moreover, if such ids are
>>>>>>>>> exposed to userspace, we need to remember which ones are allocated accross
>>>>>>>>> reboots and migrations.
>>>>>>>>>
>>>>>>>>> I'm afraid that this becomes too complex...
>>>>>>>> And I'm afraid I didn't explain myself well. So let me rephrase:
>>>>>>>>
>>>>>>>> CRIDs are always _local_ to a specific node. The local CRID counter is
>>>>>>>> bumped (atomically) with each checkpoint attempt. The main use case is
>>>>>>>> for when the checkpoint is kept is memory either shortly (until it is
>>>>>>>> written back to disk) or for a longer time (use-cases that want to keep
>>>>>>>> it there). It only remains valid as long as the checkpoint image is
>>>>>>>> still in memory and have not been committed to storage/network. Think
>>>>>>>> of it as a way to identify the operation instance.
>>>>>>>>
>>>>>>>> So they can live quite a long time, but only as long as the original
>>>>>>>> node is still alive and the checkpoint is still kept in memory. They
>>>>>>>> are meaningless across reboots and migrations. I don't think a wrap
>>>>>>>> around is a concern, but we can use 64 bit if that is the case.
>>>>>>>>
>>>>>>>> Finally, the incremental checkpoint use-case: imagine a container that
>>>>>>>> is checkpointed regularly every minutes. The first checkpoint will be
>>>>>>>> a full checkpoint, say CRID=1. The second will be incremental with
>>>>>>>> respect to the first, with CRID=2, and so on the third and the forth.
>>>>>>>> Userspace could use these CRID to name the image files (for example,
>>>>>>>> app.img.CRID). Assume that we decide (big "if") that the convention is
>>>>>>>> that the last part of the filename must be the CRID, and if we decide
>>>>>>>> (another big "if") to save the CRID as part of the checkpoint image --
>>>>>>>> the part that describe the "incremental nature" of a new checkpoint.
>>>>>>>> (That part would specify where to get state that wasn't really saved
>>>>>>>> in the new checkpoint but instead can be retrieved from older ones).
>>>>>>>> If that was the case, then the logic in the kernel would be fairly
>>>>>>>> to find (and access) the actual files that hold the data. Note, that
>>>>>>>> in this case - the CRID are guaranteed to be unique per series of
>>>>>>>> incremental checkpoints, and incremental chekcpoint is meaningless
>>>>>>>> across reboots (and we can require that across migration too).
>>>>>>> Letting the kernel guess where to find the missing data of an incremental
>>>>>>> checkpoint seems a bit hazardous indeed. What about just appending incremental
>>>>>>> checkpoints to the last full checkpoint file?
>>>>>> It isn't quite a "guess", it's like the kernel assumes that a kernel-helper
>>>>>> resides in some directory - it's a convention. I agree, though, that it may
>>>>>> not be the best method to do it.
>>>>>>
>>>>>> As for putting everything in a single file, I prefer not to do that, and it
>>>>>> may not even always possible I believe.
>>>>>>
>>>>>> An incremental would include a section that describes how to find the missing
>>>>>> data from previous checkpoints, so it must have a way to identify a previous
>>>>>> checkpoint.
>>>>>>
>>>>>> On way is like I suggested name them with this identifier, another would be,
>>>>>> for example, that the user provides a list of file-descriptors that match
>>>>>> the required identifiers. Other ways may be possible too.
>>>>>>
>>>>>> In any event, I think it is now  bit early to discuss the exact format and
>>>>>> logic, when we don't even have a simple checkpoint working :)
>>>>>>
>>>>>> Incremental checkpoint is one of a few reasons to use CRIDs, let us first
>>>>>> agree about CRIDs, and later, when we design incremental checkpoints, decide
>>>>>> on the technical details of incorporating this CRIDs.
>>>>>>
>>>>> Agreed, but since your point is to introduce CRIDs, I'd like to be convinced
>>>>> that they are needed :) At least I'd like to be convinced that they will not
>>>>> generate hard-to-manage side effects.
>>>>>
>>>>>> (Just to avoid confusion, an incremental checkpoint is _not_ a pre-copy or
>>>>>> live-migration: in a pre-copy, we repeatedly copy the state of the container
