rrun.hpp

 
#pragma once
 
#include <optional>
#include <string>
#include <vector>
 
namespace cucascade::memory {
struct system_topology_info;
}  // namespace cucascade::memory
 
namespace rapidsmpf::rrun {
 
struct bind_options {
    bool cpu{true};  
    bool memory{true};  
    bool network{true};  
    bool verify{true};  
};
 
struct resource_binding {
    int rank = -1;  
    int gpu_id = -1;  
    std::string gpu_pci_bus_id;  
    std::string cpu_affinity;  
    std::vector<int> numa_nodes;  
    std::string ucx_net_devices;  
};
 
struct expected_binding {
    std::string cpu_affinity;  
    std::vector<int> memory_binding;  
    std::vector<std::string> network_devices;  
};
 
struct binding_validation {
    bool cpu_ok = true;  
    bool numa_ok = true;  
    bool ucx_ok = true;  
    std::string expected_ucx_devices;  
 
    [[nodiscard]] bool all_passed() const {
        return cpu_ok && numa_ok && ucx_ok;
    }
};
 
resource_binding check_binding(int gpu_id_hint = -1);
 
std::optional<expected_binding> get_expected_binding(
    cucascade::memory::system_topology_info const& topology, int gpu_id
);
 
binding_validation validate_binding(
    resource_binding const& actual, expected_binding const& expected
);
 
void bind(
    std::optional<unsigned int> gpu_id = std::nullopt, bind_options const& options = {}
);
 
void bind(
    cucascade::memory::system_topology_info const& topology,
    std::optional<unsigned int> gpu_id = std::nullopt,
    bind_options const& options = {}
);
 
}  // namespace rapidsmpf::rrun