X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/0a365f3d1a19fc27b402c9c742debf8975424b3f..2dbbaaefc6a4a46a7f17b9e7799fc455cd722113:/services/nodemanager/arvnodeman/daemon.py diff --git a/services/nodemanager/arvnodeman/daemon.py b/services/nodemanager/arvnodeman/daemon.py index a65e9a0705..33b6cd58f6 100644 --- a/services/nodemanager/arvnodeman/daemon.py +++ b/services/nodemanager/arvnodeman/daemon.py @@ -103,13 +103,15 @@ class NodeManagerDaemonActor(actor_class): def __init__(self, server_wishlist_actor, arvados_nodes_actor, cloud_nodes_actor, cloud_update_actor, timer_actor, arvados_factory, cloud_factory, - shutdown_windows, min_size, min_nodes, max_nodes, + shutdown_windows, server_calculator, + min_nodes, max_nodes, poll_stale_after=600, boot_fail_after=1800, node_stale_after=7200, node_setup_class=dispatch.ComputeNodeSetupActor, node_shutdown_class=dispatch.ComputeNodeShutdownActor, - node_actor_class=dispatch.ComputeNodeMonitorActor): + node_actor_class=dispatch.ComputeNodeMonitorActor, + max_total_price=0): super(NodeManagerDaemonActor, self).__init__() self._node_setup = node_setup_class self._node_shutdown = node_shutdown_class @@ -119,12 +121,13 @@ class NodeManagerDaemonActor(actor_class): self._new_arvados = arvados_factory self._new_cloud = cloud_factory self._cloud_driver = self._new_cloud() - self._logger = logging.getLogger('arvnodeman.daemon') - self._later = self.actor_ref.proxy() + self._later = self.actor_ref.tell_proxy() self.shutdown_windows = shutdown_windows - self.min_cloud_size = min_size + self.server_calculator = server_calculator + self.min_cloud_size = self.server_calculator.cheapest_size() self.min_nodes = min_nodes self.max_nodes = max_nodes + self.max_total_price = max_total_price self.poll_stale_after = poll_stale_after self.boot_fail_after = boot_fail_after self.node_stale_after = node_stale_after @@ -139,14 +142,18 @@ class NodeManagerDaemonActor(actor_class): self.booting = {} # Actor IDs to ComputeNodeSetupActors self.booted = {} # Cloud node IDs to _ComputeNodeRecords self.shutdowns = {} # Cloud node IDs to ComputeNodeShutdownActors - self._logger.debug("Daemon initialized") + self.sizes_booting_shutdown = {} # Actor IDs or Cloud node IDs to node size + + def on_start(self): + self._logger = logging.getLogger("%s.%s" % (self.__class__.__name__, self.actor_urn[33:])) + self._logger.debug("Daemon started") def _update_poll_time(self, poll_key): self.last_polls[poll_key] = time.time() def _pair_nodes(self, node_record, arvados_node): - self._logger.info("Cloud node %s has associated with Arvados node %s", - node_record.cloud_node.id, arvados_node['uuid']) + self._logger.info("Cloud node %s is now paired with Arvados node %s", + node_record.cloud_node.name, arvados_node['uuid']) self._arvados_nodes_actor.subscribe_to( arvados_node['uuid'], node_record.actor.update_arvados_node) node_record.arvados_node = arvados_node @@ -167,11 +174,12 @@ class NodeManagerDaemonActor(actor_class): poll_stale_after=self.poll_stale_after, node_stale_after=self.node_stale_after, cloud_client=self._cloud_driver, - boot_fail_after=self.boot_fail_after).proxy() - actor.subscribe(self._later.node_can_shutdown) + boot_fail_after=self.boot_fail_after) + actorTell = actor.tell_proxy() + actorTell.subscribe(self._later.node_can_shutdown) self._cloud_nodes_actor.subscribe_to(cloud_node.id, - actor.update_cloud_node) - record = _ComputeNodeRecord(actor, cloud_node) + actorTell.update_cloud_node) + record = _ComputeNodeRecord(actor.proxy(), cloud_node) return record def update_cloud_nodes(self, nodelist): @@ -194,6 +202,7 @@ class NodeManagerDaemonActor(actor_class): except pykka.ActorDeadError: pass del self.shutdowns[key] + del self.sizes_booting_shutdown[key] record.actor.stop() record.cloud_node = None @@ -210,51 +219,117 @@ class NodeManagerDaemonActor(actor_class): self._pair_nodes(cloud_rec, arv_node) break - def _nodes_up(self): - return sum(len(nodelist) for nodelist in - [self.cloud_nodes, self.booted, self.booting]) - - def _nodes_busy(self): + def _nodes_booting(self, size): + s = sum(1 + for c in self.booting.iterkeys() + if size is None or self.sizes_booting_shutdown[c].id == size.id) + s += sum(1 + for c in self.booted.itervalues() + if size is None or c.cloud_node.size.id == size.id) + return s + + def _nodes_unpaired(self, size): + return sum(1 + for c in self.cloud_nodes.unpaired() + if size is None or c.cloud_node.size.id == size.id) + + def _nodes_booted(self, size): + return sum(1 + for c in self.cloud_nodes.nodes.itervalues() + if size is None or c.cloud_node.size.id == size.id) + + def _nodes_up(self, size): + up = self._nodes_booting(size) + self._nodes_booted(size) + return up + + def _total_price(self): + cost = 0 + cost += sum(self.server_calculator.find_size(self.sizes_booting_shutdown[c].id).price + for c in self.booting.iterkeys()) + cost += sum(self.server_calculator.find_size(c.cloud_node.size.id).price + for i in (self.booted, self.cloud_nodes.nodes) + for c in i.itervalues()) + return cost + + def _nodes_busy(self, size): return sum(1 for busy in pykka.get_all(rec.actor.in_state('busy') for rec in - self.cloud_nodes.nodes.itervalues()) + self.cloud_nodes.nodes.itervalues() + if rec.cloud_node.size.id == size.id) if busy) - def _nodes_missing(self): + def _nodes_missing(self, size): return sum(1 for arv_node in pykka.get_all(rec.actor.arvados_node for rec in self.cloud_nodes.nodes.itervalues() - if rec.actor.cloud_node.get().id not in self.shutdowns) + if rec.cloud_node.size.id == size.id and rec.actor.cloud_node.get().id not in self.shutdowns) if arv_node and cnode.arvados_node_missing(arv_node, self.node_stale_after)) - def _nodes_wanted(self): - up_count = self._nodes_up() - under_min = self.min_nodes - up_count - over_max = up_count - self.max_nodes + def _size_wishlist(self, size): + return sum(1 for c in self.last_wishlist if c.id == size.id) + + def _size_shutdowns(self, size): + sh = 0 + for c in self.shutdowns.iterkeys(): + try: + if self.sizes_booting_shutdown[c].id == size.id: + sh += 1 + except pykka.ActorDeadError: + pass + return sh + + def _nodes_wanted(self, size): + total_up_count = self._nodes_up(None) + under_min = self.min_nodes - total_up_count + over_max = total_up_count - self.max_nodes + total_price = self._total_price() + if over_max >= 0: return -over_max - elif under_min > 0: + elif under_min > 0 and size.id == self.min_cloud_size.id: return under_min + + booting_count = self._nodes_booting(size) + self._nodes_unpaired(size) + shutdown_count = self._size_shutdowns(size) + busy_count = self._nodes_busy(size) + up_count = self._nodes_up(size) - (shutdown_count + busy_count + self._nodes_missing(size)) + + self._logger.info("%s: wishlist %i, up %i (booting %i, idle %i, busy %i), shutting down %i", size.name, + self._size_wishlist(size), + up_count + busy_count, + booting_count, + up_count - booting_count, + busy_count, + shutdown_count) + + wanted = self._size_wishlist(size) - up_count + if wanted > 0 and self.max_total_price and ((total_price + (size.price*wanted)) > self.max_total_price): + can_boot = int((self.max_total_price - total_price) / size.price) + if can_boot == 0: + self._logger.info("Not booting %s (price %s) because with it would exceed max_total_price of %s (current total_price is %s)", + size.name, size.price, self.max_total_price, total_price) + return can_boot else: - up_count -= len(self.shutdowns) + self._nodes_busy() + self._nodes_missing() - return len(self.last_wishlist) - up_count - - def _nodes_excess(self): - up_count = self._nodes_up() - len(self.shutdowns) - over_min = up_count - self.min_nodes - if over_min <= 0: - return over_min - else: - return up_count - self._nodes_busy() - len(self.last_wishlist) + return wanted + + def _nodes_excess(self, size): + up_count = self._nodes_up(size) - self._size_shutdowns(size) + if size.id == self.min_cloud_size.id: + up_count -= self.min_nodes + return up_count - self._nodes_busy(size) - self._size_wishlist(size) def update_server_wishlist(self, wishlist): self._update_poll_time('server_wishlist') self.last_wishlist = wishlist - nodes_wanted = self._nodes_wanted() - if nodes_wanted > 0: - self._later.start_node() - elif (nodes_wanted < 0) and self.booting: - self._later.stop_booting_node() + for size in reversed(self.server_calculator.cloud_sizes): + try: + nodes_wanted = self._nodes_wanted(size) + if nodes_wanted > 0: + self._later.start_node(size) + elif (nodes_wanted < 0) and self.booting: + self._later.stop_booting_node(size) + except Exception as e: + self._logger.exception("while