>>>>>> without freezing it until the delta is small enough, then we freeze and then
>>>>>> we checkpoint the remaining residues. All this activity belongs to a single
>>>>>> checkpoint. In incremental checkpoints, we talk about multiple checkpoints
>>>>>> that save only the delta with respect to their preceding checkpoint).
>>>>> Don't worry, I know what incremental checkpointing is.
>>>>>
>>>>>>>> We probably don't want to use something like a pid to identify the
>>>>>>>> checkpoint (while in memory), because we may have multiple checkpoints
>>>>>>>> in memory at a time (of the same container).
>>>>>>> Agreed.
>>>>>>>
>>>>>>>>> It would be way easier if the only (kernel-level) references to a checkpoint
>>>>>>>>> were pointers to its context. Ideally, the only reference would live in a
>>>>>>>>> 'struct container' and would be easily updated at restart-time.
>>>>>>>> Consider the following scenario of calls from user-space (which is
>>>>>>>> how I envision the checkpoint optimized for minimal downtime, in the
>>>>>>>> future):
>>>>>>>>
>>>>>>>> 1)	while (syscall_to_do_precopy)		<- do precopy until ready to
>>>>>>>> 		if (too_long_already)		<- checkpoint or too long
>>>>>>>> 			break;
>>>>>>>>
>>>>>>>> 2)	freeze_container();
>>>>>>>>
>>>>>>>> 3)	crid = checkpoint(.., .., CR_CKPT_LAZY);	<- checkpoint container
>>>>>>>> 							<- don't commit to disk
>>>>>>>> 							<- (minimize owntime)
>>>>>>>>
>>>>>>>> 4)	unfreeze_container();			<- now can unfreeze container
>>>>>>>> 						<- already as soon as possible
>>>>>>>>
>>>>>>>> 5)	ckpt_writeback(crid, fd);		<- container is back running. we
>>>>>>>> 						<- can commit data to storage or
>>>>>>>> 						<- network in the background.
>>>>>>>>
>>>>>>>> #2 and #4 are done with freezer_cgroup()
>>>>>>>>
>>>>>>>> #1, #3 and #5 must be syscalls
>>>>>>>>
>>>>>>>> More specifically, syscall #5 must be able to refer to the result of syscall #3
>>>>>>>> (that is the CRID !). It is possible that another syscall #3 occur, on the same
>>>>>>>> container, between steps 4 and 5 ... but then that checkpoint will be assigned
>>>>>>>> another, unique CRID.
>>>>>>> Hm, assuming that, as proposed above, incremental checkpoints are stored in the
>>>>>>> same file as the ancestor full checkpoint, why not simply give fd as argument in
>>>>>>> #5? I'd expect that the kernel would associate the file descriptor to the
>>>>>>> checkpoint until it is finalized (written back, sent over the wire, etc.).
>>>>>> The above procedure, step 1-5 are for a _single_ checkpoint.
>>>>> This is what I understood.
>>>>>
>>>>>> Why would the kernel associate a file descriptor with the checkpoint until it
>>>>>> is finalized ?   As far as I'm concerned, the checkpoint call in step 3 can go
>>>>>> without any FD.  Also, what happens if there is another checkpoint, of the
>>>>>> same container, taken between steps 4 and 5, how would you tell the difference
>>>>>> or select which one goes in first ?   Finally, keeping that FD alive between
>>>>>> multiple checkpoints would require the checkpointer (e.g. a daemon that will
>>>>>> periodically checkpoint) to keep it alive.
>>>>>>
>>>>>> I view it differently: a checkpoint held in memory is like a kernel resource,
>>>>>> and requires a handle/identifier for user space to refer to it. Like an IPC
>>>>>> object. Why tie that object to a specific file descriptor ?
>>>>>> The only exception I can see, is the need to tie it to a some process - the
>>>>>> checkpointer for instance, such that if that process dies without completing
>>>>>> the work, the checkpoint image in memory will be cleaned up.
>>>>>> That, however, still is problematic, because it will not allow you to use
>>>>>> different procesess for different steps (above).
>>>>>>
>>>>>> Since we are not yet optimizing the checkpoint procedure, just building the
>>>>>> infrastructure, my goal is to convince that a CRID is a desired feature (and
>>>>>> I can certainly see how it will be used in various scenarios).
>>>>> Here is probably the source of the misunderstanding. I was assuming that step #3
>>>>> needed a file descriptor to dump the checkpoint progressively, but reading your
