Hi Clement,
Nice article... https://clementbera.wordpress.com/2018/06/07/free-chunk-management-in-the-co...
A random thought, I wonder if size-1 chunks can be managed packed into size-2 chunks.
Have a variable size1_active_double initially nil. First size-1 allocation, grab a size-2 chunk into size1_active_double. First half returned as the size-1-chunk. Second half marked at "empty".
Second size-1 allocation, return second half of size1_active_double as size-1-chunk. Set size1_active_double to nil.
Releasing a size-1-chunk. Dependent on whether memory alignment can inform of which half of size-2-chunk is being released. Check if other half is empty ==> add it to the size-2-chunk-list.
If other half not empty, add it to one of two singly-linked-lists depending on which half is empty, the empty half used for the linked-list. Next size-1 allocations made from these lists first to fill in the empty half.
Considerations * Added complexity * Remaining size-1 singly-linked-list might have minimal impact - added complexity not worth the gain.
____________ Or another random thought *grin* - and this is probably completely off base because of my limited understanding. Piggyback the size-1-chunk management on top of the size-3-chunk-list. Consider one end of the size-3-chunk-list to be "size1-filled-end" with references to free size-1-chunks, and the other end of the size-3-chunk-list to be "size1-empty-end".
To release a size-1-chunk, move a size-3-chunk from the "size1-empty-end" to the "size1-filled-end" and store the size-1-chunk reference there.
To allocate a size-1-chunk, get it from the "size1-filled" end of the size-3-chunk-list and move that size-3-chunk to the "size1-empty-end" of its list.
To compact a size-3-chunk that refers to a size-1-chunk, copy that into another chunk from the size1-empty-end moved to the size1-filled-end.
I wonder if that even makes sense. Anyway, it hurts less to let the ideas roam free...
cheers -ben
Hi Ben,
On Sun, Jun 17, 2018 at 4:11 PM, Ben Coman btc@openinworld.com wrote:
Hi Clement,
Nice article... https://clementbera.wordpress.com/2018/06/07/free-chunk- management-in-the-cog-vm/
A random thought, I wonder if size-1 chunks can be managed packed into size-2 chunks.
A key design feature of Spur is that every object has at least one field to function as a forwarding pointer. Hence even zero-sized objects occupy 128 bits. Each object has a 64-bit header (large objects have a 128-bit header). The allocation unit is 64-bits. All objects must have room for a forwarding pointer. Hence the minimum sized chunk is 128-bits.
Hence there isn't room for two size 1 chunks inside a size-2 chunk, only inside a size three chunk.
Have a variable size1_active_double initially nil. First size-1 allocation, grab a size-2 chunk into size1_active_double. First half returned as the size-1-chunk. Second half marked at "empty".
Second size-1 allocation, return second half of size1_active_double as size-1-chunk. Set size1_active_double to nil.
Releasing a size-1-chunk. Dependent on whether memory alignment can inform of which half of size-2-chunk is being released. Check if other half is empty ==> add it to the size-2-chunk-list.
If other half not empty, add it to one of two singly-linked-lists depending on which half is empty, the empty half used for the linked-list. Next size-1 allocations made from these lists first to fill in the empty half.
Considerations
- Added complexity
- Remaining size-1 singly-linked-list might have minimal impact - added
complexity not worth the gain.
Or another random thought *grin* - and this is probably completely off base because of my limited understanding. Piggyback the size-1-chunk management on top of the size-3-chunk-list. Consider one end of the size-3-chunk-list to be "size1-filled-end" with references to free size-1-chunks, and the other end of the size-3-chunk-list to be "size1-empty-end".
To release a size-1-chunk, move a size-3-chunk from the "size1-empty-end" to the "size1-filled-end" and store the size-1-chunk reference there.
To allocate a size-1-chunk, get it from the "size1-filled" end of the size-3-chunk-list and move that size-3-chunk to the "size1-empty-end" of its list.
To compact a size-3-chunk that refers to a size-1-chunk, copy that into another chunk from the size1-empty-end moved to the size1-filled-end.
I wonder if that even makes sense. Anyway, it hurts less to let the ideas roam free...
cheers -ben
On Mon, 18 Jun 2018, Ben Coman wrote:
A random thought, I wonder if size-1 chunks can be managed packed into
size-2 chunks.
Or, just use the XOR trick to store each node's two links in a single pointer sized field: https://everything2.com/title/Storing+a+doubly-linked+list+using+just+a+sing...
Levente
Hi Levente,
On Jun 17, 2018, at 5:36 PM, Levente Uzonyi leves@caesar.elte.hu wrote:
On Mon, 18 Jun 2018, Ben Coman wrote:
A random thought, I wonder if size-1 chunks can be managed packed into
size-2 chunks.
Or, just use the XOR trick to store each node's two links in a single pointer sized field: https://everything2.com/title/Storing+a+doubly-linked+list+using+just+a+sing...
Alas, while this lovely trick does indeed encode a doubly-linked list in a single field it only works for full traversals. The xor is of the two neighbours, so to get to the next one needs the prev, and to get to the prev one needs the next. So one can start from either end but not in the middle. Clément’s modification is to allow rapid removal without needing to traverse the entire list, so the xor trick is not fit for purpose here.
BTW the compactor I wrote before the current one (SpurPigCompactor preceded SpurPlanningCompactor) used exactly this trick. It didn’t work for reasons unrelated to the xor trick. Just mentioning it as proof that I love the technique.
Levente
Hi Eliot,
On Sun, 17 Jun 2018, Eliot Miranda wrote:
Hi Levente,
On Jun 17, 2018, at 5:36 PM, Levente Uzonyi leves@caesar.elte.hu wrote:
On Mon, 18 Jun 2018, Ben Coman wrote:
A random thought, I wonder if size-1 chunks can be managed packed into
size-2 chunks.
Or, just use the XOR trick to store each node's two links in a single pointer sized field: https://everything2.com/title/Storing+a+doubly-linked+list+using+just+a+sing...
Alas, while this lovely trick does indeed encode a doubly-linked list in a single field it only works for full traversals. The xor is of the two neighbours, so to get to the next one needs the prev, and to get to the prev one needs the next. So one can start from either end but not in the middle. Clément’s modification is to allow rapid removal without needing to traverse the entire list, so the xor trick is not fit for purpose here.
Well, it wasn't clear if that's a requirement, but I could have figured it out, because if you always iterate over the list, you can just keep a pointer to the previous node to delete the current node.
If performance is important here, random deletion can still be done in O(1) time at the cost of maintaining an external doubly linked list with each node having a pointer to the chunk and the chunk using its sole slot to point to its list node.
BTW the compactor I wrote before the current one (SpurPigCompactor preceded SpurPlanningCompactor) used exactly this trick. It didn’t work for reasons unrelated to the xor trick. Just mentioning it as proof that I love the technique.
Nice. :)
Levente
Levente
vm-dev@lists.squeakfoundation.org