Overview
A bare-metal compute service makes it possible to run tasks directly on hardware resources. Think of it as a set of multi-core hosts on which multi-threaded processes can be started using something like Ssh. The service does not perform any space-sharing among the jobs. In other words, jobs submitted to the service execute concurrently in a time-shared manner. It is the responsibility of the job submitter to pick hosts and/or numbers of cores for each task, e.g., to enforce space-sharing of cores. But by default the compute service operates as a "jungle" in which tasks share cores at will. The only resource allocation performed by the service is that it ensures that the RAM capacity of a hosts are not exceeded. Tasks that have non-zero RAM requirements are queued in FCFS fashion at each host until there is enough RAM to execute them (think of this as each host running an OS that disallows swapping and implements a FCFS access policy for RAM allocation).
Creating a bare-metal compute service
In WRENCH, a bare-metal service is defined in the wrench::BareMetalComputeService class. An instantiation of a bare-metal service requires the following parameters:
- The name of a host on which to start the service;
- A set of compute hosts in a map (
std::map), where each key is a hostname and each value is a tuple (std::tuple) with a number of cores and a RAM capacity; - A mount point (corresponding to a disk attached to the host) for the scratch space, i.e., storage local to the bare-metal service (used to store workflow files, as needed, during job executions); and
- Maps (
std::map) of configurable properties (wrench::BareMetalComputeServiceProperty) and configurable message payloads (wrench::BareMetalComputeServiceMessagePayload).
The example below creates an instance of a bare-metal service that runs on host Gateway, provides access to all cores and 1GiB of RAM on host Node1 and to 8 cores and all RAM on host Node2, and has a scratch space on the disk mounted at path /scratch on host Gateway. Furthermore, the thread startup overhead is configured to be one hundredth of a second, and the message with which the service answers resource request description requests is configured to be 1KiB:
See the documentation of wrench::BareMetalComputeServiceProperty and wrench::BareMetalComputeServiceMessagePayload for all possible configuration options.
Also see the simulators in the examples/basic-examples/bare-metal-*/ directories, which use bare-metal compute services.