>>>>> first use-case more carefully might have avoided this misunderstanding :)
>>>> Even without the first use-case (checkpoint in memory), step 3 does not need
>>>> necessarily a file-descriptor to which data will be dumped, in the case of
>>>> said optimization. Consider a scenario with periodic checkpointing of a long
>>>> running application, where we would like to minimize the downtime of the
>>>> application due to each checkpoint. The idea is to do steps 1 and 3 entirely
>>>> in memory, keep the data in a buffer (see below comment about tmpfs). The
>>>> expensive operation of streaming the data to the file-descriptor is only
>>>> done in step 5.
>>>>
>>>> (In the case of checkpoint in memory - it is never written to a file. There
>>>> are various optimization to do there for fast restart for which putting the
>>>> data in a file doesn't make sense).
>>>>
>>>> As for using tmpfs -- so during step 3 the state of all tasks is saved; part
>>>> of it is headers, task data, signals etc, but mostly the memory content. For
>>>> as long as the checkpoint is kept in memory (either because it is meant to
>>>> stay there, or because it is not committed to the file-descriptor yet), there
>>>> is no reason to make a copy of each (dirty) page. On the contrary - the pages
>>>> will be marked COW and a reference will be kept, as part of the checkpoint
>>>> context. Sure, you can put the rest of the data in a file in tmpfs; but you
>>>> probably don't want to copy all the pages to a file in tmpfs - that would be
>>>> wasteful.
>>> I think that memory pages need not to be dumped in step #3. They can be kept
>>> just as you mentioned in COW state in the checkpoint context, and be really
>>> dumped only in step #5.
>>>
>>>>> Anyway, we can still give a fd to sys_checkpoint() which will identify the
>>>>> checkpoint for the remaining operations. It's up to userspace to show the
>>>>> difference between two checkpoints taken (roughly) at the same time. From the
>>>>> kernel point of view, a file descriptor is enough to make the difference.
>>>> That is indeed an option. I haven't given a lot of thought to this approach,
>>>> because in Zap I use CRIDs. Three points against this approach are that:
>>>>
>>>> (1) as I said, that would require that the file descriptor remains alive for
>>>> as long as we want to keep the checkpoint alive (in memory), and
>>> Not sure that this is so bad. The checkpointer can transfer the descriptor
>>> to some daemon using the file descriptor transfer feature of UNIX sockets, and
>>> then freely exit.
>> Uhh.. that's an evil feature to begin with :o
>> In any case, it requires that extra logic.
>>
>>>> (2) if the checkpoint is taken by a process from within the container, we
>>>> create a situation where a resource held by the process (an FD), is referring
>>>> to the checkpoint itself and at the same time also referred to by the
>>>> checkpoint (because it is part of the state of a process that is in the
>>>> container...). In particular this will necessitate some special case treatment
>>>> during the restart operation.
>>> Interesting case. This means that the checkpointer would be checkpointed while
>>> inside sys_checkpoint(), and would possibly try to writeback the checkpoint
>>> after restart (going to step #5 as if it was not restarted). So the special
>>> handling is already needed there, right? Like making sys_checkpoint() return an
>> Not quite. See my first reply to Serge earlier in this thread. sys_checkpoint()
>> returns one of three values:  -1 for error, positive (non zero) number which is
>> the CRID on success, and 0 when it returns from restart. Logic is analogous to
>> a fork() syscall. No special handling, definitely not in kernel space.
> 
> Sorry I had not these details in mind anymore.
> 
> Returning 0 in case of a restart is what I called a special handling. You won't
> do this for the other tasks, so this is special. Since userspace must cope with
> it anyway, userspace can be clever enough to avoid using the fd on restart, or
> stupid enough to destroy its checkpoint after restart.

