CUDA 10 added runtime API calls for putting streams (= queues) in "capture mode", so that instead of executing, they are returned in a "graph". These graphs can then be made to actually execute, or they can be cloned.
But what is the rationale behind this feature? Isn't it unlikely to execute the same "graph" twice? After all, even if you do run the "same code", at least the data is different, i.e. the parameters the kernels take likely change. Or - am I missing something?
PS - I skimmed this slide deck, but still didn't get it.
My experience with graphs is indeed that they are not so mutable. You can change the parameters with 'cudaGraphHostNodeSetParams', but in order for the change of parameters to take effect, I had to rebuild the graph executable with 'cudaGraphInstantiate'. This call takes so long that any gain of using graphs is lost (in my case). Setting the parameters only worked for me when I build the graph manually. When getting the graph through stream capture, I was not able to set the parameters of the nodes as you do not have the node pointers. You would think the call 'cudaGraphGetNodes' on a stream captured graph would return you the nodes. But the node pointer returned was NULL for me even though the 'numNodes' variable had the correct number. The documentation explicitly mentions this as a possibility but fails to explain why.
Task graphs are quite mutable.
There are API calls for changing/setting the parameters of task graph nodes of various kinds, so one can use a task graph as a template, so that instead of enqueueing the individual nodes before every execution, one changes the parameters of every node before every execution (and perhaps not all nodes actually need their parameters changed).
For example, See the documentation for cudaGraphHostNodeGetParams and cudaGraphHostNodeSetParams.
Another useful feature is the concurrent kernel executions. Under manual mode, one can add nodes in the graph with dependencies. It will explore the concurrency automatically using multiple streams. The feature itself is not new but make it automatic becomes useful for certain applications.
When training a deep learning model it happens often to re-run the same set of kernels in the same order but with updated data. Also, I would expect Cuda to do optimizations by knowing statically what will be the next kernels. We can imagine that Cuda can fetch more instructions or adapt its scheduling strategy when knowing the whole graph.
CUDA Graphs is trying to solve the problem that in the presence of too many small kernel invocations, you see quite some time spent on the CPU dispatching work for the GPU (overhead).
It allows you to trade resources (time, memory, etc.) to construct a graph of kernels that you can use a single invocation from the CPU instead of doing multiple invocations. If you don't have enough invocations, or your algorithm is different each time, then it won't worth it to build a graph.
This works really well for anything iterative that uses the same computation underneath (e.g., algorithms that need to converge to something) and it's pretty prominent in a lot of applications that are great for GPUs (e.g., think of the Jacobi method).
You are not going to see great results if you have an algorithm that you invoke once or if your kernels are big; in that case the CPU invocation overhead is not your bottleneck. A succinct explanation of when you need it exists in the Getting Started with CUDA Graphs.
Where task graph based paradigms shine though is when you define your program as tasks with dependencies between them. You give a lot of flexibility to the driver / scheduler / hardware to do scheduling itself without much fine-tuning from the developer's part. There's a reason why we have been spending years exploring the ideas of dataflow programming in HPC.
TL;DR version: "What's the best way to round-robin kernel calls to multiple GPUs with Python/PyCUDA such that CPU and GPU work can happen in parallel?" with a side of "I can't have been the first person to ask this; anything I should read up on?"
Full version:
I would like to know the best way to design context, etc. handling in an application that uses CUDA on a system with multiple GPUs. I've been trying to find literature that talks about guidelines for when context reuse vs. recreation is appropriate, but so far haven't found anything that outlines best practices, rules of thumb, etc.
The general overview of what we're needing to do is:
Requests come in to a central process.
That process forks to handle a single request.
Data is loaded from the DB (relatively expensive).
The the following is repeated an arbitrary number of times based on the request (dozens):
A few quick kernel calls to compute data that is needed for later kernels.
One slow kernel call (10 sec).
Finally:
Results from the kernel calls are collected and processed on the CPU, then stored.
