|Version 3 (modified by simonmar, 8 years ago) (diff)|
GHC Commentary: The Storage Manager
GHC's storage manager is designed to be quite flexible: there are a large number of tunable parameters in the garbage collector, and partly the reason for this was because we wanted to experiment with tweaking these settings in the context of Haskell.
The Block Allocator
The block allocator is where the storage manager derives much of its flexibilty. Rather than keep our heap in a single contiguous region of memory, or one contiguous region per generation, we manage linked lists of memory blocks. Managing contiguous regions is difficult, especially when you want to change the size of some of the areas. A block-structured storage arrangement has several advantages:
- resizing areas of memory is easy: just chain more blocks onto the list.
- managing large objects without copying is easy: allocate each one a complete block, and use the block linkage to chain them together.
- free memory can be recycled faster, because a block is a block.
The concept relies on the property that most data objects are significantly smaller than a block, and only rarely do we need to allocate objects that approach or exceed the size of a block.
Structure of blocks
We want to allocate memory in units of a small block (around 4k, say). Furthermore, we want each block to have an associated small structure called a block descriptor, which contains information about the block: its link field, which generation it belongs to, and so on. We want a function Bdescr(p), that, given an arbitrary pointer into a block, returns the address of the block descriptor that corresponds to the block containing that pointer.
There are two options:
- Put the block descriptor at the start of the block. Bdescr(p) = p & ~BLOCK_SIZE. This option has problems if we need to allocate a contiguous region larger than a single block (GHC does this occasionally when allocating a large number of objects in one go).
- Allocate memory in larger units (a megablock), divide the megablock into blocks, and put all the block descriptors at the beginning. The megablock is aligned, so that the address of the block descriptor for a block is a simple function of its address. The 'Bdescr' function is more complicated than the first method, but it is easier to allocate contiguous regions (unless the contiguous region is larger than a megablock...).