t-test used to test means of two populations assumes that  each population is normally distributed. Theoretically, violation of the assumption makes the result of the test  invalid. This research evaluates the robustness of the t-test on various value of sample size using three types of distribution: normal, symmetric non-normal, and not symmetric non-normal. Different computation techniques of t-value which depend on the variance of the two populations were also employed. The simulation showed that t-test used to test mean of two populations is not influenced by non-normality of the population distribution. The exploration of distribution of the difference between two samples means showed that its distribution was normal. Therefore, the robustness against non-normality of the t-test was the consequences of the difference between two sample means that normally is distributed.