calculating nodes wanted for size %s", size) def _check_poll_freshness(orig_func): """Decorator to inhibit a method when poll information is stale. @@ -274,30 +349,28 @@ class NodeManagerDaemonActor(actor_class): return wrapper @_check_poll_freshness - def start_node(self): - nodes_wanted = self._nodes_wanted() + def start_node(self, cloud_size): + nodes_wanted = self._nodes_wanted(cloud_size) if nodes_wanted < 1: return None arvados_node = self.arvados_nodes.find_stale_node(self.node_stale_after) - try: - cloud_size = self.last_wishlist[self._nodes_up()] - except IndexError: - cloud_size = self.min_cloud_size - self._logger.info("Want %s more nodes. Booting a %s node.", + self._logger.info("Want %i more %s nodes. Booting a node.", nodes_wanted, cloud_size.name) new_setup = self._node_setup.start( timer_actor=self._timer, arvados_client=self._new_arvados(), arvados_node=arvados_node, cloud_client=self._new_cloud(), - cloud_size=cloud_size).proxy() + cloud_size=cloud_size).tell_proxy() self.booting[new_setup.actor_ref.actor_urn] = new_setup + self.sizes_booting_shutdown[new_setup.actor_ref.actor_urn] = cloud_size + if arvados_node is not None: self.arvados_nodes[arvados_node['uuid']].assignment_time = ( time.time()) new_setup.subscribe(self._later.node_up) if nodes_wanted > 1: - self._later.start_node() + self._later.start_node(cloud_size) def _get_actor_attrs(self, actor, *attr_names): return pykka.get_all([getattr(actor, name) for name in attr_names]) @@ -305,28 +378,34 @@ class NodeManagerDaemonActor(actor_class): def node_up(self, setup_proxy): cloud_node = setup_proxy.cloud_node.get() del self.booting[setup_proxy.actor_ref.actor_urn] + del self.sizes_booting_shutdown[setup_proxy.actor_ref.actor_urn] + setup_proxy.stop() - record = self.cloud_nodes.get(cloud_node.id) - if record is None: - record = self._new_node(cloud_node) - self.booted[cloud_node.id] = record - self._timer.schedule(time.time() + self.boot_fail_after, - self._later.shutdown_unpaired_node, cloud_node.id) + if cloud_node is not None: + record = self.cloud_nodes.get(cloud_node.id) + if record is None: + record = self._new_node(cloud_node) + self.booted[cloud_node.id] = record + self._timer.schedule(time.time() + self.boot_fail_after, + self._later.shutdown_unpaired_node, cloud_node.id) @_check_poll_freshness - def stop_booting_node(self): - nodes_excess = self._nodes_excess() + def stop_booting_node(self, size): + nodes_excess = self._nodes_excess(size) if (nodes_excess < 1) or not self.booting: return None for key, node in self.booting.iteritems(): - if node.stop_if_no_cloud_node().get(): + if node.cloud_size.get().id == size.id and node.stop_if_no_cloud_node().get(): del self.booting[key] + del self.sizes_booting_shutdown[key] + if nodes_excess > 1: - self._later.stop_booting_node() + self._later.stop_booting_node(size) break def _begin_node_shutdown(self, node_actor, cancellable): - cloud_node_id = node_actor.cloud_node.get().id + cloud_node_obj = node_actor.cloud_node.get() + cloud_node_id = cloud_node_obj.id if cloud_node_id in self.shutdowns: return None shutdown = self._node_shutdown.start( @@ -334,11 +413,12 @@ class NodeManagerDaemonActor(actor_class): arvados_client=self._new_arvados(), node_monitor=node_actor.actor_ref, cancellable=cancellable).proxy() self.shutdowns[cloud_node_id] = shutdown - shutdown.subscribe(self._later.node_finished_shutdown) + self.sizes_booting_shutdown[cloud_node_id] = cloud_node_obj.size + shutdown.tell_proxy().subscribe(self._later.node_finished_shutdown) @_check_poll_freshness def node_can_shutdown(self, node_actor): - if self._nodes_excess() > 0: + if self._nodes_excess(node_actor.cloud_node.get().size) > 0: self._begin_node_shutdown(node_actor, cancellable=True) def shutdown_unpaired_node(self, cloud_node_id): @@ -359,10 +439,10 @@ class NodeManagerDaemonActor(actor_class): if not success: if cancel_reason == self._node_shutdown.NODE_BROKEN: self.cloud_nodes.blacklist(cloud_node_id) - del self.shutdowns[cloud_node_id] elif cloud_node_id in self.booted: self.booted.pop(cloud_node_id).actor.stop() - del self.shutdowns[cloud_node_id] + del self.shutdowns[cloud_node_id] + del self.sizes_booting_shutdown[cloud_node_id] def shutdown(self): self._logger.info("Shutting down after signal.")