It's a different "special hanlding" :)   In the case of a single task that wants
to checkpoint itself - there are no other tasks.  In the case of a container -
there will be only a single task that calls sys_checkpoint(), so only that task
will either get the CRID or the 0 (or an error). The other tasks will resume
whatever it was that they were doing (lol, assuming of course restart works).

So this "special handling" ends up being a two-liner: setting the return
value of the syscall for the task that called sys_checkpoint() (well, actually
it will call sys_restart() to restart, and return from sys_checkpoint() with
a value of 0 ...).

If you use an FD, you will have to checkpoint that resource as part of the
checkpoint, and restore it as part of the restart. In doing so you'll need
to specially handle it, because it has a special meaning. I agree, of course,
that it is feasible.

> 
>>> error upon restart. I'm not sure that the checkpoint fd should really need a
>>> special handling in the special case of self-checkpoiting, because the
>>> checkpointer shoud probably not try to do anything with this checkpoint after a
>>> restart, unless it reopens the checkpoint file for appending new incremental
>>> checkpoints.
>>>
>>> Anyway, we are trying to solve an issue that was explicitly forbidden in
>>> previous discussions IIUC, because the whole container is assumed to be frozen
>>> before calling sys_checkpoint(), which means that the checkpointer should live
>>> outside of the container.
>> Actually, I made the point in the mini-summit that such a functionality will be
>> useful, and I have several use cases, and two of them actually implemented
>> with Zap. The main change from a regular, freeze-entire-container checkpoint
>> is that one task - the checkpointer - will be allowed not to freeze. Since
>> it will be doing the checkpoint itself, there is no concern about it not being
>> frozen (after all, we freeze them so they don't change their state). I already
> 
> I had no doubt that self-checkpoint is feasible, since we are doing this in
> Kerrighed (it's a signal that is handled at kernel-level only).
> 
>> implemented this is Zap and it proved quite useful. See this paper, for example:
>> http://www.ncl.cs.columbia.edu/publications/sosp2007_dejaview.pdf
> 
> Nice paper :)
> 
>>>> (3) if a give tasks wants to keep many checkpoints in memory (again, either
>>>> permanently or shortly), it will have to keep, forever, a lot of open file
>>>> descriptors.
>>> The only problem I see here is the limitation on the number of file descriptors.
>>> Hm, hundreds of checkpoints in memory looks like memory wastage in some way.
>> "640K ought to be enough for anybody." - Bill Gates, 1981  (actually, according
>> to this page http://en.wikiquote.org/wiki/Talk:Bill_Gates, it may not have been
>> him at all ...)
>>
>> Now seriously, I have at least one use case (the details weren't published yet).
> 
> So sad! We'll have to wait...
> 
>>>> On the other hand, using an FD provide the advantage of a simple cleanup (FD
>>>> closed -> checkpoint data discarded) and ridding us from the need to come up
>>>> with a cleanup strategy.
>>> We would not get this for free unless we add data for this to the file
>>> descriptor. Adding something like an inotify listener (only used by the kernel)
>>> should also make it.
>> Lol .. then we stick to CRID if we have to implemented something anyway :)
> 
> My comment did not aim at saying "it's bad", and it actually didn't. Just giving
> an idea on how to do it.
> 
>>>>> Let's consider the three use cases of CRID you mentioned earlier:
>>>>>
>>>>> 1) Checkpointing in memory:
>>>>> Actually, checkpointing in memory could also be done from userspace using tmpfs.
>>>>> Again, I agree that this kind of optimization should be discussed later. I'm
>>>>> just not convinced that this needs a CRID...
>>>> See my comment about regarding tmpfs. You are right, however, in that we could
>>>> use FD to tmpfs where the rest of the data (not pages) will be stored.
>>> See my comment above ;)
>>>
>>>>> 2) Reducing downtime of the checkpoint:
>>>>> If reducing downtime is just a matter of avoiding disk accesses, tmpfs is again
>>>>> a kind of solution. It even allows to swap if the checkpoint size is too big.
>>>>> What kind of scenario (other than incremental checkpointing) do you envision
>>>>> where multiple calls to sys_checkpoint() would use the same checkpoint object?
>>>> Again, see the comment regarding tmpfs. The actual memory copy operation between
>>>> the real pages and the space allocated in tmpfs can take substantial time for
>>>> applications with large memory (compared to merely marking the pages COW, and
>>>> amortizing the cost during regular execution of the application), besides the
>>>> extra space overhead. Also, writing tmpfs incurs visible overhead when you care
>>>> about milliseconds of downtime; I've seen that with Zap.
>>> Are those milliseconds related to pages or to the kernel structures also?
>> It's a visible overhead. I can't remember exactly how much because once I saw
>> it was expensive, I dropped that path. Even buffer allocation (page allocation
>> in case of tmpfs) could become an annoyance when it comes to low downtime, so
>> one optimization in Zap was the pre-allocate the buffers using a good estimate
>> on their sizes based on past checkpoints.
>>
>> Finally, there are use-cases in which you'd like a reall-super-ultra-fast
>> checkpoint (e.g. in context), that is under a millesecond (like a partial
>> fork, to some extent); you do feel the difference then.
>>
>>>>> 3) Incremental checkpoint:
>>>>> I agree that maintaing a fd alive (in a checkpointer daemon for instance) may
>>>>> look restrictive, but I'm not sure that it is really needed to keep it alive
>>>>> between consecutive incremental checkpoints. I'd really like to see incremental
>>>>> checkpointing as an append operation to a checkpoint file. This way the file
>>>> Why ?  What's the advantage of having all data in a single file as opposed to
>>>> multiple files ?
>>> - You do not have to look for the previous checkpoints using a to-be-defined
>>>   naming scheme, since they are all in the file.
>> but if you *want* to look for a previous checkpoint -- you wanna return to an
>> arbitrary checkpoint in the past ?  now you need to look for it.
> 
> I think I already sketched how to do it.
> 
>>> - Userspace makes less errors when managing incremental checkpoints.
>> have you implemented this ?  did you experience issues in real life ?  user
>> space will need a way to manage all of it anyway in many aspects. This will
>> be the last/least of the issues ...
> 
> No it was not implemented, and I'm not going to enter a discussion about the
> weight of arguments whether they are backed by implementations or not. It just
> becomes easier to create a mess with things depending on each other created as
> separate, "freely" (userspace-decided)-named objects.

