The runtime implementation of "Scheduling I/O Latency-Hiding Futures in Task-Parallel Platforms" by Kyle Singer, Kunal Agrawal, and I-Ting Angelina Lee (https://doi.org/10.1137/1.9781611976021.11).

In this version of Cilk-L, the default configuration is that I/O threads and work threads are bound to the same logical CPU core. In the corresponding paper, the configuration is such that I/O threads and work threads are bound to the same physical CPU, but such that the work thread is on one hyperthread (logical CPU on the same physical core) and the I/O thread is on the other hyperthread. The configuration can be changed by modifying the g->io_mode variable around line 430 of cilk-l/runtime/global_state.cpp

It is also important to note that because the Cilk ABI has been modified to support futures, the entirety of this repo must be built using the provided custom clang implementation that, when given the -fcilk-no-inline, will not inline ABI functions and will instead call the functions linked in.

In order to build and run the code, first set up the Docker container system. Once that has completed, run ./build_docker.sh followed by ./start_docker.sh. This will put you in an interactive Docker container shell. For convenience, the ./setup.sh script can now be used to build everything but the benchmarks. Once setup.sh has finished building the compiler and the runtime, the benchmarks can be built from within the Docker container with the following commands: cd bench; make