RFR: 8333658: NMT: Use an allocator with 4-byte pointers to save memory in NativeCallStackStorage

[Thomas Stuefe]

On Thu, 6 Jun 2024 14:06:22 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

>> Hi,
>> 
>> This PR introduces a new allocator, `IndexedFreelistAllocator`. It uses a `GrowableArray` in order to have 4-byte "pointers" to its elements and also works as a freelist as unused slots form a linked list. This allocator cannot shrink its used memory. I'm always open for better names.
>> 
>> We then use this allocator in order to store the `NativeCallStackStorage` hash table. This saves `4 * 4099 = ~16KiB` of memory as each table element is now only 4 bytes. For each stored stack it also saves 4 bytes. The fact that the allocator cannot shrink is fine, we do not ever remove stacks from the storage.
>> 
>> The main point of this is to introduce a pointer-generic experimential API, so I also implemented CHeap and Arena allocators. It's currently placed in NMT, but we might want to move it into utilities. It uses a bit of template-magic, but my IDE (Emacs+clangd) was actually able to figure things out when the types didn't line up correctly etc, so it's not an enemy to IDE help.
>> 
>> It sounded expensive to have the GrowableArray continously realloc its underlying data array, so I did a basic test where we allocate 1 000 000 stacks and push them into NativeCallStackStorage backed by different allocators. This is available in the PR.
>> 
>> The results are as follows on linux-x64-slowdebug:
>> 
>> 
>> Generate stacks... Done
>> Time taken with GrowableArray: 8341.240945
>> Time taken with CHeap: 12189.031318
>> Time taken with Arena: 8800.703092
>> Time taken with GrowableArray again: 8295.508829
>> 
>> 
>> And on linux-x64:
>> 
>> 
>> Time taken with GrowableArray: 8378.018135
>> Time taken with CHeap: 12437.347868
>> Time taken with Arena: 8758.064717
>> Time taken with GrowableArray again: 8391.076291
>> 
>> 
>> Obviously, this is a very basic benchmark, but it seems like we're faster than CHeap and Arena for this case.
>
> src/hotspot/share/nmt/indexedFreeListAllocator.hpp line 58:
> 
>> 56:       // Follow the link to the next free element
>> 57:       free_start = be.link;
>> 58:     }
> 
> This feels a bit uncommon and contrary to the normal linked list implementation. If possible, I would opt for a more traditional approach that is easier to understand and does not rely on at_grow (e.g. in case we want adapt this to some other form of backing storage).
> 
> e.g.
> 
> 
> if (_freelist == -1) {
>   allocate new slot   
> } else {
>   reuse slot at freelist
> }

Note that at_grow also could have the problem that an error, resulting in a high index, would cause a large allocation to happen.

I think this class could use an optional max value.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/18979#discussion_r1629643025