If I were to write a user-space tool to handle this, I would keep each chain
of checkpoints (from "base" and on) in a separate subdir, for example. In fact,
that's how I did it :)

>>> - You can easily create new branches by just copying the file, restarting from
>>>   it, and adding incremental checkpoints to it.  (Not sure this branch feature
>>>   is really interesting, but I it sounds funny :))
>> Using multiple files, you can create branches by adding hard-links (or soft-
>> links) to previous files. Saves space, time, and - I'd argue - easier to
>> understand and manage.
> 
> Again, no doubt about the feasibility with multiple files. I admit that this
> also saves space since the common parts are shared.
> 
>> Branches features is really interesting, as a matter of fact; Again I refer
>> you to the paper mentioned above.
>>
>>>> Recall that the data can be streamed, so when you start to read a file you
>>>> don't know a-priori how long is the checkpoint image, until you have parsed
>>>> it all; So you can't easily find the beginning of the, say 15th checkpoint
>>>> int that case.
>>> Good point: in append-only mode, we do not know that there are 15 checkpoints
>>> until we reach the 15th one. Perhaps append-only is too restrictive for
>>> incremental checkpoint. OTOH, do we really want to support a unique stream
>>> having multiple checkpoints? Probably not. So rewrite and append looks like a
>>> better option. An incremental checkpoint procedure could look like this:
>>>
>>> 	err = sys_checkpoint(base_fd, out_fd, ...)
>> Re-write + append will end up being very costly (imagine you save the data
>> on a network filel system), both in time and (at least for some time) in
>> space.
> 
> Hm, I'd bet that you have to read the previous checkpoints anyway, unless after
> some time things differ so much that the oldest images are not needed anymore.

Read, yes. Not re-write. And you don't need to read all of them, but cherry-pick
the pieces of interest (as indicated in the "current" checkpoint image).

> 
>> Besides, this scheme begins to sound much more complex than a single file.
>> Do you really gain so much from not having multiple files, one per checkpoint ?
> 
> Well, at least you are not limited by the number of open file descriptors
> (assuming that, as you mentioned earlier, you pass an array of previous images
> to compute the next incremental checkpoint).

You aren't limited by the number of open file. User space could provide an array
of <CRID, pathname> (or <serial#, pathname>) to the kernel, the kernel will
access the files as necessary.

