I just encountered a problem I cannot really understand. I suspect my benchmarking results are not accurate. The first time i ran it, as expected, the SyncMap + Load > LoadOrStore implementation outperforms the rest in every metric. Well as usual, i would run it once more just to be safe and guess what Map + Mutex outperforms it by almost 50%… Well maybe it was a fluke… But no! After a few more runs, it became more certain that the results always varies. They simply are not consistent, heck look at the below result, even in the same go test , the same implementation can have such different results (just look at SyncMap_+_LoadOrStore ).