The exploration of brain networks with resting-state fMRI (rs-fMRI) combined with graph theoretical approaches has become popular, with the perspective of finding network graph metrics as biomarkers in the context of clinical studies. A preliminary requirement for such findings is to assess the reliability of the graph based connectivity metrics. In previous test-retest (TRT) studies, this reliability has been explored using intraclass correlation coefficient (ICC) with heterogeneous results. But the issue of sample size has not been addressed. Using the large TRT rs-fMRI dataset from the Human Connectome Project (HCP), we computed ICCs and their corresponding p-values (applying permutation and bootstrap techniques) and varied the number of subjects (from 20 to 100), the scan duration (from 400 to 1200 time points), the cost and the graph metrics, using the Anatomic-Automatic Labelling (AAL) parcellation scheme. We quantified the reliability of the graph metrics computed both at global and regional level depending, at optimal cost, on two key parameters, the sample size and the number of time points or scan duration. In the cost range between 20% to 35%, most of the global graph metrics are reliable with 40 subjects or more with long scan duration (14 min 24 s). In large samples (for instance, 100 subjects) most global and regional graph metrics are reliable for a minimum scan duration of 7 min 14s. Finally, for 40 subjects and long scan duration (14 min 24 s), the reliable regions are located in the main areas of the default mode network (DMN), the motor and the visual networks.