Uhh .. hold on:  you need the array of previous checkpoint to _restart_ from
an incremental checkpoint. You don't care about it when you checkpoint: instead,
you keep track in memory of (1) what changed (e.g. which pages where touched),
and (2) where to find unmodified pages in previous checkpoints. You save this
information with each new checkpoint.  The data structure to describe #2 is
dynamic and changes with the execution, and easily keeps track of when older
checkpoint images become irrelevant (because all the pages they hold have been
overwritten already).


>>> where:
>>> - base_fd is a regular file containing the base checkpoint, or -1 if a full
>>>   checkpoint should be done. The checkpoint could actually also live in memory,
>>>   and the kernel should check that it matches the image pointed to by base_fd.
>>> - out_fd is whatever file/socket/etc. on which we should dump the checkpoint. In
>>>   particular, out_fd can equal base_fd and should point to the beginning of the
>>>   file if it's a regular file.
>> Excellent example. What if the checkpoint data is streamed over the network;
>> so you cannot rewrite the file after it has been streamed...  Or you will have
>> to save the entire incremental history in memory :(
> 
> I'm not sure to have expressed myself well: as was explained later, streaming
> output is ok for an incremental checkpoint, since you need the base checkpoint
> anyway. Unless you have a solution to build an incremental checkpoint out of
> streamed earlier checkpoints, I don't see what kind of limitation this would
> introduce.

I suspect we need to clarify the terminology: by "streamed" I mean that
the format does not require seeks (going back and forth), so that it can be
sent over a socket and make sense. While this is useful for migration, it
does not imply a migration. Consider, for instance, if you want to store the
checkpoint elsewhere you transfer the data via a socket to a daemon.

I actually wasn't thinking of streaming a series of incremental checkpoints
(from base and on) to implement migration... I simply didn't have a use-case
for that :)

>> The checkpoint - may, or may not live in memory for a long time. Usually not,
>> by the way, for the usual case it doesn't really make sense to use up memory
>> for nothing.
> 
> Definitely agreed.
> 
>>> If base_fd is a valid file descriptor, sys_checkpoint() would do this:
>>>
>>> #1 check the validity of the checkpoint image (possibly compare with in-memory
>>>     checkpoint states),
>>>
>>> #2 (over)write the position of the next (coming) checkpoint on out_fd (see
>>>    explanations below) and its sequence number as well (this actually makes
>>>    sequence counters live in the checkpoint image),
>>>
>>> #3 write the contents of base_fd to out_fd, marking the records invalidated by
>>>    the current checkpoint on the fly (see explanations below),
>>>
>>> #4 write the new incremental checkpoint records.
>> I truly don't think this scheme is simpler or easier to manage compared to
>> a using multiple files; and I really wonder what is the big advantage of
>> going through this non-trivial logic ?
>>
>>> This assumes that a checkpoint image has a place in the header to tell where the
>>> last checkpoint image is. Eventually, each record (task struct, vma, page, etc.)
>>> should contain a field telling which later incremental checkpoint invalidates
>>> it, so that we can restart from any intermediate checkpoint if we like.
>> My experience is that you really need incremental for memory, but not that
>> necessary for the rest of the state. So the way I did it is - whenever a
>> vma is saved, if some of its pages are found in previous checkpoints, a
>> pointer to where the page data resides is given (CRID, position) instead of
>> the page contents.
> 
> So in the case I described, say we restart from checkpoint #7, the page would be
> found at the first page record of same (mm,address) that is not invalidated by a
> checkpoint having id <= 7.

Ehhh... I'm confused with this. Invalidated by checkpoint having id <= 7 ?  only
a later checkpoint can invalidate a page and provide a newer version of that
page.

So to restart from checkpoint #7, you first restart from checkpoint #7 *as is*.
At this point you'll have everything setup, except that some memory contents
(hopefully much, because that means you saved a lot by doing incremental) will
be incorrect, because they weren't actually saved with checkpoint #7.  But
checkpoint #7 will also have a section that describes this remaining memory and
where it can be found, e.g many entries like this:

	<mm_struct id, page addr, checkpoint image id, position in file>

Now the code will scan this array, and fetch the required pages from where
they are stored.

(As mentioned before, the data structure that describes this array will be
dynamically updated as applications modify their memory).

