On restore the "main" thread is started before the Resource's afterRestore has completed

Thu Apr 6 15:03:42 UTC 2023

Another aspect where OpenJ9 differs from CRaC is that the latter allows the
user to take a checkpoint at random point in time using jcmd.
I don't think OpenJ9 supports that. With the jcmd approach, it is almost
impossible for the user to envisage the coordination needed between
different entities in the system.
Compare this to the case where the user has to explicitly call a checkpoint
API in the code and knows to a large extent what the state of the system is
and what actions it can take to maintain the integrity of the system after
restore.
I think using checkpoint API makes it easier for the user by making the
problem of coordination more manageable.

- Ashutosh Mehra

On Thu, Apr 6, 2023 at 10:31 AM Dan Heidinga <heidinga at redhat.com> wrote:

>
>
> On Thu, Apr 6, 2023 at 9:59 AM Christian Tzolov <
> christian.tzolov at gmail.com> wrote:
>
>> Hi Dan and Radim,
>>
>>
>> Thanks for the feedback and suggestions!
>>
>> It is the first time I’m facing the java.lang.invoke.* API and it might
>> take some time to wrap my head around it.
>>
>> So be prepared plese for lame questions, as those inlined below.
>>
>>
>> On Wed, Apr 5, 2023 at 4:28 PM Dan Heidinga <heidinga at redhat.com> wrote:
>>
>>> Hi Radim,
>>>
>>> Thanks for the write up of the various options in this space.
>>>
>>> On Tue, Apr 4, 2023 at 2:49 AM Radim Vansa <rvansa at azul.com> wrote:
>>>
>>>> Hi Christian,
>>>>
>>>> I believe this is a common problem when porting existing architecture
>>>> under CRaC; the obvious solution is to guard access to the resource
>>>> (ProcessorContext in this case) with a RW lock that'd be read-acquired
>>>> by 'regular' access and acquired for write in beforeCheckpoint/released
>>>> in afterRestore. However this introduces extra synchronization (at
>>>> least
>>>> in form of volatile writes) even in case that C/R is not used at all,
>>>> especially if the support is added into libraries.
>>>>
>>>
>>> I've seen variations of this approach go by in code reviews but have we
>>> written up a good example of how to do this well?  Having a
>>> canonical pattern would help to highlight the best way to do it today and
>>> make the tradeoffs explicit.
>>>
>>
>> @Radim, your “guard access” suggestion made me realise that perhaps I’ve
>> oversimplified my  sample.
>>
>> So I’ve modified it a bit:
>> https://github.com/tzolov/crac-demo/blob/main/src/main/java/com/example/crac/CrackDemoExt.java
>> by introducing a new ProcessorState used by the Processor for its
>> computation.
>> At the same time I’ve removed the direct Processor dependency on the
>> ProcessorContext. Instead the ProcessorContext is responsible for managing
>> the lifecycle of the ProcessorState before the Processor can use it.
>> Then given your original suggestion is it right to assume that the “guard
>> access to the resource” now should guard the ProcessorState not the
>> ProcessorContext?
>>
>
> I think the example is still too simple as there is no state being
> protected.  Typically, the beforeCheckpoint/afterRestore methods are used
> to modify the state of a Class so that the class's invariants continue to
> hold across the restore.  This often (though not always) has to do with the
> external environment - if the application has captured a view of the
> environment (particular ports, # of cpus, env vars, etc) and made decisions
> based on that view, then after restore, that view needs to be updated.
>
> The lifecycle works by giving the application an opportunity prior to
> checkpoint to stop using the old state.  It also gives an application an
> opportunity to update that state after restore.  Those are the
> {beforeChecpoint, afterRestore} apis on Resource.  This produces an
> indeterminate length of time from the start of the checkpoint (and first
> call to beforeCheckpoint) thru to the completion of the last afterRestore
> call.  During this period, threads may see the original value, the
> beforeCheckpoint updated value, the afterRestore updated value, or some
> combination of all three depending on timing and thread scheduling.
>
>
>> And if this is true then how one would be able to identify all possible
>> “resources” to be guarded?
>>
>
> That's the million dollar question.  The answer so far has been code
> inspection or trial-and-error.  And the answer of which "resources" depends
> a bit on the use case - the set of resources for a desktop application that
> will be checkpointed/restored on the same machine may be very different
> than a server application that will be spread across a K8 cluster or a
