Memory Regions
A significant improvement to Maple's memory allocator is the addition of multiple memory regions.
Benefits of multiple memory regions include:
More efficient allocation of memory, leading to more memory being available to solve bigger problems without the need to initially specify the size of the reserved region.
Better resource sharing between different processes running on the same machine, including multiple instances of Maple and multiple Maple kernels launched as parallel grid nodes.
Performance improvements due to less fragmentation and cache locality, as well as performance improvements when running parallel code.
Previously, Maple initially reserved a large contiguous heap of memory that was subsequently managed by additional allocations and garbage collection as required. Initially, only a small fraction of this reserved memory was allotted to handle memory allocation requests. Whenever a given allotment was exhausted, a garbage collection would occur to reclaim any unused memory. As the computation required more memory, the memory management system would progressively commit more of the heap to be allocated.
With the addition of memory regions, Maple is no longer constrained by the limitation imposed by a single large contiguous memory region (heap). Instead, Maple is able to incorporate additional regions upon demand. Various types of memory regions are provided: small allocations come from thread-local regions (512 words and less), medium-sized allocations reside in global thread-shared regions (less than 1MB), and finally large blocks of memory are individually allocated as distinct regions of their own. Overall, this allows Maple to more effectively and efficiently manage the available memory resources, and more specifically, this offers improvements to caching via better data locality and reduced fragmentation.
In earlier releases of Maple, when a chunk of memory was committed, it remained under Maple's control for the remainder of the session. Handling memory requests larger than 1MB by allocating an individually tailored memory region enables Maple 17 to return this large block of memory back to the operating system when it is no longer needed.
For example, the following code snippet allocates an array of matrices, does some work, and then clobbers the array thereby rendering the allocated memory garbage. Notice that memory usage drops after the explicit call to garbage collection.*
restart;
size ≔ 1024:elems ≔ 10:A ≔ Array 1.. elems, 1..elems :for i to elems do for j to elems do Ai, j ≔ Matrix size, size, datatype = float8 : end do: end do:
Report how many KB Maple is using.
kerneloptsbytesalloc1024;
850704
A ≔ 0:
gc;kerneloptsbytesalloc1024;
88476
* The explicit call to gc() is used here to illustrate a specific behavior of Maple's memory management system. In general, it is better to allow Maple to decide when to initiate a garbage collection cycle.
See Also
gc, kernelopts
Download Help Document
