代码调用流程:
1. nova.scheduler.client.query.SchedulerQueryClient#select_destinations 2. nova.scheduler.rpcapi.SchedulerAPI#select_destinations 3. nova.scheduler.manager.SchedulerManager#select_destinations 4. nova.scheduler.filter_scheduler.FilterScheduler#select_destinations
scheduler的rpcapi和manager属于同步调用。
在第三步中scheduler会调用placement提供的API,对所有的`compute node`进行初步的筛选,placement的API会返回一个字典,格式如下:
{ "provider_summaries": { "4cae2ef8-30eb-4571-80c3-3289e86bd65c": { "resources": { "VCPU": { "used": 2, "capacity": 64 }, "MEMORY_MB": { "used": 1024, "capacity": 11374 }, "DISK_GB": { "used": 2, "capacity": 49 } } } }, "allocation_requests": [ { "allocations": [ { "resource_provider": { "uuid": "4cae2ef8-30eb-4571-80c3-3289e86bd65c" }, "resources": { "VCPU": 1, "MEMORY_MB": 512, "DISK_GB": 1 } } ] } ] }
对于placement API筛选出的节点,scheduler会再度进行筛选,大概的筛选过程:all hosts => filtering => weighting => random
1. get all hosts:这里的all host当然不是指环境中所有的host,而是在通过placement API,返回的所有host的详细信息;
2. filtering:首先过滤ignore host和force host,如果force host或者force node直接返回即可。然后结合nova的配置文件中available_filters和enabled_filters参数,依次执行所有的filter。下面我们举几个filter的例子,执行filter的入口:
nova.filters.BaseFilterHandler#get_filtered_objects def get_filtered_objects(self, filters, objs, spec_obj, index=0): list_objs = list(objs) LOG.debug("Starting with %d host(s)", len(list_objs)) part_filter_results = [] full_filter_results = [] log_msg = "%(cls_name)s: (start: %(start)s, end: %(end)s)" # 循环遍历配置文件中指定的filters for filter_ in filters: if filter_.run_filter_for_index(index): cls_name = filter_.__class__.__name__ # 记录开始该执行filter之前的host的个数 start_count = len(list_objs) # 对所有的host执行该filter,返回只有经过该filter的host objs = filter_.filter_all(list_objs, spec_obj) if objs is None: LOG.debug("Filter %s says to stop filtering", cls_name) return list_objs = list(objs) end_count = len(list_objs) part_filter_results.append(log_msg % {"cls_name": cls_name, "start": start_count, "end": end_count}) if list_objs: remaining = [(getattr(obj, "host", obj), getattr(obj, "nodename", "")) for obj in list_objs] full_filter_results.append((cls_name, remaining)) else: LOG.info(_LI("Filter %s returned 0 hosts"), cls_name) full_filter_results.append((cls_name, None)) break LOG.debug("Filter %(cls_name)s returned " "%(obj_len)d host(s)", {'cls_name': cls_name, 'obj_len': len(list_objs)}) # 下边是一些日志中打印一些详细信息,不在赘述 ………… return list_objs
接下来介绍几个filter。
class AvailabilityZoneFilter(filters.BaseHostFilter): # 如果是一次创建多个虚机,则AvailabilityZoneFilter指执行一次 run_filter_once_per_request = True # 所有的filter都需要实现该方法 def host_passes(self, host_state, spec_obj): # 获取request_spec中指定的availability_zone,这里需要强调一下,如果创建时,没有指定--availability-zone 参数,request_sepc中的availability_zone就是空的。 availability_zone = spec_obj.availability_zone # 如果request_spec中availability_zone值为空,那么也就是这个操作是允许跨AZ操作的。 if not availability_zone: return True # 获取host的availability_zone信息,首先获取该host所属的aggregate信息,aggregate信息中有availability_zone相关的信息 metadata = utils.aggregate_metadata_get_by_host( host_state, key='availability_zone') if 'availability_zone' in metadata: # 判断request_spec中指定的availability_zone是否在该host所属的availability_zone中。 hosts_passes = availability_zone in metadata['availability_zone'] host_az = metadata['availability_zone'] else: hosts_passes = availability_zone == CONF.default_availability_zone host_az = CONF.default_availability_zone if not hosts_passes: LOG.debug("Availability Zone '%(az)s' requested. " "%(host_state)s has AZs: %(host_az)s", {'host_state': host_state, 'az': availability_zone, 'host_az': host_az}) return hosts_passes
