V5.2.2 - hostnames, Failed to initialize NVML fix for ubuntu 22 hosts

clore_hosting/docker_configurator.py

@@ -4,6 +4,7 @@ from lib import custom_entrypoint
 from lib import networking
 from lib import wireguard
 from lib import logging as logging_lib
+from clore_hosting import utils as hosting_utils
 import shutil
 import os
 import re
@@ -53,9 +54,12 @@ def configure(containers):
 
         for index, container in enumerate(containers):
             ok_custom_entrypoint = False
+            invalid_hostname = False
             if index < len(custom_entrypoint_state):
                 ok_custom_entrypoint = custom_entrypoint_state[index]
             startup_script_name = f"{container['name']}.sh"
+            if "hostname" in container and not hosting_utils.validate_hostname(container["hostname"]):
+                invalid_hostname = True
             if "ip" in container and len(container["ip"])>6 and type(container["ip"])==str:
                 if container["ip"][:8] == "; echo '":
                     last_occurrence, text_after_last_ip = get_last_ip_occurrence_and_text(container["ip"])
@@ -95,14 +99,14 @@ def configure(containers):
                             newly_created_networks.append(container["network"])
                         else:
                             any_fail=True
-                    if not any_fail and ok_custom_entrypoint:
+                    if not any_fail and ok_custom_entrypoint and not invalid_hostname:
                         valid_containers.append(container)
             elif "network" in container and container["network"][:len(config.clore_network_name_prefix)]==config.clore_network_name_prefix: # Subnet & gateway not defined, must be some of default networks, otherwise dump it
                 if container["network"] in default_network_names:
                     for docker_network in docker_networks:
                         if docker_network["Name"]==container["network"]:
                             for ipam in docker_network["IPAM"]:
-                                if not ok_custom_entrypoint:
+                                if not ok_custom_entrypoint or invalid_hostname:
                                     break
                                 elif not "ip" in container:
                                     valid_containers.append(container)

clore_hosting/main.py

@@ -106,6 +106,15 @@ class CloreClient:
                 "expiration":"immune"
             }
 
+        self.os_release = get_specs.get_os_release()
+        self.restart_docker = False
+        if "use_cgroupfs" in self.os_release:
+            self.updated_exec_opts = True if docker_interface.configure_exec_opts("native.cgroupdriver","cgroupfs") else False
+            if self.updated_exec_opts:
+                docker_info = docker_interface.get_info()
+                if "CgroupDriver" in docker_info and docker_info["CgroupDriver"]=="systemd":
+                    self.restart_docker = True # Restart docker when it's loaded under systemd (accual restart will happen only if no orders running to not disrupt workload)
+
         docker_interface.verify_docker_version()
         nvml.init()
 
@@ -333,6 +342,7 @@ class CloreClient:
                     print("STEP",step,'|',self.containers_set, self.containers if config.log_containers_strings else '')
 
                 tasks = []
+                running_order = False
 
                 container_conf = WebSocketClient.get_containers()
 
@@ -345,6 +355,7 @@ class CloreClient:
                             log_pull = False
                             if "name" in container:
                                 if "-order-" in container["name"]:
+                                    running_order=True
                                     log_pull=True
                             image_config = {
                                 "image":container["image"],
@@ -362,6 +373,12 @@ class CloreClient:
 
                             if not image_config in tmp_images:
                                 tmp_images.append(image_config)
+
+                    if self.restart_docker and not running_order and len(self.containers)>0:
+                        log.debug("Sending docker restart command")
+                        utils.run_command_v2("systemctl restart docker")
+                        self.restart_docker=False
+                    
                     if tmp_images!=self.needed_images:
                         self.needed_images=tmp_images
                         await pull_list.put(self.needed_images)
@@ -426,7 +443,7 @@ class CloreClient:
     async def submit_specs(self, current_specs):
         try:
             if type(current_specs) == dict:
-                current_specs["backend_version"]=11
+                current_specs["backend_version"]=12
                 current_specs["update_hw"]=True
                 smallest_pcie_width = 999
                 for gpu in current_specs["gpus"]["nvidia"]:

clore_hosting/utils.py

@@ -10,5 +10,13 @@ def is_valid_websocket_url(url):
             return True
     return False
 
+def validate_hostname(hostname):
+    # Define the regular expression pattern for a valid hostname
+    pattern = re.compile(r'^[a-zA-Z0-9._-]{1,63}$')
+    if pattern.match(hostname):
+        return True
+    else:
+        return False
+
 def unix_timestamp():
     return int(time.time())

lib/docker_cli_wrapper.py

@@ -21,6 +21,9 @@ def create_container(container_options, ip=None, docker_gpus=False, timeout=30):
     if "network_mode" in container_options:
         command.extend(["--network", container_options["network_mode"]])
 
