Implement memory_usage() builtin function #25
Reference in New Issue
Block a user
No description provided.
Delete Branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
In scope of this ticket, we need to implement additional builtin function - memory_usage(). It should return a number of bytes allocated by Aer Interpreter.
Implement get_memory_usage() builtin functionto Implement memory_usage() builtin functionIs it relevant (and possible) as well the max memory usage ?
Yeah, why not. I think this task can be even bigger. If we need to modify the memory allocation mechanism, to save such information, we can also already check if current memory usage + what we are going to allocate does not exceed some limit set.
I made a basic version, of the first part only, locally but not sure about the approach.
1/ Storing the info into VM allocator.
2/ Counting occurs within MemBackendAlloc/Realloc (i.e the "old" address)/Free.
3/ But then it s counting the total size of the block rather than just the requested size.
Generally speaking, I was thinking about some structure storing all necessary information, like for example:
This is some sort of minimal version storing all data in bytes.
SyOSHeapAlloc() and related takes sxu32 nByte argument which tells how much memory should be allocated. This is finally passed to malloc() or realloc() under Unix-family OS. I think we can this value simply add to ph7_heap.size and check if it is bigger than peak. If so, overwrite the peak with current value.
The SyOSHeapRealloc() and SyOSHeapFree() takes a pointer which I think contains information about current size, doesn't it?
Also having the limit set, we can check if size + requested amount of memory does not exceed the limit.
I think I don't understand why it should be counting size of block?
Yes pointer contains info about the size this is how xChunkSize function pointer count on.
In fact that was my question, counting the real size allocated vs just the bytes requested.
Ok I see your point but maybe the type of the heap struct are a bit small if we count in byte and why signed types ?
The structure was just a quick example. AFAIK this is how PHP counts the memory, however its zend_mm_heap structure contains more information, like about list of huge blocks, unused chinks, cached chunks, and so on...
Ok I ve made a quick test (did not do the limit part yet)
`
class Main {
}
new Main();
`
resulting in my case
283377 283297 TRUEAs a test would display only the latter for the .exp counterpart ...
Basically what I did :
1/ ph7_heap type within SyMemBackend, all zeroed at init already.
2/ Changing the MemOS* signatures to accept this type, updating the peak accordingly too.
3/ I named similarly to php but it does not please, easy to rename.
`
I just like to add, that it might be hard to use above code as test. I imagine that memory usage values can be different every script run, and thus test will fail. If we want to test it, we might need to think about some script not showing the integer values.
I know exactly what I said -> showing only the last case :-) the values was only for demo here.
Do you think we could check if a value returned by both functions are positive integers and simply output true or false? Does such a test make sense?
You mean > 0 (in my case types are 64 types unsigned) ? Yes we could add these too.
A little progress, the memory limit for the moment passed by the command line (did not copy yet the multiplier behavior of php i.e 100M etc)
binary/aer -M100000 tests/memory_usage.aer Out of memory, max allowed 100000 bytes, allocated 90328 bytesFinally got a relatively working version of these, thanks to @belliash for early tests !
I moved all work to memory_limit branch.
There is still some issues to address. There is a problem with displaying an 'out of memory' error message if some limit is set.
I have reviewed this once again and I think there is no need to pass SyMemBackend structure down to MemOS* functions as the calculations can be done in MemBackedn* functions. They already have access to SyMemBackend, what simplifies whole implementation.
I also wonder if we could use pointers, in SyMemBackendInit() and SyMemBackendInitFromParent() to avoid copying the memory usage statistics data (ph7_heap structure)...
I have tried but the values are very inconsistent in this case. Plus not sure the statistics are copied between backends ? Even looking at InitFromParent, only the memory limit is copied.
This implementation is based on @devnexen work and the attached patch can be applied on top of master branch. I think everything is implemented and working, except there is no SAPI parameter to set the memory limit. Still needs some testing.
@devnexen, looking forward for your review.
The MemBackendCalculate could be a macro to avoid one function call otherwise looks reasonably good for a first stage. Not against calculating the chunk size instead of the fcn. pointer. Also yes needs testing (i.e. realloc data for example).