photoncloud-monorepo/nix/test-cluster/common.nix

# PhotonCloud 6-Node Test Cluster
#
# Common configuration shared by all nodes
#
# Usage: Import this from individual node configurations

{ config, lib, pkgs, modulesPath, ... }:

let
  cfg = config.photonTestCluster;
in
{
  imports = [
    (modulesPath + "/virtualisation/qemu-vm.nix")
    ../modules/plasmacloud-cluster.nix
  ];

  options.photonTestCluster = {
    sshBasePort = lib.mkOption {
      type = lib.types.port;
      default = 2200;
      description = "Base host port used for guest SSH forwarding.";
    };

    vdeSock = lib.mkOption {
      type = lib.types.str;
      default = "/tmp/photoncloud-test-cluster-vde.sock";
      description = "VDE control socket path used for the east-west cluster NIC.";
    };

    chainfireControlPlaneAddrs = lib.mkOption {
      type = lib.types.str;
      default = "10.100.0.11:2379,10.100.0.12:2379,10.100.0.13:2379";
      description = "Comma-separated ChainFire client endpoints for multi-endpoint failover.";
    };

    flaredbControlPlaneAddrs = lib.mkOption {
      type = lib.types.str;
      default = "10.100.0.11:2479,10.100.0.12:2479,10.100.0.13:2479";
      description = "Comma-separated FlareDB client endpoints for multi-endpoint failover.";
    };
  };

  config = {
    virtualisation = let
      # Extract node index (e.g., "node01" -> "1")
      nodeIndex = lib.strings.toInt (lib.strings.removePrefix "node0" config.networking.hostName);
      macSuffix = lib.strings.fixedWidthString 2 "0" (toString nodeIndex);
      vdeSock = cfg.vdeSock;
    in {
      graphics = false;
      cores = 2;
      forwardPorts =
        [
          { from = "host"; host.port = cfg.sshBasePort + nodeIndex; guest.port = 22; }
        ]
        ++ lib.optionals (config.networking.hostName == "node06") [
          { from = "host"; host.port = 8080; guest.port = 8080; }
          { from = "host"; host.port = 8443; guest.port = 8443; }
          { from = "host"; host.port = 9090; guest.port = 9090; }
          { from = "host"; host.port = 3000; guest.port = 3000; }
        ];
      qemu.options = [
        # Nested KVM validation requires hardware acceleration and host CPU flags.
        "-enable-kvm"
        "-cpu host"
        # eth1: Cluster network shared across all VMs. VDE is materially faster
        # than multicast sockets for this nested-QEMU storage lab.
        "-netdev vde,id=n1,sock=${vdeSock}"
        "-device virtio-net-pci,netdev=n1,mac=52:54:00:10:00:${macSuffix}"
      ];
    };

    networking.firewall.enable = false;
    services.openssh = {
      enable = true;
      settings = {
        KbdInteractiveAuthentication = false;
        PasswordAuthentication = true;
        PermitRootLogin = "yes";
      };
    };
    users.mutableUsers = false;
    users.users.root.hashedPassword = "$6$photoncloud$aUJCEE5wm/b5O.9KIKGm84qUWdWXwnebsFEiMBF7u9Y7AOWodaMrjbbKGMOf0X59VJyJeMRsgbT7VWeqMHpUe.";

    # qemu-vm.nix provides the default SLiRP NIC as eth0.
    # The extra multicast NIC above becomes eth1 and carries intra-cluster traffic.
    networking.interfaces.eth0.useDHCP = true;

    boot.loader.grub.device = "nodev";
    boot.kernelModules = [ "nbd" ];
    boot.extraModprobeConfig = ''
      options nbd nbds_max=16 max_part=8
    '';
    fileSystems."/" = { device = "/dev/disk/by-label/nixos"; fsType = "ext4"; };

    system.stateVersion = "24.05";

    systemd.services.photon-test-cluster-net-tuning = {
      description = "Tune cluster NIC offloads for nested-QEMU storage tests";
      wantedBy = [ "multi-user.target" ];
      after = [ "network-online.target" ];
      wants = [ "network-online.target" ];
      serviceConfig = {
        Type = "oneshot";
        RemainAfterExit = true;
      };
      path = [ pkgs.ethtool pkgs.iproute2 pkgs.coreutils ];
      script = ''
        set -eu
        iface="eth1"
        for _ in $(seq 1 30); do
          if ip link show "$iface" >/dev/null 2>&1; then
            break
          fi
          sleep 1
        done
        if ! ip link show "$iface" >/dev/null 2>&1; then
          echo "photon-test-cluster-net-tuning: $iface not present, skipping" >&2
          exit 0
        fi