+    if "hostname" in container_options:
+        command.extend(["--hostname", container_options["hostname"]])
+
     if "cap_add" in container_options:
         for cap in container_options["cap_add"]:
             command.extend(["--cap-add", cap])

lib/docker_deploy.py

@@ -76,6 +76,13 @@ def deploy(validated_containers):
                 )
             }
 
+            if "hostname" in validated_container:
+                container_options["hostname"]=validated_container["hostname"]
+            elif "clore-order-" in validated_container["name"]:
+                try:
+                    container_options["hostname"] = f"O-{int(validated_container["name"][12:])}"
+                except Exception as eon:
+                    pass
             if "network" in validated_container:
                 container_options["network_mode"]=validated_container["network"]
             if "ip" in validated_container and config.creation_engine=="sdk":

lib/docker_interface.py

@@ -50,6 +50,14 @@ class DockerNetwork(BaseModel):
 
 client = docker.from_env()
 low_level_client = docker.APIClient(base_url='unix://var/run/docker.sock')
+daemon_config_path = "/etc/docker/daemon.json"
+
+def get_info():
+    try:
+        client_info = client.info()
+        return client_info
+    except Exception as e:
+        return {}
 
 def check_docker_connection():
     try:
@@ -346,16 +354,15 @@ def validate_and_secure_networks():
 
 
 def get_daemon_config():
-    config_path = "/etc/docker/daemon.json"
     try:
-        with open(config_path, 'r') as file:
+        with open(daemon_config_path, 'r') as file:
             config_data = json.load(file)
             return config_data
     except FileNotFoundError:
-        print(f"Error: {config_path} not found.")
+        print(f"Error: {daemon_config_path} not found.")
         return None
     except json.JSONDecodeError:
-        print(f"Error: Failed to parse JSON from {config_path}.")
+        print(f"Error: Failed to parse JSON from {daemon_config_path}.")
         return None
     
 def verify_docker_version(min_version="17.06"):
@@ -368,3 +375,41 @@ def verify_docker_version(min_version="17.06"):
     except Exception as e:
         log.error(f"Failed to verify docker version | {e}")
         os._exit(1)
+
+def configure_exec_opts(key="native.cgroupdriver", value="cgroupfs"):
+    deamon_config = get_daemon_config()
+    if deamon_config:
+        try:
+            if (not "exec-opts" in deamon_config or type(deamon_config["exec-opts"])!=list) and value!=None:
+                deamon_config["exec-opts"]=[f"{key}={value}"]
+            elif "exec-opts" in deamon_config:
+                new_exec_opts=[]
+                matched_key=False
+                for exec_opt in deamon_config["exec-opts"]:
+                    if '=' in exec_opt:
+                        exec_opt_key, exec_opt_value = exec_opt.split('=',1)
+                        if exec_opt_key==key:
+                            matched_key=True
+                            if value!=None:
+                                new_exec_opts.append(f"{key}={value}")
+                        else:
+                            new_exec_opts.append(exec_opt)
+                    else:
+                        new_exec_opts.append(exec_opt)
+                if not matched_key:
+                    new_exec_opts.append(f"{key}={value}")
+                if len(new_exec_opts)==0:
+                    del deamon_config["exec-opts"]
+                else:
+                    if deamon_config["exec-opts"] == new_exec_opts:
+                        return "Same"
+                    deamon_config["exec-opts"]=new_exec_opts
+            json_string = json.dumps(deamon_config, indent=4)
+            with open(daemon_config_path, 'w') as file:
+                file.write(json_string)
+            return True
+        except Exception as e:
+            log.error(f"Failed 'configure_exec_opts' | {e}")
+            return False
+    else:
+        return False

lib/get_specs.py

@@ -43,6 +43,28 @@ def get_kernel():
 def is_hive():
     return "hive" in get_kernel()
 
+def get_os_release():
+    try:
+        with open("/etc/os-release") as f:
+            os_info = f.read()
+        os_release = {}
+        for line in os_info.split('\n'):
+            if '=' in line:
+                key, value = line.split('=', 1)
+                if value[:1]=='"' and value.endswith('"'):
+                    value = value[1:len(value)-1]
+                    os_release[key]=value
+
+        needed_cgroupfs_versions = ["22.04", "22.10"] # Mitigate issue https://github.com/NVIDIA/nvidia-docker/issues/1730
+
+        if "NAME" in os_release and "VERSION_ID" in os_release:
+            if os_release["NAME"].lower() == "ubuntu" and os_release["VERSION_ID"] in needed_cgroupfs_versions:
+                os_release["use_cgroupfs"]=True
+
+        return os_release
+    except Exception as e:
+        return {}
+
 def drop_caches():
     try:
         with open('/proc/sys/vm/drop_caches', 'w') as f: