Improving the performance of speculatively parallel applications on the Hydra CMP