Fig. 1. CPU time (milliseconds) consumed to perform RSWR algorithms. (a) CPU times of all three RSWR algorithms (left) and (b) CPU times of RSWR-batch and RSWR-skip (right). For each algorithm, the reservoir size is set to 1000 and stream size goes up to 1 million. CPU times are incrementally measured in every 10K time stamps. Note that the plots of RSWR-batch and RSWR-skip are redrawn in a separate graph due to their low complexities compared to that of RSWR-naive.

Fig. 2. CPU times (milliseconds) of RSWR algorithms with different reservoir sizes. (a) All three RSWR algorithms (left) and (2) RSWR-batch and RSWR-skip (right). For each case, the stream size is fixed at 100K. Note that the plots of RSWR-batch and RSWR-skip are redrawn in a separate graph due to their low complexities compared to that of RSWR-naive.

Fig. 3. (a) CPU time (milliseconds) of RSWR-batch and RSWR-skip measured at early time stamps (Left). (b) The number of skipped elements measured at different time intervals (Right). In (a), the reservoir size is fixed at 10K, and CPU times are measured in every 5K time stamps. The time stamp goes up to 100K. In (b), the reservoir size is set to 1000, the stream size goes up to 1 million, and the number of skipped elements are measured in every 10K.

Fig. 4. Time stamp distribution of elements in the reservoir after the 1 millionth time stamp. For each RSWR algorithm, the entire interval of 1 million time stamps is split into 20 equal intervals, and the number of elements from each interval that are in the final reservoir is counted. The reservoir size is set to 1000. The frequencies are averaged after 10 runs.