        # Nested QEMU over VDE is sensitive to guest-side offloads; disabling
        # them reduces retransmits and keeps the storage benchmarks closer to
        # raw TCP throughput.
        ethtool -K "$iface" tso off gso off gro off tx off rx off sg off || true
        ip link set dev "$iface" txqueuelen 10000 || true
      '';
    };

    environment.systemPackages = with pkgs; [
      awscli2
      curl
      dnsutils
      ethtool
      fio
      jq
      grpcurl
      htop
      iperf3
      (python3.withPackages (ps: [ ps.boto3 ]))
      vim
      netcat
      iproute2
      tcpdump
      pciutils # lspci for debugging
      qemu
    ];

    plasmacloud.cluster = {
      enable = true;
      name = "photoncloud-test";
      
      nodes = {
        node01 = {
          role = "control-plane";
          ip = "10.100.0.11";
          services = [ "chainfire" "flaredb" "iam" "prismnet" "flashdns" "fiberlb" "k8shost" "plasmavmc" "lightningstor" "coronafs" ];
          labels = {
            tier = "control-plane";
          };
          pool = "control";
          nodeClass = "control-plane";
          failureDomain = "zone-a";
          raftPort = 2380;
          apiPort = 2379;
        };
        node02 = {
          role = "control-plane";
          ip = "10.100.0.12";
          services = [ "chainfire" "flaredb" "iam" ];
          labels = {
            tier = "control-plane";
          };
          pool = "control";
          nodeClass = "control-plane";
          failureDomain = "zone-b";
          raftPort = 2380;
          apiPort = 2379;
        };
        node03 = {
          role = "control-plane";
          ip = "10.100.0.13";
          services = [ "chainfire" "flaredb" "iam" ];
          labels = {
            tier = "control-plane";
          };
          pool = "control";
          nodeClass = "control-plane";
          failureDomain = "zone-c";
          raftPort = 2380;
          apiPort = 2379;
        };
        node04 = {
          role = "worker";
          ip = "10.100.0.21";
          services = [ "plasmavmc-agent" "lightningstor-data" "node-agent" ];
          labels = {
            runtime = "native";
          };
          pool = "general";
          nodeClass = "worker-linux";
          failureDomain = "zone-b";
          state = "provisioning";
          raftPort = 2380;
          apiPort = 2379;
        };
        node05 = {
          role = "worker";
          ip = "10.100.0.22";
          services = [ "plasmavmc-agent" "lightningstor-data" "node-agent" ];
          labels = {
            runtime = "native";
          };
          pool = "general";
          nodeClass = "worker-linux";
          failureDomain = "zone-c";
          state = "provisioning";
          raftPort = 2380;
          apiPort = 2379;
        };
        node06 = {
          role = "control-plane";
          ip = "10.100.0.100";
          services = [ "apigateway" "nightlight" "creditservice" "deployer" "fleet-scheduler" ];
          labels = {
            tier = "control-plane";
            ingress = "true";
          };
          pool = "control";
          nodeClass = "control-plane";
          failureDomain = "zone-a";
          raftPort = 2380;
          apiPort = 2379;
        };
      };

      deployer = {
        clusterId = "test-cluster";
        environment = "test";

        nodeClasses = {
          control-plane = {
            description = "Control-plane services and management endpoints";
            roles = [ "control-plane" ];
            labels = {
              tier = "control-plane";
            };
          };

          worker-linux = {
            description = "General-purpose native runtime workers";
            roles = [ "worker" ];
            labels = {
              tier = "general";
              runtime = "native";
            };
          };
        };

        pools = {
          control = {
            description = "Control-plane pool";
            nodeClass = "control-plane";
            labels = {
              plane = "control";
            };
          };

          general = {
            description = "General-purpose native worker pool";
            nodeClass = "worker-linux";
            labels = {
              "pool.photoncloud.io/name" = "general";
            };
          };
        };

        services = {
          native-web = {
            protocol = "http";
            ports.http = 18190;
            schedule = {
              replicas = 2;
              placement = {
                roles = [ "worker" ];
                pools = [ "general" ];
                nodeClasses = [ "worker-linux" ];
                matchLabels = {
                  runtime = "native";
                };
                spreadByLabel = "failure_domain";
                maxInstancesPerNode = 1;
              };
              instancePort = 18190;
              process = {
                command = "python3";
                args = [
                  "-m"
                  "http.server"
                  "\${INSTANCE_PORT}"
                  "--bind"
                  "\${INSTANCE_IP}"
                ];
              };
              healthCheck = {
                type = "http";
                path = "/";
                intervalSecs = 5;
                timeoutSecs = 3;
              };
            };
            publish = {
              dns = {
                zone = "native.cluster.test";
                name = "web";
                ttl = 30;
                mode = "load_balancer";
              };
              loadBalancer = {
                orgId = "native-services";
                projectId = "test-cluster";
                listenerPort = 18191;
                protocol = "http";
                poolProtocol = "http";
              };
            };
          };

          native-container = {
            protocol = "http";
            ports.http = 18192;
            schedule = {
              replicas = 1;
              placement = {
                roles = [ "worker" ];
                pools = [ "general" ];
                nodeClasses = [ "worker-linux" ];
                matchLabels = {
                  runtime = "native";
                };
                maxInstancesPerNode = 1;
              };
              instancePort = 18192;
              container = {
                image = "docker.io/library/nginx:1.27-alpine";
                runtime = "podman";
                pullPolicy = "if-not-present";
                ports = [
                  {
                    containerPort = 80;
                    hostPort = 18192;
                    protocol = "tcp";
                  }
                ];
              };
              healthCheck = {
                type = "http";
                path = "/";
                intervalSecs = 5;
                timeoutSecs = 5;
                startupGraceSecs = 120;
              };
            };
          };
        };
      };
      
      bootstrap.initialPeers = [ "node01" "node02" "node03" ];
      bgp.asn = 64512;
    };
  };
}