> different set of Lambda endpoints.
>
>
>>
>>
>>
>>> Anton Kozlov proposed techniques like RCU [1] but at this point there's
>>>> no support for this in Java. Even the Linux implementation might
>>>> require
>>>> some additional properties from the code in critical (read) section
>>>> like
>>>> not calling any blocking code; this might be too limiting.
>>>>
>>>> The situation is simpler if the application uses a single threaded
>>>> event-loop; beforeCheckpoint can enqueue a task that would, upon its
>>>> execution, block on a primitive and notify the C/R notification thread
>>>> that it may now deinit the resource; in afterRestore the resource is
>>>> initialized and the eventloop is unblocked. This way we don't impose
>>>> any
>>>> extra overhead when C/R is happening.
>>>>
>>>
>>> That's a nice idea!
>>>
>>>
>>>>
>>>> To avoid extra synchronization it could be technically possible to
>>>> modify CRaC implementation to keep all other threads frozen during
>>>> restore. There's a risk of some form of deadlock if the thread
>>>> performing C/R would require other threads to progress, though, so any
>>>> such solution would require extra thoughts. Besides, this does not
>>>> guarantee exclusivity so the afterRestore would need to restore the
>>>> resource to the *exactly* same state (as some of its before-checkpoint
>>>> state might have leaked to the thread in Processor). In my opinion this
>>>> is not the best way.
>>>>
>>>
>>> This is the approach that OpenJ9 took to solve the consistency problems
>>> introduced by updating resources before / after checkpoints.  OpenJ9 enters
>>> "single threaded mode" when creating the checkpoint and executing the
>>> before checkkpoint fixups.  On restore, it continues in single-threaded
>>> mode while executing the after checkpoint fixups.  This makes it easier to
>>> avoid additional runtime costs related to per-resource locking for
>>> checkpoints, but complicates locking and wait/notify in general.
>>>
>>> This means a checkpoint hook operation can't wait on another thread
>>> (would block indefinitely as other threads are paused), can't wait on a
>>> lock being held by another thread (again, would deadlock), and sending
>>> notify may result in inconsistent behaviour (wrong number of notifies
>>> received by other threads).  See "The checkpointJVM() API" section of their
>>> blog post on CRIU for more details [0].
>>>
>>
>> The "single thread mode", imo, corresponds to the
>> "serializable isolation" approach in data processing and DB transactions.
>> The OpenJ9 blogs are very informative and like the jdk invoke API would
>> need time to digest.
>> But I have one conceptual question. What part of this should/cloud be
>> implemented by the CRaC inself and what abstractions should be exposed to
>> the CRaC users?
>>
>
> If CRaC were to adopt the single-threaded mode, then almost all of the
> work for that would be in the CRaC project (ie: Hotspot) itself.  Users
> would only need to be sure their before/after checkpoint methods were
> "safe" to run.
>
>
>>
>>
>>>> The problem with RCU is tracking which threads are in the critical
>>>> section. I've found RCU-like implementations for Java that avoid
>>>> excessive overhead using a spread out array - each thread marks
>>>> entering/leaving the critical section by writes to its own counter,
>>>> preventing cache ping-pong (assuming no false sharing). Synchronizer
>>>> thread uses another flag to request synchronization; reading this by
>>>> each thread is not totally without cost but reasonably cheap, and in
>>>> that case worker threads can enter a blocking slow path. The simple
>>>> implementation assumes a fixed number of threads; if the list of
>>>> threads
>>>> is dynamic the solution would be probably more complicated. It might
>>>> also make sense to implement this in native code with a per-CPU
>>>> counters, rather than per-thread. A downside, besides some overhead in
>>>> terms of both cycles and memory usage, is that we'd need to modify the
>>>> code and explicitly mark the critical sections.
>>>>
>>>> Another solution could try to leverage existing JVM mechanics for code
>>>> deoptimization, replacing the critical sections with a slower, blocking
>>>> stub, and reverting back after restore. Or even independently
>>>> requesting
>>>> a safe-point and inspecting stack of threads until the synchronization
>>>> is possible.
>>>>
>>>
>>> This will have a high risk of livelock.  The OpenJ9 experience
>>> implementing single-threaded mode for CRIU indicates there are a lot of
>>> strange locking patterns in the world.
>>>
>>>
>>>>
>>>> So I probably can't offer a ready-to-use performant solution; pick your
>>>> poison. The future, though, offers a few possibilities and I'd love to
>>>> hear others' opinions about which one would look the most feasible.
>>>> Because unless we offer something that does not harm a no-CRaC use-case
>>>> I am afraid that the adoption will be quite limited.
>>>>
>>>
>>> Successful solutions will push the costs into the checkpoint / restore
>>> paths as much as possible.  Going back to the explicit lock mechanism you
>>> first mentioned, I wonder if there's a role for
>>> java.lang.invoke.Switchpoint [1] here?  Switchpoint was added as a tool for
>>> language implementers that wanted to be able speculate on a particular
>>> condition (ie: CHA assumptions) and get the same kind of low cost state
>>> change that existing JITTED code gets.  I'm not sure how well that vision
>>> worked in practice or how well Hotspot optimizes it yet, but this might be
>>> a reason to push on its performance.
>>>
>>> Roughly the idea would be to add a couple of Switchpoints to
>>> jdk.crac.Core:
>>>
>>>    public SwitchPoint getBeforeSwitchpoint();
>>>    public SwitchPoint getAfterSwitchpoint();
>>>
>>> and users could then write their code using MethodHandles to
>>> implementing the branching logic:
>>>
>>>     MethodHandle normalPath = ...... // existing code
>>>     MethodHandle fallbackPath = ..... // before Checkpoint extra work
>>>     MethodHandle guardWithTest =
>>> getBeforeSwitchPoint.guardWithTest(normalPath, fallbackPath);
>>>
>>> and the jdk.crac.Core class would invalidate the "before" SwitchPoint
>>> prior to the checkpoint and "after" one after the restore.  Aside from the
>>> painful programming model, this might give us the tools we need to make it
>>> performant.
>>>
>>
>> @Dan, this is very interesting!
>> Could you please elaborate a bit further. Perhaps in the context of the
>> CrackDemoExt.java sample?
>>
>
> Let me think on that.  I'll see if I can pull something together that
> shows the api use.
>
> --Dan
>
>
>>
>>
>>>
>>> Needs more exploration and prototyping but would provide a potential
>>> path to reasonable performance by burying the extra locking in the fallback
>>> paths.  And it would be a single pattern to optimize, rather than all the
>>> variations users could produce.
>>> --Dan
>>> [0]
>>> https://blog.openj9.org/2022/10/14/openj9-criu-support-a-look-under-the-hood/
>>> [1]
>>> https://docs.oracle.com/en/java/javase/17/docs/api/java.base/java/lang/invoke/SwitchPoint.html
>>>
>>
>> Thank you,
>>  - Christian
>>
>>
>>
>>>
>>>> Cheers,
>>>>
>>>> Radim
>>>>
>>>> [1] https://en.wikipedia.org/wiki/Read-copy-update
>>>>
>>>> On 03. 04. 23 22:30, Christian Tzolov wrote:
>>>> > Hi, I'm testing CRaC in the context of long-running applications
>>>> (e.g. streaming, continuous processing ...) and I've stumbled on an issue
>>>> related to the coordination of the resolved threads.
>>>> >
>>>> > For example, let's have a Processor that performs continuous
>>>> computations. This processor depends on a ProcessorContext and later must
>>>> be fully initialized before the processor can process any data.
>>>> >
>>>> > When the application is first started (e.g. not from checkpoints) it
>>>> ensures that the ProcessorContext is initialized before starting the
>>>> Processor loop.
>>>> >
>>>> > To leverage CRaC I've implemented a ProcessorContextResource
>>>> gracefully stops the context on beforeCheckpoint and then re-initialized it
>>>> on afterRestore.
>>>> >
>>>> > When the checkpoint is performed, CRaC calls the
>>>> ProcessorContextResource.beforeCheckpoint and also preserves the current
>>>> Processor call stack. On Restore processor's call stack is expectedly
>>>> restored at the point it was stopped but unfortunately it doesn't wait for
>>>> the ProcessorContextResource.afterRestore complete. This expectedly crashes
>>>> the processor.
>>>> >
>>>> > The https://github.com/tzolov/crac-demo illustreates this issue. The
>>>> README explains how to reproduce the issue. The OUTPUT.md (
>>>> https://github.com/tzolov/crac-demo/blob/main/OUTPUT.md ) offers
>>>> terminal snapshots of the observed behavior.
>>>> >
>>>> > I've used latest JDK CRaC release:
>>>> >    openjdk 17-crac 2021-09-14
>>>> >    OpenJDK Runtime Environment (build 17-crac+5-19)
>>>> >    OpenJDK 64-Bit Server VM (build 17-crac+5-19, mixed mode, sharing)
>>>> >
>>>> > As I'm new to CRaC, I'd appreciate your thoughts on this issue.
>>>> >
>>>> > Cheers,
>>>> > Christian
>>>> >
>>>> >
>>>> >
>>>> >
>>>>
>>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/crac-dev/attachments/20230406/64559b9b/attachment-0001.htm>