©　2017.　Thread-level　speculation　(TLS)　is　an　automatic　parallelization　technique　for　serial　programs　on　multicore　platforms,　and　it　permits　to　generate　multiple　threads　during　compiling　as　well　as　to　run　them　at　runtime.　Thread　partition　plays　an　important　role　in　TLS.　Heuristic　rules-based　(HR-based)　partition　and　machine　learning-based　(ML-　based)　partition　are　two　commonly　used　approaches.　During　the　partition　of　these　two　approaches,　what　seriously　affect　the　partition　effect　are　five　thresholds,　i.e.,　upper　limit　of　spawning　distance　(ULoSD),　lower　limit　of　spawning　distance　(LLoSD),　data　dependence　count　(DDC),　upper　limit　of　thread　granularity　(ULoTG),　lower　limit　of　thread　granularity　(LLoTG).　This　paper　proposes　to　optimize　these　five　thresholds　with　a　level-based　traversal　method,　in　order　to　find　the　optimal　partition　thresholds　for　every　procedure,　so　as　to　obtain　the　optimal　partition.　Prophet,　which　consists　of　an　automatic　partition　compiler　and　a　simulator,　is　used　to　perform　partition　and　achieve　speedups.　Experimental　results　show　that　11　Olden　benchmarks　obtain　speedup　improvements.