At the moment, each kernel call creates and then destroys a context, which seems wasteful. Setup is taking about 0.1 sec per context and kernel load, and while that's not huge, it is precluding us from moving other quicker tasks to the GPU.
I am trying to figure out the best way to manage contexts, etc. so that we can use the machine efficiently. I think that in the single-gpu case, it's relatively simple:
Create a context before starting any of the GPU work.
Launch the kernels for the first set of data.
Record an event for after the final kernel call in the series.
Prepare the second set of data on the CPU while the first is computing on the GPU.
Launch the second set, repeat.
Insure that each event gets synchronized before collecting the results and storing them.
That seems like it should do the trick, assuming proper use of overlapped memory copies.
However, I'm unsure what I should do when wanting to round-robin each of the dozens of items to process over multiple GPUs.
The host program is Python 2.7, using PyCUDA to access the GPU. Currently it's not multi-threaded, and while I'd rather keep it that way ("now you have two problems" etc.), if the answer means threads, it means threads. Similarly, it would be nice to just be able to call event.synchronize() in the main thread when it's time to block on data, but for our needs efficient use of the hardware is more important. Since we'll potentially be servicing multiple requests at a time, letting other processes use the GPU when this process isn't using it is important.
I don't think that we have any explicit reason to use Exclusive compute modes (ie. we're not filling up the memory of the card with one work item), so I don't think that solutions that involve long-standing contexts are off the table.
Note that answers in the form of links to other content that covers my questions are completely acceptable (encouraged, even), provided they go into enough detail about the why, not just the API. Thanks for reading!
Caveat: I'm not a PyCUDA user (yet).
With CUDA 4.0+ you don't even need an explicit context per GPU. You can just call cudaSetDevice (or the PyCUDA equivalent) before doing per-device stuff (cudaMalloc, cudaMemcpy, launch kernels, etc.).
If you need to synchronize between GPUs, you will need to potentially create streams and/or events and use cudaEventSynchronize (or the PyCUDA equivalent). You can even have one stream wait on an event inserted in another stream to do sophisticated dependencies.
So I suspect the answer to day is quite a lot simpler than talonmies' excellent pre-CUDA-4.0 answer.
You might also find this answer useful.
(Re)Edit by OP: Per my understanding, PyCUDA supports versions of CUDA prior to 4.0, and so still uses the old API/semantics (the driver API?), so talonmies' answer is still relevant.
I was thinking about garbage collection on the way home, and I began wondering, why does the garbage collector totally freeze execution of a program? Personally I would have designed it to block any threads which try to allocate a new object, but threads which were running would be left alone.
I can't imagine any situation where this would be a problem compared to how a garbage collector currently works.
I was thinking about garbage collection on the way home, and I began wondering, why does the garbage collector totally freeze execution of a program?
There is a trade-off between latency and throughput in GC design. You can either process heap-allocated blocks individually ("incremental") or you can batch them up and process them all at the same time ("stop the world"). Fully incremental collection never totally freezes a program and it has very low latency but it also has very poor throughput. Stop the world garbage collectors have the worst possible latency (freezing the program for seconds or even minutes at a time) but near-optimal throughput.
All of the major production GCs today provide a middle ground, typically with generational collection with the per-thread nursery generations collected in batches and incremental or concurrent collection of the shared old generation. Thus, only nursery collections incur pauses and nursery size is bounded so pause times are kept low, e.g. 10-100ms in .NET with the workstation GC.
For a simple GC algorithm that never pauses, see Baker's Treadmill. For more information on garbage collection I highly recommend the Memory Management Reference and the Garbage Collection Handbook.
