The alg5_rand_mem_bench.c is a parallel memory benchmark for random memory accesses in four domains:
- Domain 1: each core perfroms memory accesses to its local L3 cache,
- Domain 2: each core perfroms memory accesses to all L3 caches of its NUMA node,
- Domain 3: each core perfroms memory accesses to all L3 caches, and
- Domain 4: each core perfroms memory accesses that is missed by all L3 caches.
The code evaluates 3 types of memory accesses:
- Atomic write
- Write
- Read
To run the benchmark you need to run make alg5_rand_mem_bench.
To remove hyper threading, run make alg5_rand_mem_bench no_ht=1.
The results are shown as the Million Transactions per second (MT/s). A sample output is shown in the following.
