# SPDX-License-Identifier: MIT # Copyright © 2024 Intel Corporation import logging from enum import Enum from pathlib import Path from bench import exceptions from bench.configurators.pci import GpuModel, get_vgpu_profiles_file from bench.configurators.vgpu_profile import (VgpuProfile, VgpuProfilesDefinitions, VgpuProfilesJsonReader, VgpuResourcesConfig, VgpuSchedulerConfig, VgpuSecurityConfig) logger = logging.getLogger('DeviceConfigurator') class VfSchedulingMode(str, Enum): INFINITE = 'Infinite' # Infinite EQ/PT - HW default DEFAULT_PROFILE = 'Default_Profile' # Default vGPU scheduler profile FLEXIBLE_30FPS = 'Flexible_30fps_GPUTimeSlicing' FIXED_30FPS = 'Fixed_30fps_GPUTimeSlicing' FLEXIBLE_BURSTABLE_QOS = 'Flexible_BurstableQoS_GPUTimeSlicing' def __str__(self) -> str: return str.__str__(self) class VgpuProfileConfigurator: def __init__(self, vgpu_profiles_dir: Path, gpu_model: GpuModel = GpuModel.Unknown) -> None: self.gpu_model: GpuModel = gpu_model self.vgpu_profiles_dir: Path = vgpu_profiles_dir self.supported_vgpu_profiles: VgpuProfilesDefinitions = self.query_vgpu_profiles() def __helper_create_vgpu_json_path(self, vgpu_resource_dir: Path) -> Path: vgpu_device_file = get_vgpu_profiles_file(self.gpu_model) vgpu_json_file_path = vgpu_resource_dir / vgpu_device_file if not vgpu_json_file_path.exists(): logger.error("vGPU profiles JSON file not found in %s", vgpu_resource_dir) raise exceptions.VgpuProfileError(f'vGPU profiles JSON file not found in {vgpu_resource_dir}') return vgpu_json_file_path def query_vgpu_profiles(self) -> VgpuProfilesDefinitions: """Get all vGPU profiles supported for a given GPU device.""" json_reader = VgpuProfilesJsonReader(self.__helper_create_vgpu_json_path(self.vgpu_profiles_dir)) return json_reader.vgpu_profiles def select_vgpu_resources_profile(self, requested_num_vfs: int) -> VgpuResourcesConfig: """Find vGPU profile matching requested number of VFs. In case exact match cannot be found, try to fit similar profile with up to 2 more VFs, for example: - if requested profile with 3 VFs is not available, return close config with 4 VFs. - if requested profile with neither 9 VFs, nor with 10 or 11 VFs is available - throw 'not found' exeception. """ vgpu_resources_config = VgpuResourcesConfig() for pf_resource in self.supported_vgpu_profiles.pf_resources: if pf_resource.profile_name == self.supported_vgpu_profiles.pf_resource_default: vgpu_resources_config.pfLmem = pf_resource.local_memory_ecc_on vgpu_resources_config.pfContexts = pf_resource.contexts vgpu_resources_config.pfDoorbells = pf_resource.doorbells vgpu_resources_config.pfGgtt = pf_resource.ggtt_size is_vf_resource_found = False for vf_resource in self.supported_vgpu_profiles.vf_resources: current_num_vfs = vf_resource.vf_count if current_num_vfs == requested_num_vfs: is_vf_resource_found = True # Exact match elif requested_num_vfs < current_num_vfs <= requested_num_vfs + 2: logger.debug("Unable to find accurate vGPU profile but have similar: %s", vf_resource.profile_name) is_vf_resource_found = True # Approximate match if is_vf_resource_found: vgpu_resources_config.vfLmem = vf_resource.local_memory_ecc_on vgpu_resources_config.vfContexts = vf_resource.contexts vgpu_resources_config.vfDoorbells = vf_resource.doorbells vgpu_resources_config.vfGgtt = vf_resource.ggtt_size break if not is_vf_resource_found: logger.error("vGPU VF resources profile %sxVF not found!", requested_num_vfs) raise exceptions.VgpuProfileError(f'vGPU VF resources profile {requested_num_vfs}xVF not found!') return vgpu_resources_config def select_vgpu_scheduler_profile(self, requested_num_vfs: int, requested_scheduler: VfSchedulingMode) -> VgpuSchedulerConfig: # Function eval is needed to calculate VF EQ/PT for num_vfs # Disable eval warning # pylint: disable=W0123 vgpu_scheduler_config = VgpuSchedulerConfig() if requested_scheduler is VfSchedulingMode.INFINITE: return vgpu_scheduler_config for scheduler in self.supported_vgpu_profiles.scheduler_configs: if scheduler.profile_name == requested_scheduler: vgpu_scheduler_config.scheduleIfIdle = scheduler.schedule_if_idle vgpu_scheduler_config.pfExecutionQuanta = scheduler.pf_execution_quanta vgpu_scheduler_config.pfPreemptionTimeout = scheduler.pf_preemption_timeout lambda_vf_eq = eval(scheduler.vf_execution_quanta) lambda_vf_eq_result = lambda_vf_eq(requested_num_vfs) lambda_vf_pt = eval(scheduler.vf_preemption_timeout) lambda_vf_pt_result = lambda_vf_pt(requested_num_vfs) vgpu_scheduler_config.vfExecutionQuanta = lambda_vf_eq_result vgpu_scheduler_config.vfPreemptionTimeout = lambda_vf_pt_result return vgpu_scheduler_config def select_vgpu_security_profile(self) -> VgpuSecurityConfig: # Currently supports only default security profile vgpu_security_config = VgpuSecurityConfig() for security_profile in self.supported_vgpu_profiles.security_configs: if security_profile.profile_name == self.supported_vgpu_profiles.security_config_default: vgpu_security_config.reset_after_vf_switch = security_profile.reset_after_vf_switch vgpu_security_config.guc_sampling_period = security_profile.guc_sampling_period vgpu_security_config.guc_threshold_cat_error = security_profile.guc_threshold_cat_error vgpu_security_config.guc_threshold_page_fault = security_profile.guc_threshold_page_fault vgpu_security_config.guc_threshold_h2g_storm = security_profile.guc_threshold_h2g_storm vgpu_security_config.guc_threshold_db_storm = security_profile.guc_threshold_db_storm vgpu_security_config.guc_treshold_gt_irq_storm = security_profile.guc_treshold_gt_irq_storm vgpu_security_config.guc_threshold_engine_reset = security_profile.guc_threshold_engine_reset return vgpu_security_config def get_vgpu_profile(self, requested_num_vfs: int, requested_scheduler: VfSchedulingMode) -> VgpuProfile: """Get vGPU profile for requested number of VFs, scheduler and security modes.""" logger.info("Requested vGPU profile: %s VFs / scheduling: %s", requested_num_vfs, requested_scheduler) vgpu_profile: VgpuProfile = VgpuProfile() vgpu_profile.num_vfs = requested_num_vfs vgpu_profile.resources = self.select_vgpu_resources_profile(requested_num_vfs) if requested_scheduler is VfSchedulingMode.DEFAULT_PROFILE: requested_scheduler = VfSchedulingMode(self.supported_vgpu_profiles.scheduler_config_default) vgpu_profile.scheduler = self.select_vgpu_scheduler_profile(requested_num_vfs, requested_scheduler) vgpu_profile.security = self.select_vgpu_security_profile() return vgpu_profile