There is a lot of misinformation in the other answers here. Jon Skeet wrote some source code and started discussing it from the point of view of garbage collection. You need to be very careful doing this because there is little correspondence between source code and what the GC sees. The compiler does instruction block rearrangements, register allocation, promotion and so on, all of which affect what is visible to the GC at run time. In particular, scope in source code is not carried through to compiled code and is typically replaced with the related concept of liveness. Jon also wrote that you must pause in order to get the global roots. That is not strictly true although it is the most efficient way to get the global roots and the resulting pause is almost always tiny (sub-millisecond) because you're just copying less than a kB of stack from each thread.
Powerlord wrote that moving collectors must block reads and, therefore, all threads that read. This is also not true. The simplest counter example is immutable data: referential transparency means you can read from any copy safely.
Kico wrote that pauses are required to determine reachability. This is also not true. See Dijkstra's research about "on-the-fly" collectors and any recent real-time GC such as Stacatto.
Jerry Coffin wrote the best answer but moving isn't the reason GCs pause. There are GCs that don't move but do pause (e.g. HLVM's) and those that do move but don't pause (e.g. Stacatto).
Modern garbage collectors (in .NET and Java, anyway) don't actually "stop the world" - they do all kinds of clever things to collect concurrently.
However, you might want to consider a situation like this:
object x = null;
object y = new object();
...
x = y;
y = null;
Now, suppose the GC looks at x, then the lines below the ... run, and then the GC looks at y - it won't have seen any live objects... but the object should still be live.
Basically there needs to be a certain amount of pausing in order to get a consistent set of references. Then there's compaction, reference reassignment etc. However, it isn't nearly as bad as it used to be in terms of requiring everything to be stopped for the whole of the GC cycle. It does, however, get painful to think about :)
In addition to what Kico Lobo said, Garbage Collectors can also move things around in memory.
Therefore, they don't just have to block threads that write to memory, but also threads that read from memory.
Which is every thread.
Most GCs stop execution because objects can move in memory during a collection cycle (at least with most reasonably recent designs). That means either reading or writing almost any object at the wrong time can cause a problem.
There are collectors that have been designed around the idea of just blocking reads (or writes) to the specific parts of memory being modified at a given time, so as long as execution only uses objects that aren't (currently) being moved around, it can proceed unhindered. The problem is that most typical hardware doesn't provide efficient support for this, so even though they work in principle, they're fairly inefficient in practice. There has been at least one attempt at adapting that type of algorithm to use the write protection available in a typical paging unit, but I'm not aware of its having been used for much other than research and experimentation.
The primary alternative is to make the collector incremental -- i.e. have it do only a small amount of work at a time, so even though other execution gets stopped, it only has to stop for a little while at any given time.
With multi-core machines becoming so common, however, I'd expect to see more work put into garbage collection algorithms that can run in parallel with other execution. Up until recently, the primary emphasis was on minimizing the total time/effort spent on garbage collection. The growing number of cores available is likely to (often) mean that doing more total work in garbage collection may be easily justified, if doing so allows the mainstream of the code to run with fewer hindrances.
Edit: You might want to read Paul Wilson's Survey of Uniprocessor Garbage Collection Techniques. This isn't definitive (especially any more, given its age), but it's at least a reasonable starting point.
Because that's the only way it can assure that the refereces it is going to clean are not been used by anyone else.
If it didn´t freezed the execution, it could not assure that.
Lets say you want to write a high performance method which processes a large data set.
Why shouldn't developers have the ability to turn on manual memory management instead of being forced to move to C or C++?
void Process()
{
unmanaged
{
Byte[] buffer;
while (true)
{
buffer = new Byte[1024000000];
// process
delete buffer;
}
}
}
Because allowing you to manually delete a memory block while there may still be references to it (and the runtime has no way of knowing that without doing a GC cycle) can produce dangling pointers, and thus break memory safety. GC languages are generally memory-safe by design.
That said, in C#, in particular, you can do what you want already:
void Process()
{
unsafe
{
byte* buffer;
while (true)
{
buffer = Marshal.AllocHGlobal(1024000000);
// process
Marshal.FreeHGlobal(buffer);
}
}
}
Note that, as in C/C++, you have full pointer arithmetic for raw pointer types in C# - so buffer[i] or buffer+i are valid expressions.
