diff options
Diffstat (limited to 'node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodePrioritizer.java')
| -rw-r--r-- | node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodePrioritizer.java | 96 |
1 files changed, 69 insertions, 27 deletions
diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodePrioritizer.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodePrioritizer.java index 960d0b9d729..3df79174b5c 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodePrioritizer.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodePrioritizer.java @@ -30,7 +30,7 @@ import java.util.stream.Collectors; * * @author smorgrav */ -public class NodePrioritizer { +class NodePrioritizer { private final Map<Node, PrioritizableNode> nodes = new HashMap<>(); private final List<Node> allNodes; @@ -39,10 +39,10 @@ public class NodePrioritizer { private final ApplicationId appId; private final ClusterSpec clusterSpec; + private final boolean isDocker; private final boolean isAllocatingForReplacement; private final Set<Node> spareHosts; - private final Map<Node, Boolean> headroomHosts; - private final boolean isDocker; + private final Map<Node, ResourceCapacity> headroomHosts; NodePrioritizer(List<Node> allNodes, ApplicationId appId, ClusterSpec clusterSpec, NodeSpec nodeSpec, NodeFlavors nodeFlavors, int spares) { this.allNodes = Collections.unmodifiableList(allNodes); @@ -50,8 +50,8 @@ public class NodePrioritizer { this.clusterSpec = clusterSpec; this.appId = appId; - spareHosts = findSpareHosts(allNodes, spares); - headroomHosts = findHeadroomHosts(allNodes, spareHosts, nodeFlavors); + this.spareHosts = findSpareHosts(allNodes, spares); + this.headroomHosts = findHeadroomHosts(allNodes, spareHosts, nodeFlavors); this.capacity = new DockerHostCapacity(allNodes); @@ -68,14 +68,14 @@ public class NodePrioritizer { .filter(node -> node.allocation().get().membership().cluster().id().equals(clusterSpec.id())) .count(); - isAllocatingForReplacement = isReplacement(nofNodesInCluster, nofFailedNodes); - isDocker = isDocker(); + this.isAllocatingForReplacement = isReplacement(nofNodesInCluster, nofFailedNodes); + this.isDocker = isDocker(); } /** * From ipAddress - get hostname * - * @return hostname or null if not able to do the loopup + * @return hostname or null if not able to do the lookup */ private static String lookupHostname(String ipAddress) { try { @@ -104,14 +104,14 @@ public class NodePrioritizer { } /** - * Headroom are the nodes with the least but sufficient space for the requested headroom. + * Headroom hosts are the host with the least but sufficient capacity for the requested headroom. * - * If not enough headroom - the headroom violating hosts are the once that are closest to fulfull + * If not enough headroom - the headroom violating hosts are the once that are closest to fulfill * a headroom request. */ - private static Map<Node, Boolean> findHeadroomHosts(List<Node> nodes, Set<Node> spareNodes, NodeFlavors flavors) { + private static Map<Node, ResourceCapacity> findHeadroomHosts(List<Node> nodes, Set<Node> spareNodes, NodeFlavors flavors) { DockerHostCapacity capacity = new DockerHostCapacity(nodes); - Map<Node, Boolean> headroomNodesToViolation = new HashMap<>(); + Map<Node, ResourceCapacity> headroomHosts = new HashMap<>(); List<Node> hostsSortedOnLeastCapacity = nodes.stream() .filter(n -> !spareNodes.contains(n)) @@ -121,20 +121,25 @@ public class NodePrioritizer { .sorted((a, b) -> capacity.compareWithoutInactive(b, a)) .collect(Collectors.toList()); + // For all flavors with ideal headroom - find which hosts this headroom should be allocated to for (Flavor flavor : flavors.getFlavors().stream().filter(f -> f.getIdealHeadroom() > 0).collect(Collectors.toList())) { Set<Node> tempHeadroom = new HashSet<>(); Set<Node> notEnoughCapacity = new HashSet<>(); + + ResourceCapacity headroomCapacity = ResourceCapacity.of(flavor); + + // Select hosts that has available capacity for both headroom and for new allocations for (Node host : hostsSortedOnLeastCapacity) { - if (headroomNodesToViolation.containsKey(host)) continue; - if (capacity.hasCapacityWhenRetiredAndInactiveNodesAreGone(host, flavor)) { - headroomNodesToViolation.put(host, false); + if (headroomHosts.containsKey(host)) continue; + if (capacity.hasCapacityWhenRetiredAndInactiveNodesAreGone(host, headroomCapacity)) { + headroomHosts.put(host, headroomCapacity); tempHeadroom.add(host); } else { notEnoughCapacity.add(host); } if (tempHeadroom.size() == flavor.getIdealHeadroom()) { - continue; + break; } } @@ -145,14 +150,13 @@ public class NodePrioritizer { .limit(flavor.getIdealHeadroom() - tempHeadroom.size()) .collect(Collectors.toList()); - for (Node nodeViolatingHeadrom : violations) { - headroomNodesToViolation.put(nodeViolatingHeadrom, true); + for (Node hostViolatingHeadrom : violations) { + headroomHosts.put(hostViolatingHeadrom, headroomCapacity); } - } } - return headroomNodesToViolation; + return headroomHosts; } /** @@ -197,14 +201,14 @@ public class NodePrioritizer { } } - if (!conflictingCluster && capacity.hasCapacity(node, getFlavor())) { + if (!conflictingCluster && capacity.hasCapacity(node, ResourceCapacity.of(getFlavor(requestedNodes)))) { Set<String> ipAddresses = DockerHostCapacity.findFreeIps(node, allNodes); if (ipAddresses.isEmpty()) continue; String ipAddress = ipAddresses.stream().findFirst().get(); String hostname = lookupHostname(ipAddress); if (hostname == null) continue; Node newNode = Node.createDockerNode("fake-" + hostname, Collections.singleton(ipAddress), - Collections.emptySet(), hostname, Optional.of(node.hostname()), getFlavor(), NodeType.tenant); + Collections.emptySet(), hostname, Optional.of(node.hostname()), getFlavor(requestedNodes), NodeType.tenant); PrioritizableNode nodePri = toNodePriority(newNode, false, true); if (!nodePri.violatesSpares || isAllocatingForReplacement) { nodes.put(newNode, nodePri); @@ -249,7 +253,7 @@ public class NodePrioritizer { pri.node = node; pri.isSurplusNode = isSurplusNode; pri.isNewNode = isNewNode; - pri.preferredOnFlavor = requestedNodes.specifiesNonStockFlavor() && node.flavor().equals(getFlavor()); + pri.preferredOnFlavor = requestedNodes.specifiesNonStockFlavor() && node.flavor().equals(getFlavor(requestedNodes)); pri.parent = findParentNode(node); if (pri.parent.isPresent()) { @@ -260,14 +264,29 @@ public class NodePrioritizer { pri.violatesSpares = true; } - if (headroomHosts.containsKey(parent)) { - pri.violatesHeadroom = headroomHosts.get(parent); + if (headroomHosts.containsKey(parent) && isPreferredNodeToBeReloacted(allNodes, node, parent)) { + ResourceCapacity neededCapacity = headroomHosts.get(parent); + + // If the node is new then we need to check the headroom requirement after it has been added + if (isNewNode) { + neededCapacity = ResourceCapacity.composite(neededCapacity, new ResourceCapacity(node)); + } + pri.violatesHeadroom = !capacity.hasCapacity(parent, neededCapacity); } } return pri; } + static boolean isPreferredNodeToBeReloacted(List<Node> nodes, Node node, Node parent) { + NodeList list = new NodeList(nodes); + return list.childNodes(parent).asList().stream() + .sorted(NodePrioritizer::compareForRelocation) + .findFirst() + .filter(n -> n.equals(node)) + .isPresent(); + } + private boolean isReplacement(long nofNodesInCluster, long nodeFailedNodes) { if (nodeFailedNodes == 0) return false; @@ -280,7 +299,7 @@ public class NodePrioritizer { return (wantedCount > nofNodesInCluster - nodeFailedNodes); } - private Flavor getFlavor() { + private static Flavor getFlavor(NodeSpec requestedNodes) { if (requestedNodes instanceof NodeSpec.CountNodeSpec) { NodeSpec.CountNodeSpec countSpec = (NodeSpec.CountNodeSpec) requestedNodes; return countSpec.getFlavor(); @@ -289,7 +308,7 @@ public class NodePrioritizer { } private boolean isDocker() { - Flavor flavor = getFlavor(); + Flavor flavor = getFlavor(requestedNodes); return (flavor != null) && flavor.getType().equals(Flavor.Type.DOCKER_CONTAINER); } @@ -299,4 +318,27 @@ public class NodePrioritizer { .filter(n -> n.hostname().equals(node.parentHostname().orElse(" NOT A NODE"))) .findAny(); } + + private static int compareForRelocation(Node a, Node b) { + // Choose smallest node + int capacity = ResourceCapacity.of(a).compare(ResourceCapacity.of(b)); + if (capacity != 0) return capacity; + + // Choose unallocated over allocated (this case is when we have ready docker nodes) + if (!a.allocation().isPresent() && b.allocation().isPresent()) return -1; + if (a.allocation().isPresent() && !b.allocation().isPresent()) return 1; + + // Choose container over content nodes + if (a.allocation().isPresent() && b.allocation().isPresent()) { + if (a.allocation().get().membership().cluster().type().equals(ClusterSpec.Type.container) && + !b.allocation().get().membership().cluster().type().equals(ClusterSpec.Type.container)) + return -1; + if (!a.allocation().get().membership().cluster().type().equals(ClusterSpec.Type.container) && + b.allocation().get().membership().cluster().type().equals(ClusterSpec.Type.container)) + return 1; + } + + // To get a stable algorithm - choose lexicographical from hostname + return a.hostname().compareTo(b.hostname()); + } } |