nova.scheduler.filters.image_props_filter.ImagePropertiesFilter#host_passes # 主要是根据镜像中的property的值进行过滤,在ironic的调度中会使用到。 def host_passes(self, host_state, spec_obj): image_props = spec_obj.image.properties if spec_obj.image else {} # 判断该compute_node是否支持image的property属性中指定的参数值。 if not self._instance_supported(host_state, image_props, host_state.hypervisor_version): LOG.debug("%(host_state)s does not support requested " "instance_properties", {'host_state': host_state}) return False return True def _instance_supported(self, host_state, image_props, hypervisor_version): img_arch = image_props.get('hw_architecture') # 架构,i686或x86_64 img_h_type = image_props.get('img_hv_type') # hypervisor 类型 img_vm_mode = image_props.get('hw_vm_mode') # 虚拟化类型 ………… # 获取该compute_node支持的instance类型,返回值为列表。比如: [["x86_64", "baremetal", "hvm"]] [["i686", "qemu", "hvm"], ["i686", "kvm", "hvm"], ["x86_64", "qemu", "hvm"], ["x86_64", "kvm", "hvm"]] supp_instances = host_state.supported_instances ………… 比较规则 def _compare_props(props, other_props): # 对image的property指定的所有值进行遍历 for i in props: 查看该property是否是该compute_node支持的 if i and i not in other_props: return False return True # 对该compute_node支持的所有类型进行遍历 for supp_inst in supp_instances: if _compare_props(checked_img_props, supp_inst)
对于Ironic的调度需要我们着重使用到ImagePropertiesFilter,虚机使用的镜像和裸机使用的镜像中的property的值是不同的,再结合相关的placement的调度,实现虚机不会调度到ironic node,同时创建裸机不会调度到qemu的node。
3. 把过滤后的hosts计算权重并且进行最优排序,下面我们举几个weight的例子:
class BaseWeightHandler(loadables.BaseLoader): object_class = WeighedObject def get_weighed_objects(self, weighers, obj_list, weighing_properties): """Return a sorted (descending), normalized list of WeighedObjects.""" # obj_list 表示filter筛选出的所有hosts # weighing_properties 表示request_sepc信息 weighed_objs = [self.object_class(obj, 0.0) for obj in obj_list] # 如果经过filter筛选只剩一个host,则无需进行权重的比较,直接返回该host即可 if len(weighed_objs) <= 1: return weighed_objs # 根据配置文件中指定的weigher_classes,逐个计算权重 for weigher in weighers: # 以RAMWeigher为例进行说明 weights = weigher.weigh_objects(weighed_objs, weighing_properties) # Normalize the weights weights = normalize(weights, minval=weigher.minval, maxval=weigher.maxval) for i, weight in enumerate(weights): obj = weighed_objs[i] # 将计算后的权重值,保存到host信息中,并且将所有类型的权重加到一块,如果我们想要增加某种类型的权重比例,我们可以修改配置文件中*_weight_multiplier的值,比如我们想要在权重的计算中有关内存的权重占更大的作用,那么我们可以通过调节ram_weight_multiplier的值达到效果。 obj.weight += weigher.weight_multiplier() * weight # 按照权重进行性排序(倒序) return sorted(weighed_objs, key=lambda x: x.weight, reverse=True) class RAMWeigher(weights.BaseHostWeigher): minval = 0 def weight_multiplier(self): """Override the weight multiplier.""" return CONF.filter_scheduler.ram_weight_multiplier def _weigh_object(self, host_state, weight_properties): """Higher weights win. We want spreading to be the default.""" # 直接返回该节点的剩余内存,也就是剩余内存越多的节点,有关内存的权重越大。 return host_state.free_ram_mb
4. random,这个过程我们通过代码进行详细的分析。
host_subset_size = CONF.filter_scheduler.host_subset_size if host_subset_size < len(weighed_hosts): weighed_subset = weighed_hosts[0:host_subset_size] else: weighed_subset = weighed_hosts # 使用随机算法,从N个中抽取1个 chosen_host = random.choice(weighed_subset) weighed_hosts.remove(chosen_host) return [chosen_host] + weighed_hosts
对于host_subset_size参数,默认值为1。官方是这样解释的:如果设置大于1的正整数,当有多个scheduler进程处理相同的请求是会减少调度到同一台host的可能性,创造了一种竞争机制。从N个host中挑选最适合请求的一个host,会减少冲突。然而,如果该值设置的越大,对于给定的请求,选择的主机可能不太优化。