If you need high performance and detailed control, maybe you should write what you're doing in C or C++. Not all languages are good for all things.
Edited to add: A single language is not going to be good for all things. If you add up all the useful features in all the good programming languages, you're going to get a really big mess, far worse than C++, even if you can avoid inconsistency.
Features aren't free. If a language has a feature, people are likely to use it. You won't be able to learn C# well enough without learning the new C# manual memory management routines. Compiler teams are going to implement it, at the cost of other compiler features that are useful. The language is likely to become difficult to parse like C or C++, and that leads to slow compilation. (As a C++ guy, I'm always amazed when I compile one of our C# projects. Compilation seems almost instantaneous.)
Features conflict with each other, sometimes in unexpected ways. C90 can't do as well as Fortran at matrix calculations, since the possibility that C pointers are aliased prevents some optimizations. If you allow pointer arithmetic in a language, you have to accept its consequences.
You're suggesting a C# extension to allow manual memory management, and in a few cases that would be useful. That would mean that memory would have to be allocated in separate ways, and there would have to be a way to tell manually managed memory from automatically managed memory. Suddenly, you've complicated memory management, there's more chance for a programmer to screw up, and the memory manager itself is more complicated. You're gaining a little performance that matters in a few cases in exchange for more complication and slower memory management in all cases.
It may be that at some time we'll have a programming language that's good for almost all purposes, from scripting to number crunching, but there's nothing popular that's anywhere near that. In the meantime, we have to be willing to accept the limitations of using only one language, or the challenge of learning several and switching between them.
In the example you posted, why not just erase the buffer and re-use it?
The .NET garbage collector is very very good at working out which objects aren't referenced anymore and freeing the associated memory in a timely manner. In fact, the garbage collector has a special heap (the large object heap) in which it puts large objects like this, that is optimized to deal with them.
On top of this, not allowing references to be explicitly freed simply removes a whole host of bugs with memory leaks and dangling pointers, that leads to much safer code.
Freeing each unused block individually as done in a language with explicit memory management can be more expensive than letting a Garbage Collector do it, because the GC has the possibility to use copying schemes to spend a time linear to the number of blocks left alive (or the number of recent blocks left alive) instead of having to handle each dead block.
The same reason most kernels won't let you schedule your own threads. Because 99.99+% of time you don't really need to, and exposing that functionality the rest of the time will only tempt you to do something potentially stupid/dangerous.
If you really need fine grain memory control, write that section of code in something else.
I've been programming a long time, and the programs I see, when they run out of memory, attempt to clean up and exit, i.e. fail gracefully. I can't remember the last time I saw one actually attempt to recover and continue operating normally.
So much processing relies on being able to successfully allocate memory, especially in garbage collected languages, it seems that out of memory errors should be classified as non-recoverable. (Non-recoverable errors include things like stack overflows.)
What is the compelling argument for making it a recoverable error?
It really depends on what you're building.
It's not entirely unreasonable for a webserver to fail one request/response pair but then keep on going for further requests. You'd have to be sure that the single failure didn't have detrimental effects on the global state, however - that would be the tricky bit. Given that a failure causes an exception in most managed environments (e.g. .NET and Java) I suspect that if the exception is handled in "user code" it would be recoverable for future requests - e.g. if one request tried to allocate 10GB of memory and failed, that shouldn't harm the rest of the system. If the system runs out of memory while trying to hand off the request to the user code, however - that kind of thing could be nastier.
In a library, you want to efficiently copy a file. When you do that, you'll usually find that copying using a small number of big chunks is much more effective than copying a lot of smaller ones (say, it's faster to copy a 15MB file by copying 15 1MB chunks than copying 15'000 1K chunks).
But the code works with any chunk size. So while it may be faster with 1MB chunks, if you design for a system where a lot of files are copied, it may be wise to catch OutOfMemoryError and reduce the chunk size until you succeed.