This, of course, assumes that an incremental restart is _not_ stream-able,
and that all the files (or the entire single file) is available and seek-able.
(Still, being able to stream the (regular) checkpoint/restart operation is one
of our goals).

> I see where multiple files provide more performance however: you do not have to
> read the whole history to restart. At least this is true for non-streamed
> checkpoints. As soon as they are streamed, you can only hope that you won't need
> data living at the end of the images.

Exactly.

> 
>>> Moreover, each intermediate checkpoint would contain a pointer to the start of
>>> the previous and the next one, so that any intermediate checkpoint can be easily
>>> found. This actually makes step #2 and #3 modify the checkpoint image in place,
>>> whenever based_fd and out_fd point to the same file. This disables streaming for
>>> restarts from an intermediate checkpoint, but I don't think this is a real
>>> issue, unless there are use cases outside live-migration?
>> This is not quite possible to do when the data has been streamed through a
>> socket, for example (can't rewrite); or expensive to do with a network file
>> system.
> 
> Again, how do you build an incremental checkpoint out of streamed-only previous
> checkpoints?

I hope the clarification above explains that what I meant by "data being
steamed" is that the file is not seek-able.

> 
>> Live migration is orthogonal to incremental checkpoint, they have nothing
>> in common. There are use cases for restarting from an intermediate checkpoint
>> like the paper I mentioned, as well as fault tolerance, debugging, forensics,
>> and more.
> 
> I'm definitely sure that intermediate checkpoints are interesting. I was only
> wondering if streaming was so interesting for them.

Not in the sense of streaming for migration :)

> 
>> "Streaming" also means, as I mentioned above, to the case where you send
>> the data over a socket (even if not for a live migration, but to a daemon
>> that would hold it in memory on another node, for example). In that media
>> you cannot easily rewrite the file.
> 
> The point is that you need previous data when building an incremental
> checkpoint, so you will read it at least. And since it was previously stored (in

The scheme that I described above and is implemented in Zap does not require
access to previous checkpoints when building a new incremental checkpoint.
Instead, you keep some data structure in the kernel that describes the pieces
that you need to carry with you (what pages were saved, and where; when a task
exits, the data describing its mm will be discarded, of course, and so on).

> memory or whatever), you can even get its size before actually reading it,
> unless you checkpoint at such a rate that the previous chekpoint was not
> completely sent when you start the next one. If a remote daemon should really
> host the checkpoints, you can even tell the daemon which checkpoint to overwrite
> with the new one.
> 
>>>> Depending on the size of your checkpoint, a single file may eventually become
>>>> very large in a short time. I have one system that takes a checkpoint every
>>>> second of en entire user-desktop ...
>>>>
>>>> One single large file is harder to manager, parse, and inspect, even with
>>>> proper user tools. If you wanted to change something inside (for whatever
>>>> reasons), that would be a difficult to do. Same goes for when you want to
>>>> coalesce multiple checkpoints into a single checkpoint (e.g. to save space,
>>>> or because you don't care about some of your past)
>>> Ok, this becomes more complex, but feasible I think (see above).
>> Heh ... of course it is feasible. The question is which alternative is better ?
> 
> Definitely, and probably none of them alone ;)
> 
>>> Coalescing checkpoints seems rather easy as soon as checkpoints records are
>>> tagged with the first checkpoint number that invalidates them.
>>>
>>>> Ahh.. ok.. I stop here. This is not related to CRID vs. FD anymore :)
>>> You're right. Hopefully it is interesting, although a bit early to discuss :)
>> lol .. I couldn't help it.
> 
> I could also have simply shut up and kept on lurking... but it was so temptating
> to enter the discussion :)
> 
>>>>> could contain the entire checkpoint history. On the other hand, you are not sure
>>>>> that we could do incremental checkpoint this way, which justifies your need for
>>>>> a CRID. Perhaps you have an example?
>>>> Arguments given above. Note that even with multiple files we don't _need_
>>>> CRID, they are merely helpful. Instead, the user could be required to provide
>>>> the kernel with an array of file names, corresponding to checkpoint#0 (base),
>>>> checkpoint#2, checkpoint#3 etc; In this case, the "incremental state" that
>>>> is saved with checkpoint#4, is (a) that it is #4, and (b) for each part of
>>>> state that is found in a previous checkpoint, a reference to the serial no.
>>>> of that checkpoint is kept.
>>> See above for a solution based on a single file.
>>>
>>>> (The proposal for CRID was that instead of a serial number that starts from
>>>> 0 with every full (base) checkpoint, we use the CRID).
>>>>
>>>>> Anyway, do not take this as an attack. I just want to be well convinced that
>>>> On the contrary; your comments are definitely in place.
>>>>
>>>>> CRIDs are really needed, and are worth the effort of managing them cleanly.
>>>>> Exposing them to userspace just scares me a bit.
>>>> I'm not sure why is there an "effort of managing" them ?  It's a simple
>>>> atomic counter, that won't wrap around (use 64 bit if we wish). All in-memory
>>>> checkpoint contexts will be (also) in global linked list and easily located
>>>> there by their CRID.
>>> Ok, as long as no userspace task holds such IDs accross reboot or migration. How
>>> would you check this?
>> Ahhhh.... once again:  CRIDs do _not_ make sense across a reboot. Not in the
>> kernel anyway.  (For incremental, they can be used as hints, and userspace
>> brains are needed there anyway).  A CRID identifies a checkpoint _in memory_
>> and goes away when the checkpoint is removed from memory (canceled, commited)
>> or when the container goes away, or when the RAM goes away (e.g. reboot).
> 
> Again, I must have failed expressing myself well. I really understand that your
> CRIDs have no sense accross reboot or migration, and I do not want to give them
> such sense. What annoys me is that userspace gets a CRID as a result of
> sys_checkpoint(), and then can give it back to the kernel to write back the
> checkpoint. IIUC, a correct userspace checkpointer would give this CRID to
> the kernel to write the checkpoint (your step #5), and then would never give it
> again to the kernel (or only if the kernel would keep it internally for later
> incremental checkpoints). The problem is not well-behaving userspace apps. The
> problem is: how does the kernel check that userspace does not give a crappy
> CRID (actually the CRID of a checkpoint in an un-related container, it would
> probably not hurt for CRIDs generated on a another node/life mistakenly refering
> to locally computed checkpoints of the same container)?

