本文共 5078 字，大约阅读时间需要 16 分钟。

在实际应用中，为了实现基于利用率的调度（utilization-based scheduler）对Ceph OSD系统盘数量的优先调度，我们需要对Nova计算监控机制进行相应的扩展和配置。以下将详细介绍实现这一目标的具体步骤和技术细节。

首先，在配置文件中添加新的监控驱动。我们需要编辑Nova的配置文件nova-4.0rc6-py2.7.egg-info/entry_points.txt，并在nova.compute.monitors.osd对应的位置添加新的驱动实现。具体来说，我们需要在文件中添加如下内容：

[nova.compute.monitors.osd]virt_driver = nova.compute.monitors.osd.virt_driver:Monitor

接下来，我们需要将新的监控驱动注册到Nova的监控系统中。为此，我们需要编辑nova/compute/monitors/__init__.py，并在NAMESPACES列表中添加新的监控类型。具体修改如下：

NAMESPACES = [    'nova.compute.monitors.cpu',    'nova.compute.monitors.osd',  # 增加osd监控]

然后，我们需要实现新的监控功能。具体来说，我们需要在nova/compute/monitors/osd/virt_driver.py中定义新的监控类。以下是该文件的核心代码逻辑：

import pyudevfrom oslo_config import cfgfrom oslo_log import log as loggingfrom oslo_utils import timeutilsfrom nova.compute.monitors import basefrom nova import exceptionfrom nova.i18n import _LECONFLOG = logging.getLogger(__name__)class Monitor(base.OSDMonitorBase):    """CPU monitor that uses the virt driver's get_host_cpu_stats() call."""    def __init__(self, resource_tracker):        super(Monitor, self).__init__(resource_tracker)        self.source = cfg.CONF.compute_driver        self.driver = resource_tracker.driver        self._data = {}        self._cpu_stats = {}    def get_metric(self, name):        self._update_data()        return self._data[name], self._data["timestamp"]    def _update_data(self):        now = timeutils.utcnow()        if self._data.get("timestamp") is not None:            delta = now - self._data.get("timestamp")            if delta.seconds <= 1:                return self._data, now        # Extract node's CPU statistics.        try:            context = pyudev.Context()            pattern = re.compile('^/dev/vd[a-z]')            osd_devices = []            for device in context.list_devices(DEVTYPE='disk'):                major = device['MAJOR']                if major == '8' or re.search(pattern, device.device_node):                    osd_devices.append(device.device_node)            self._data["osd.num"] = len(osd_devices)        except (NotImplementedError, TypeError, KeyError):            LOG.exception(_LE("Not all properties needed are implemented"))            raise exception.ResourceMonitorError(                monitor=self.__class__.__name__            )

此外，我们还需要创建新的监控对象类。为此，我们需要编辑nova/objects/monitor_osd_metric.py，并定义新的对象类。以下是该文件的核心代码：

from oslo_serialization import jsonutilsfrom oslo_utils import timeutilsfrom nova.objects import basefrom nova.objects import fieldsfrom nova import utilsNOTE(jwcroppe): Used to determine which fields whose value we need to adjust(READ: divide by 100.0 before sending information to the RPC notifier sincethese values were expected to be within the range [0, 1])FIELDS_REQUIRING_CONVERSION = []@base.NovaObjectRegistry.registerclassclass OSDMonitorMetric(base.NovaObject):    """Version 1.0: Initial version    Version 1.1: Added Numa support    VERSION = '1.1'    fields = {        'name': fields.OSDMonitorMetricTypeField(nullable=False),        'value': fields.IntegerField(nullable=False),        'numa_membw_values': fields.DictOfIntegersField(nullable=True),        'timestamp': fields.DateTimeField(nullable=False),        # This will be the full class name for the plugin from which the metric originates        'source': fields.StringField(nullable=False),    }    def obj_make_compatible(self, primitive, target_version):        super(OSDMonitorMetric, self).obj_make_compatible(primitive, target_version)        target_version = utils.convert_version_to_tuple(target_version)        if target_version < (1, 1):            del primitive['numa_membw_values']    def to_dict(self):        dict_to_return = {            'name': self.name,            # NOTE(jaypipes): This is what jsonutils.dumps() does            # for datetime.datetime objects, which is what timestamp is in            # this object as well            'timestamp': timeutils.strtime(self.timestamp),            'source': self.source,        }        if self.obj_attr_is_set('value'):            if self.name in FIELDS_REQUIRING_CONVERSION:                dict_to_return['value'] = self.value / 100.0            else:                dict_to_return['value'] = self.value        elif self.obj_attr_is_set('numa_membw_values'):            dict_to_return['numa_membw_values'] = self.numa_membw_values        return dict_to_return    @classmethod    def from_json(cls, metrics):        """Converts a legacy json object into a list of MonitorMetric objs        and finally returns of MonitorMetricList"""        metrics = jsonutils.loads(metrics) if metrics else []        metric_list = [OSDMonitorMetric(**metric) for metric in metrics]        return OSDMonitorMetricList(objects=metric_list)

最后，我们需要定义新的监控指标类型字段。为此，我们需要编辑nova/objects/fields.py，并添加新的枚举类型。以下是该文件的核心代码：

class OSDMonitorMetricTypeField(BaseEnumField):    AUTO_TYPE = OSDMonitorMetricType()class OSDMonitorMetricType(Enum):    OSD_NUM = "osd.num"    ALL = (OSD_NUM,)

通过以上步骤，我们可以实现对Ceph OSD系统盘数量的基于利用率的调度优先级。需要注意的是，上述实现仅为演示 purposes，实际应用中需要根据具体需求进行调整和优化。此外，在实际操作中，建议参考相关文档和官方指南，确保所有配置和修改都符合Nova和Ceph的官方规范。

转载地址：http://tljfk.baihongyu.com/