Another place is a cache for Object stored in a database. You want to keep as many objects in the cache as possible but you don't want to interfere with the rest of the application. Since these objects can be recreated, it's a smart way to conserve memory to attach the cache to an out of memory handler to drop entries until the rest of the app has enough room to breathe, again.
Lastly, for image manipulation, you want to load as much of the image into memory as possible. Again, an OOM-handler allows you to implement that without knowing in advance how much memory the user or OS will grant your code.
[EDIT] Note that I work under the assumption here that you've given the application a fixed amount of memory and this amount is smaller than the total available memory excluding swap space. If you can allocate so much memory that part of it has to be swapped out, several of my comments don't make sense anymore.
Users of MATLAB run out of memory all the time when performing arithmetic with large arrays. For example if variable x fits in memory and they run "x+1" then MATLAB allocates space for the result and then fills it. If the allocation fails MATLAB errors and the user can try something else. It would be a disaster if MATLAB exited whenever this use case came up.
OOM should be recoverable because shutdown isn't the only strategy to recovering from OOM.
There is actually a pretty standard solution to the OOM problem at the application level.
As part of you application design determine a safe minimum amount of memory required to recover from an out of memory condition. (Eg. the memory required to auto save documents, bring up warning dialogs, log shutdown data).
At the start of your application or at the start of a critical block, pre-allocate that amount of memory. If you detect an out of memory condition release your guard memory and perform recovery. The strategy can still fail but on the whole gives great bang for the buck.
Note that the application need not shut down. It can display a modal dialog until the OOM condition has been resolved.
I'm not 100% certain but I'm pretty sure 'Code Complete' (required reading for any respectable software engineer) covers this.
P.S. You can extend your application framework to help with this strategy but please don't implement such a policy in a library (good libraries do not make global decisions without an applications consent)
I think that like many things, it's a cost/benefit analysis. You can program in attempted recovery from a malloc() failure - although it may be difficult (your handler had better not fall foul of the same memory shortage it's meant to deal with).
You've already noted that the commonest case is to clean up and fail gracefully. In that case it's been decided that the cost of aborting gracefully is lower than the combination of development cost and performance cost in recovering.
I'm sure you can think of your own examples of situations where terminating the program is a very expensive option (life support machine, spaceship control, long-running and time-critical financial calculation etc.) - although the first line of defence is of course to ensure that the program has predictable memory usage and that the environment can supply that.
I'm working on a system that allocates memory for IO cache to increase performance. Then, on detecting OOM, it takes some of it back, so that the business logic could proceed, even if that means less IO cache and slightly lower write performance.
I also worked with an embedded Java applications that attempted to manage OOM by forcing garbage collection, optionally releasing some of non-critical objects, like pre-fetched or cached data.
The main problems with OOM handling are:
1) being able to re-try in the place where it happened or being able to roll back and re-try from a higher point. Most contemporary programs rely too much on the language to throw and don't really manage where they end up and how to re-try the operation. Usually the context of the operation will be lost, if it wasn't designed to be preserved
2) being able to actually release some memory. This means a kind of resource manager that knows what objects are critical and what are not, and the system be able to re-request the released objects when and if they later become critical
Another important issue is to be able to roll back without triggering yet another OOM situation. This is something that is hard to control in higher level languages.
Also, the underlying OS must behave predictably with regard to OOM. Linux, for example, will not, if memory overcommit is enabled. Many swap-enabled systems will die sooner than reporting the OOM to the offending application.
And, there's the case when it is not your process that created the situation, so releasing memory does not help if the offending process continues to leak.
Because of all this, it's often the big and embedded systems that employ this techniques, for they have the control over OS and memory to enable them, and the discipline/motivation to implement them.
It is recoverable only if you catch it and handle it correctly.
In same cases, for example, a request tried to allocate a lot memory. It is quite predictable and you can handle it very very well.
