11. MVT (Manual Variable Temporal)

MVT manually manages variable-sized, unformatted objects. It uses the temporal fit allocation policy.

11.1. Temporal fit

Temporal fit attempts to place consecutive allocations next to each other. It relies on delaying re-use as long as possible to permit freed blocks to coalesce, thus maximizing the number of consecutive allocations that can be co-located. Temporal fit permits a very fast allocator and a deallocator competitive in speed with all other known policies.

Temporal fit is intended to take advantage of knowledge of object lifetimes: either a priori knowledge, or knowledge acquired by profiling. The best performance will be achieved by allocating objects with similar expected death times together.

A simple policy can be implemented to take advantage of MVT. Object size is typically well-correlated with object life-expectancy, and birth time plus lifetime gives death time, so allocating objects of similar size sequentially from the same pool instance should result in objects allocated close to each other dying at about the same time.

An application that has several classes of objects of widely differing life expectancy will best be served by creating a different MVT pool instance for each life-expectancy class. A more sophisticated policy can use either the programmer’s knowledge of the expected lifetime of an object, or any characteristic of objects that correlates with lifetime, to choose an appropriate pool to allocate in.

Allocating objects with unknown or very different death times together will pessimize the space performance of MVT.

11.2. MVT properties

11.3. MVT interface

#include "mpscmvt.h"
mps_pool_class_t mps_class_mvt(void)

Return the pool class for an MVT (Manual Variable Temporal) pool.

When creating an MVT pool, mps_pool_create_k() accepts six optional keyword arguments:

  • MPS_KEY_ALIGN (type mps_align_t, default is MPS_PF_ALIGN) is the alignment of the addresses allocated (and freed) in the pool. The minimum alignment supported by pools of this class is sizeof(void *) and the maximum is the arena grain size (see MPS_KEY_ARENA_GRAIN_SIZE).

  • MPS_KEY_MIN_SIZE (type size_t, default is MPS_PF_ALIGN) is the predicted minimum size of blocks that will be allocated from the pool.

  • MPS_KEY_MEAN_SIZE (type size_t, default 32) is the predicted mean size of blocks that will be allocated from the pool.

  • MPS_KEY_MAX_SIZE (type size_t, default 8192) is the predicted maximum size of blocks that will be allocated from the pool. Partial freeing is not supported for blocks larger than this; doing so will result in the storage of the block never being reused.

The three SIZE arguments above are hints to the MPS: the pool will be less efficient if they are wrong, but the only thing that will break is the partial freeing of large blocks.

  • MPS_KEY_MVT_RESERVE_DEPTH (type mps_word_t, default 1024) is the expected hysteresis of the population of the pool. When blocks are freed, the pool will retain sufficient storage to allocate this many blocks of the mean size for near term allocations (rather than immediately making that storage available to other pools).

    If a pool has a stable population, or one which only grows over the lifetime of the pool, or one which grows steadily and then shrinks steadily, use a reserve depth of 0.

    It is always safe to use a reserve depth of 0, but if the population typically fluctuates in a range (for example, the client program repeatedly creates and destroys a subset of blocks in a loop), it is more efficient for the pool to retain enough storage to satisfy that fluctuation. For example, if a pool has an object population that typically fluctuates between 8,000 and 10,000, use a reserve depth of 2,000.

    The reserve will not normally be available to other pools for allocation, even when it is not used by the pool. If this is undesirable, a reserve depth of 0 may be used for a pool whose object population does vary, at a slight cost in efficiency. The reserve does not guarantee any particular amount of allocation.

  • MPS_KEY_MVT_FRAG_LIMIT (type double, default 0.3) may range from 0.0 to 1.0 (inclusive). It sets an upper limit on the space overhead of an MVT pool, in case block death times and allocations do not correlate well. If the free space managed by the pool as a ratio of all the space managed by the pool exceeds the fragmentation limit, the pool falls back to a first fit allocation policy, exploiting space more efficiently at a cost in time efficiency. A fragmentation limit of 0.0 would cause the pool to operate as a first-fit pool, at a significant cost in time efficiency: therefore this is not permitted.

    A fragmentation limit of 1.0 causes the pool to always use temporal fit (unless resources are exhausted). If the objects allocated in the pool have similar lifetime expectancies, this mode will have the best time- and space-efficiency. If the objects have widely varying lifetime expectancies, this mode will be time-efficient, but may be space-inefficient. An intermediate setting can be used to limit the space-inefficiency of temporal fit due to varying object life expectancies.

For example:

MPS_ARGS_BEGIN(args) {
    MPS_ARGS_ADD(args, MPS_KEY_MIN_SIZE, 4);
    MPS_ARGS_ADD(args, MPS_KEY_MEAN_SIZE, 32);
    MPS_ARGS_ADD(args, MPS_KEY_MAX_SIZE, 1024);
    MPS_ARGS_ADD(args, MPS_KEY_MVT_RESERVE_DEPTH, 256);
    MPS_ARGS_ADD(args, MPS_KEY_MVT_FRAG_LIMIT, 0.5);
    res = mps_pool_create_k(&pool, arena, mps_class_mvt(), args);
} MPS_ARGS_END(args);