Increasing Concurrency in Deterministic Runtimes with Conversion

Experimental results are presented for several benchmark programs, identifying quantum size imbalance as a major source of inefficiency in Dthreads. A two-pronged approach is proposed to address this problem. First, Dthreads is ported to a versioned memory subsystem, so that the fence may be removed. Second, we find that Dthreads’ roundrobin token order is more rigid than necessary for the revised memory model. We propose a more efficient, yet determinism-preserving order based on approximation of program execution time using a determinstic clock.