However, in many cases in multi-thread application, OOE may also happen on background thread (including created by system/3rd-party library).
It is almost imposable to predict and you may unable to recover the state of all your threads.
No.
An out of memory error from the GC is should not generally be recoverable inside of the current thread. (Recoverable thread (user or kernel) creation and termination should be supported though)
Regarding the counter examples: I'm currently working on a D programming language project which uses NVIDIA's CUDA platform for GPU computing. Instead of manually managing GPU memory, I've created proxy objects to leverage the D's GC. So when the GPU returns an out of memory error, I run a full collect and only raise an exception if it fails a second time. But, this isn't really an example of out of memory recovery, it's more one of GC integration. The other examples of recovery (caches, free-lists, stacks/hashes without auto-shrinking, etc) are all structures that have their own methods of collecting/compacting memory which are separate from the GC and tend not to be local to the allocating function.
So people might implement something like the following:
T new2(T)( lazy T old_new ) {
T obj;
try{
obj = old_new;
}catch(OutOfMemoryException oome) {
foreach(compact; Global_List_Of_Delegates_From_Compatible_Objects)
compact();
obj = old_new;
}
return obj;
}
Which is a decent argument for adding support for registering/unregistering self-collecting/compacting objects to garbage collectors in general.
In the general case, it's not recoverable.
However, if your system includes some form of dynamic caching, an out-of-memory handler can often dump the oldest elements in the cache (or even the whole cache).
Of course, you have to make sure that the "dumping" process requires no new memory allocations :) Also, it can be tricky to recover the specific allocation that failed, unless you're able to plug your cache dumping code directly at the allocator level, so that the failure isn't propagated up to the caller.
It depends on what you mean by running out of memory.
When malloc() fails on most systems, it's because you've run out of address-space.
If most of that memory is taken by cacheing, or by mmap'd regions, you might be able to reclaim some of it by freeing your cache or unmmaping. However this really requires that you know what you're using that memory for- and as you've noticed either most programs don't, or it doesn't make a difference.
If you used setrlimit() on yourself (to protect against unforseen attacks, perhaps, or maybe root did it to you), you can relax the limit in your error handler. I do this very frequently- after prompting the user if possible, and logging the event.
On the other hand, catching stack overflow is a bit more difficult, and isn't portable. I wrote a posixish solution for ECL, and described a Windows implementation, if you're going this route. It was checked into ECL a few months ago, but I can dig up the original patches if you're interested.
Especially in garbage collected environments, it's quote likely that if you catch the OutOfMemory error at a high level of the application, lots of stuff has gone out of scope and can be reclaimed to give you back memory.
In the case of single excessive allocations, the app may be able to continue working flawlessly. Of course, if you have a gradual memory leak, you'll just run into the problem again (more likely sooner than later), but it's still a good idea to give the app a chance to go down gracefully, save unsaved changes in the case of a GUI app, etc.
Yes, OOM is recoverable. As an extreme example, the Unix and Windows operating systems recover quite nicely from OOM conditions, most of the time. The applications fail, but the OS survives (assuming there is enough memory for the OS to properly start up in the first place).
I only cite this example to show that it can be done.
The problem of dealing with OOM is really dependent on your program and environment.
For example, in many cases the place where the OOM happens most likely is NOT the best place to actually recover from an OOM state.
Now, a custom allocator could possibly work as a central point within the code that can handle an OOM. The Java allocator will perform a full GC before is actually throws a OOM exception.
The more "application aware" that your allocator is, the better suited it would be as a central handler and recovery agent for OOM. Using Java again, it's allocator isn't particularly application aware.
This is where something like Java is readily frustrating. You can't override the allocator. So, while you could trap OOM exceptions in your own code, there's nothing saying that some library you're using is properly trapping, or even properly THROWING an OOM exception. It's trivial to create a class that is forever ruined by a OOM exception, as some object gets set to null and "that never happen", and it's never recoverable.
So, yes, OOM is recoverable, but it can be VERY hard, particularly in modern environments like Java and it's plethora of 3rd party libraries of various quality.
The question is tagged "language-agnostic", but it's difficult to answer without considering the language and/or the underlying system. (I see several toher hadns
If memory allocation is implicit, with no mechanism to detect whether a given allocation succeeded or not, then recovering from an out-of-memory condition may be difficult or impossible.
For example, if you call a function that attempts to allocate a huge array, most languages just don't define the behavior if the array can't be allocated. (In Ada this raises a Storage_Error exception, at least in principle, and it should be possible to handle that.)
On the other hand, if you have a mechanism that attempts to allocate memory and is able to report a failure to do so (like C's malloc() or C++'s new), then yes, it's certainly possible to recover from that failure. In at least the cases of malloc() and new, a failed allocation doesn't do anything other than report failure (it doesn't corrupt any internal data structures, for example).
Whether it makes sense to try to recover depends on the application. If the application just can't succeed after an allocation failure, then it should do whatever cleanup it can and terminate. But if the allocation failure merely means that one particular task cannot be performed, or if the task can still be performed more slowly with less memory, then it makes sense to continue operating.
A concrete example: Suppose I'm using a text editor. If I try to perform some operation within the editor that requires a lot of memory, and that operation can't be performed, I want the editor to tell me it can't do what I asked and let me keep editing. Terminating without saving my work would be an unacceptable response. Saving my work and terminating would be better, but is still unnecessarily user-hostile.
This is a difficult question. On first sight it seems having no more memory means "out of luck" but, you must also see that one can get rid of many memory related stuff if one really insist. Let's just take the in other ways broken function strtok which on one hand has no problems with memory stuff. Then take as counterpart g_string_split from the Glib library, which heavily depends on allocation of memory as nearly everything in glib or GObject based programs. One can definitly say in more dynamic languages memory allocation is much more used as in more inflexible languages, especially C. But let us see the alternatives. If you just end the program if you run out of memory, even careful developed code may stop working. But if you have a recoverable error, you can do something about it. So the argument, making it recoverable means that one can choose to "handle" that situation differently (e.g putting aside a memory block for emergencies, or degradation to a less memory extensive program).
So the most compelling reason is. If you provide a way of recovering one can try the recoverying, if you do not have the choice all depends on always getting enough memory...
Regards
It's just puzzling me now.
At work, we have a bundle of applications working together, and memory is running low. While the problem is either make the application bundle go 64-bit (and so, be able to work beyond the 2 Go limits we have on a normal Win32 OS), and/or reduce our use of memory, this problem of "How to recover from a OOM" won't quit my head.
Of course, I have no solution, but still play at searching for one for C++ (because of RAII and exceptions, mainly).
Perhaps a process supposed to recover gracefully should break down its processing in atomic/rollback-able tasks (i.e. using only functions/methods giving strong/nothrow exception guarantee), with a "buffer/pool of memory" reserved for recovering purposes.
Should one of the task fails, the C++ bad_alloc would unwind the stack, free some stack/heap memory through RAII. The recovering feature would then salvage as much as possible (saving the initial data of the task on the disk, to use on a later try), and perhaps register the task data for later try.
I do believe the use of C++ strong/nothrow guanrantees can help a process to survive in low-available-memory conditions, even if it would be akin memory swapping (i.e. slow, somewhat unresponding, etc.), but of course, this is only theory. I just need to get smarter on the subject before trying to simulate this (i.e. creating a C++ program, with a custom new/delete allocator with limited memory, and then try to do some work under those stressful condition).
Well...
Out of memory normally means you have to quit whatever you were doing. If you are careful about cleanup, though, it can leave the program itself operational and able to respond to other requests. It's better to have a program say "Sorry, not enough memory to do " than say "Sorry, out of memory, shutting down."