Excellent point. (same as with IPC identifiers ...)

> Ok, the answer is probably here: CRIDs are local to containers, and userspace
> always gives them to the kernel with a reference on the container (whatever
> struct it is based on).

Exactly.

> 
>> When I said "hints" for user space, I refer to two use cases actually. One
>> is the incremental checkpoint where this CRID will be part of the header of
>> the checkpoint file, and user space will have that number returned by the
>> syscall and could use it (e.g. to name the files, but also to keep a record
>> of when/what was checkpointed).
>> Another is when we will add the capability of file-system snapshot, then
>> we'll have a way to identify each snapshot (let's say there will be some
>> identifier to each). Then user space could keep a table with the tuples:
>> <time, filename, CRID, FSID> to keep track of checkpoint data (FSID stands
>> for filesystem snapshot identifier).
> 
> Ok. I'd bet that userspace could figure out itself what is the sequence number
> of the next checkpoint, but why not.
> 
> It's probably time to conclude: I am now convinced that CRIDs can be managed
> correctly without userspace being able to crash everything. I'm not strongly
> against incremental checkpoints having their own files, so I won't debate to
> death on their advantages and drawbacks. You recognized (IIUC your words) the
> feasibility of single files hosting chains of incremental checkpoints, so I will
> consider myself satisfied with your proposal.

All agreed :)

> 
> Anyway, other proposals are coming (eg the one from openvz), and things may
> still move. So the discussion will probably come back in some way.
> 
> Thanks for the discussion (and I'm still interested in your answers to questions
> left above).

I tried my best.

Oren.

> 
> Louis
> 
>>>>> Btw, if we ever decide to use CRIDs, I'd propose to manage them in some
>>>>> pseudo-filesystem, like SYSV IPC objects actually are.
>>>> Eventually, yes ;)
>>>>
>>>>> Thanks,
>>>>>
>>>>> Louis
>>>>>
>>>> Thanks for the comments and stimulating the discussion.
>>> I should have had many more discussions like this during my PhD. Your's is going
>>> to be definitely better than mine :)
>> :)
>>
>> Oren.
>>
>>> Thanks,
>>>
>>> Louis
>>>
> 
_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers