Analysis of various scalar , vector , and parallel implementations of RandomAccess ∗

RandomAccess test (previously called GUPS and guppie) measures a sustained rate (measured in GUPS) of updates to random locations in main memory. The random numbers may be generated by a variety of methods. In this paper two random generators are considered: one based on shift registers and one based on linear congruence relation. The quality of the main memory address distribution is considered together with the performance of the generation process. Bothe aspect are measured in various ways to create an objective setting for comparison of the generators. The tests and comparisons include both sequential as well as parallel execution each possibly using code vectorization. RandomAccess relies on a pseudo Random Number Generator (RNG) to create a sequence of random memory locations that are updated using a simple bit-wise operation. The computational load of the update is very low. Instead, the fetch and store of the data required for the update are time consuming due to the randomness of the sequence of memory locations that span at least half ot the main memory of the tested computer. The RNG used in RandomAccess needs to preserve this balance: generating the random numbers needs to be computationally cheap and take less time than a cache miss (and possibly a Translation Look-aside Buffer – TLB – miss) that can only be resolved by the main memory and not by any of the cache levels.