Out of memory can be caused either by free memory depletion or by trying to allocate an unreasonably big block (like one gig). In "depletion" cases memory shortage is global to the system and usually affects other applications and system services and the whole system might become unstable so it's wise to forget and reboot. In "unreasonably big block" cases no shortage actually occurs and it's safe to continue. The problem is you can't automatically detect which case you're in. So it's safer to make the error non-recoverable and find a workaround for each case you encounter this error - make your program use less memory or in some cases just fix bugs in code that invokes memory allocation.
There are already many good answers here. But I'd like to contribute with another perspective.
Depletion of just about any reusable resource should be recoverable in general. The reasoning is that each and every part of a program is basically a sub program. Just because one sub cannot complete to it's end at this very point in time, does not mean that the entire state of the program is garbage. Just because the parking lot is full of cars does not mean that you trash your car. Either you wait a while for a booth to be free, or you drive to a store further away to buy your cookies.
In most cases there is an alternative way. Making an out of error unrecoverable, effectively removes a lot of options, and none of us like to have anyone decide for us what we can and cannot do.
The same applies to disk space. It's really the same reasoning. And contrary to your insinuation about stack overflow is unrecoverable, i would say that it's and arbitrary limitation. There is no good reason that you should not be able to throw an exception (popping a lot of frames) and then use another less efficient approach to get the job done.
My two cents :-)
If you are really out of memory you are doomed, since you can not free anything anymore.
If you are out of memory, but something like a garbage collector can kick in and free up some memory you are non dead yet.
The other problem is fragmentation. Although you might not be out of memory (fragmented), you might still not be able to allocate the huge chunk you wanna have.
I know you asked for arguments for, but I can only see arguments against.
I don't see anyway to achieve this in a multi-threaded application. How do you know which thread is actually responsible for the out-of-memory error? One thread could allocating new memory constantly and have gc-roots to 99% of the heap, but the first allocation that fails occurs in another thread.
A practical example: whenever I have occurred an OutOfMemoryError in our Java application (running on a JBoss server), it's not like one thread dies and the rest of the server continues to run: no, there are several OOMEs, killing several threads (some of which are JBoss' internal threads). I don't see what I as a programmer could do to recover from that - or even what JBoss could do to recover from it. In fact, I am not even sure you CAN: the javadoc for VirtualMachineError suggests that the JVM may be "broken" after such an error is thrown. But maybe the question was more targeted at language design.
uClibc has an internal static buffer of 8 bytes or so for file I/O when there is no more memory to be allocated dynamically.
What is the compelling argument for making it a recoverable error?
In Java, a compelling argument for not making it a recoverable error is because Java allows OOM to be signalled at any time, including at times where the result could be your program entering an inconsistent state. Reliable recoery from an OOM is therefore impossible; if you catch the OOM exception, you can not rely on any of your program state. See
No-throw VirtualMachineError guarantees
I'm working on SpiderMonkey, the JavaScript VM used in Firefox (and gnome and a few others). When you're out of memory, you may want to do any of the following things:
Run the garbage-collector. We don't run the garbage-collector all the time, as it would kill performance and battery, so by the time you're reaching out of memory error, some garbage may have accumulated.
Free memory. For instance, get rid of some of the in-memory cache.
Kill or postpone non-essential tasks. For instance, unload some tabs that haven't be used in a long time from memory.
Log things to help the developer troubleshoot the out-of-memory error.
Display a semi-nice error message to let the user know what's going on.
...
So yes, there are many reasons to handle out-of-memory errors manually!
I have this:
void *smalloc(size_t size) {
void *mem = null;
for(;;) {
mem = malloc(size);
if(mem == NULL) {
sleep(1);
} else
break;
}
return mem;
}
Which has saved a system a few times already. Just because you're out of memory now, doesn't mean some other part of the system or other processes running on the system have some memory they'll give back soon. You better be very very careful before attempting such tricks, and have all control over every memory you do allocate in your program though.