In computer science, locality of reference, also known as the principle of locality, is the tendency of a processor to access the same set of memory locations repetitively over a short period of time. There are two basic types of reference locality – temporal and spatial locality. Temporal locality refers to the reuse of specific data, and/or resources, within a relatively small time duration. Spatial locality (also termed data locality) refers to the use of data elements within relatively close storage locations. Sequential locality, a special case of spatial locality, occurs when data elements are arranged and accessed linearly, such as, traversing the elements in a one-dimensional array.
Locality is a type of predictable behavior that occurs in computer systems. Systems that exhibit strong locality of reference are great candidates for performance optimization through the use of techniques such as the caching, prefetching for memory and advanced branch predictors at the pipelining stage of a processor core.
There are several different types of locality of reference:
In order to benefit from the very frequently occurring temporal and spatial locality, most of the information storage systems are hierarchical. The equidistant locality is usually supported by the diverse nontrivial increment instructions of the processors. For branch locality, the contemporary processors have sophisticated branch predictors, and on the basis of this prediction the memory manager of the processor tries to collect and preprocess the data of the plausible alternatives.
There are several reasons for locality. These reasons are either goals to achieve or circumstances to accept, depending on the aspect. The reasons below are not disjoint; in fact, the list below goes from the most general case to special cases:
If most of the time the substantial portion of the references aggregate into clusters, and if the shape of this system of clusters can be well predicted, then it can be used for performance optimization. There are several ways to benefit from locality using optimization techniques. Common techniques are:
Hierarchical memory is a hardware optimization that takes the benefits of spatial and temporal locality and can be used on several levels of the memory hierarchy. Paging obviously benefits from temporal and spatial locality. A cache is a simple example of exploiting temporal locality, because it is a specially designed, faster but smaller memory area, generally used to keep recently referenced data and data near recently referenced data, which can lead to potential performance increases.
Data elements in a cache do not necessarily correspond to data elements that are spatially close in the main memory; however, data elements are brought into cache one cache line at a time. This means that spatial locality is again important: if one element is referenced, a few neighboring elements will also be brought into cache. Finally, temporal locality plays a role on the lowest level, since results that are referenced very closely together can be kept in the machine registers. Some programming languages (such as C) allow the programmer to suggest that certain variables be kept in registers.
Data locality is a typical memory reference feature of regular programs (though many irregular memory access patterns exist). It makes the hierarchical memory layout profitable. In computers, memory is divided into a hierarchy in order to speed up data accesses. The lower levels of the memory hierarchy tend to be slower, but larger. Thus, a program will achieve greater performance if it uses memory while it is cached in the upper levels of the memory hierarchy and avoids bringing other data into the upper levels of the hierarchy that will displace data that will be used shortly in the future. This is an ideal, and sometimes cannot be achieved.
Typical memory hierarchy (access times and cache sizes are approximations of typical values used as of 2013for the purpose of discussion; actual values and actual numbers of levels in the hierarchy vary):
Modern machines tend to read blocks of lower memory into the next level of the memory hierarchy. If this displaces used memory, the operating system tries to predict which data will be accessed least (or latest) and move it down the memory hierarchy. Prediction algorithms tend to be simple to reduce hardware complexity, though they are becoming somewhat more complicated.
A common example is matrix multiplication:
1 for i in 0..n 2 for j in 0..m 3 for k in 0..p 4 C[i][j] = C[i][j] + A[i][k] * B[k][j];
By switching the looping order for
k, the speedup in large matrix multiplications becomes dramatic, at least for languages that put contiguous array elements in the last dimension. This will not change the mathematical result, but it improves efficiency. In this case, "large" means, approximately, more than 100,000 elements in each matrix, or enough addressable memory such that the matrices will not fit in L1 and L2 caches.
1 for i in 0..n 2 for k in 0..p 3 for j in 0..m 4 C[i][j] = C[i][j] + A[i][k] * B[k][j];
The reason for this speedup is that in the first case, the reads of
A[i][k] are in cache (since the
k index is the contiguous, last dimension), but
B[k][j] is not, so there is a cache miss penalty on
C[i][j] is irrelevant, because it can be hoisted out of the inner loop -- the loop variable there is
1 for i in 0..n 2 for j in 0..m 3 temp = C[i][j] 4 for k in 0..p 5 temp = temp + A[i][k] * B[k][j]; 6 C[i][j] = temp
In the second case, the reads and writes of
C[i][j] are both in cache, the reads of
B[k][j] are in cache, and the read of
A[i][k] can be hoisted out of the inner loop.
1 for i in 0..n 2 for k in 0..p 3 temp = A[i][k] 4 for j in 0..m 5 C[i][j] = C[i][j] + temp * B[k][j];
Thus, the second example has no cache miss penalty in the inner loop while the first example has a cache penalty.
On a year 2014 processor, the second case is approximately five times faster than the first case, when written in C and compiled with
gcc -O3. (A careful examination of the disassembled code shows that in the first case, GCC uses SIMD instructions and in the second case it does not, but the cache penalty is much worse than the SIMD gain.)
Temporal locality can also be improved in the above example by using a technique called blocking. The larger matrix can be divided into evenly sized sub-matrices, so that the smaller blocks can be referenced (multiplied) several times while in memory.
1 for (ii = 0; ii < SIZE; ii += BLOCK_SIZE) 2 for (kk = 0; kk < SIZE; kk += BLOCK_SIZE) 3 for (jj = 0; jj < SIZE; jj += BLOCK_SIZE) 4 maxi = min(ii + BLOCK_SIZE, SIZE); 5 for (i = ii; i < maxi; i++) 6 maxk = min(kk + BLOCK_SIZE, SIZE); 7 for (k = kk; k < maxk; k++) 8 maxj = min(jj + BLOCK_SIZE, SIZE); 9 for (j = jj; j < maxj; j++) 10 C[i][j] = C[i][j] + A[i][k] * B[k][j];
The temporal locality of the above solution is provided because a block can be used several times before moving on, so that it is moved in and out of memory less often. Spatial locality is improved because elements with consecutive memory addresses tend to be pulled up the memory hierarchy together.