xCAT/xcat-core
2
0
Fork 0
mirror of https://github.com/xcat2/xcat-core.git synced 2024-11-21 09:02:00 +00:00
Code Issues Releases Wiki Activity
Page: Hierarchical_Design
Pages
2014_10_01_Bash_vulnerability 2014_11_04_Poodle_Attack AIX_NIM_NFSv4_support Add_Extra_Parameters_for_confignics_Postscript CRHS like_function CRHS like_function_enhancements CSM_CFM like_function_to_sync_config_files_to_the_nodes Capture_Linux_Image_from_running_node_to_create_a_stateless statelite_image Chef_databag_support_for_deploying_openstack_in_xCAT Cluster_config_file Conserver_Replacement Coordinated_Cluster_Shutdown_and_Bringup Copycds_enhancement Coral PDU supports Create_osimage_definition_for_Management_Node Cumulus_OS_Upgrade_Support_Test DFM_Service_Node_Hierarchy_support DHCPv6_management DNS_Enhancments DNS_master_slave_support Def_commands_to_display set_nic_attributes_more_easily Deploying_OpenStack_Baremetal_Nodes_Using_xCAT Disable_ssh_root_passwordless_node to node_access Document_large_cluster_tunables,_etc. Dynamic_node_groups_support Energy_Management_for_Flex Energy_management_support_for_blade Enhance_NTP_setup_support Enhancement_of_lsslp_ flexdiscover Enhancement_on_makeroutes_and_snmove_commands Exporting_more_table_attributes_to_nodes Filling_the_osimage_attributes_with_the_files_that_are_found_in_the_search_path Full_Automation_of_DB2_Setup_on_Service_Nodes Function Verification Test Design of xCAT High Availability GPFS_statelite_nodes Getmacs_and_rnetboot_enhancement Getnodetype_performance_imprvoments_ _phase_I HA_NFS_on_AIX_service_nodes HA_Service_Nodes_for_AIX HA_Service_Nodes_for_AIX_ _using_External_NFS_Server HFI_MAC_address_failover HPC_Integration_Phase_1 Hardware_Inventory_for_Node Health_Check_Script_Framework Hierarchical_Design Home Image_Provisioning Image_import_and_export_enhancement_for_Linux Improve_perl_array_search_efficiency_for_scalability Improvement_of_pre creating_mypostscripts_on_linux Initrd_Enhancement_separate_initrd_hacking_from_noteset_get_drivers_from_os_update_distro Kdump_for_Linux_diskless_nodes Linux_AIX_mixed_Clusters_postscripts_cleanup Local_Disk_support_for_statelite Lstree_for_Servicenode_Hardware_and_VM_hierarchies Management_Node_Pool Management_Node_Setup_Verification Managing_Multiple_Sites_and_Clusters Mellanox_Switch_Support Mini design for openbmc python framework Mini design_of_updating_pre defined_groups_for_nodes Minimize_the_install setup_steps_of_xcat_on_the_MN Mkvlan_enhancements Multiple_Zone_Support2 Multiple_pkgdir_support_with_full_installation NFS_Redundancy_Investigation NFS_redundancy New_Cloud_Tables_in_xCAT Node_Status_and_Application_Status Node_definition_installation_service_node_setup_verification Nodeset_defaults_when_node_provmethod=osimage_is_set OS_Deployment_Monitoring_And_Automatic_Retry OS_Distro_Update_xCAT_command_osdistroupdate OpenStack Chef Cookbook_xCAT_Integration OpenStack_Deployment_using_Chef OpenStack_Deployment_using_Puppet Ospkgs_enhancements_to_install upgrade_OS_packages_from_Updates Overview_of_HA_Services_for_Diskless_AIX_Nodes Power8_BE_support Predict_long_nicname_for_RH7 Proposal_for_docker_Support_in_xCAT RHEL7_Support RHEV_support Rcons_hierachy_support Redfish_support Refine_the_usage_of_installnic_and_primarynic_attributes Release processs automation investigation Resolving_Nodenames_for_System_p_HW_Ctrl Run_parallel_commands_for_Ethernet_switches Run_postbootscripts_after_each_diskful_reboot SLES12_Support SLES_11_SP1_SDK_support Secure root support Sequential_Discovery Sequential_Discovery_phase_2 Service_Node_Scenarios_and_Retry_Logic_in_Postscript_Process Service_node_manual_takeover_on_AIX Service_node_take_over Setup_IP_routes_between_the_MN_and_the_CNs Shared data based xCAT MN HA mini design Shared data based xCAT MN HA user case Shared data based xCAT MN HA Specify_the_disk_partitions_during_diskfull_install Steps to install xcat_openbmc_py3 on RH8 Support uploading openbmc firmware in parallel Support_Management_Node_in_the_servicenode_table Support_for_Redundant_BPAs_and_FSPs Support_multiple_hardware_control_points Support_osimage.pkgdir_with_multiple_paths_in_genimage_and_diskfull_installation Support_postscripts_associated_with_rsync_files_to_be_run_after_rsync_in_xdcp Support_to_inject_drivers_from_Driver_RPM_Package Support_xcatd_running_user_defined_sql_scripts_on_start_and_foreign_keys Supporting_driver_update_disk_for_Linux Supporting_the_scalable_functions_for_HX5_blade Switch_Discovery_And_Switch_Management System_p_Discovery_Enhancements System_z_hierarchical_tree Test Design of HA Test Design of TestCase Label Test Design of `goconserver` Regression Testing Test Design of `goconserver` Support Test Design of configib Test Design of cumulus osimage Test Design of cumulus_osimage Test Design of gonconserver redirect logs Test Design of hardware control commands regression Test Design of makegocons option l Test Design of rbeacon python version Test Design of rcons based goconserver Test Design of revenlog python version Test Design of rflash python version Test Design of rinv python version Test Design of rpower python version Test Design of rsetboot python version Test Design of rspconfig python version Test Design of rvitals python version Test Design of syncfiles Test Design of xcat inventory diff Test Design of xcat inventory export import performance Test Design of xcat inventory option e Test Design of xcat inventory osimage environment variables export part Test Design of xcat inventory osimage environment variables Test Design of xcat inventory osimage phase2 Test Design of xcat inventory osimage support share file directory Test Design of xcat inventory osimage Test Design of xcat inventory phase 1 Test Design of xcat inventory regular expression Test Design of xcat inventory save keys Test Design of xcat inventory user case Test Design of xcat terraform provider phase1 Test Report of `goconserver` Test cases against nodeset error handling Test_Case_Designs_FVT The mini design of osimage version control(initial draft) The mini design of syncfiles for cumulus The mini design of using xcat inventory templates to setup xCAT cluster with ansible The_remote_console_support_of_hmc Trace_Facility_for_xCAT Ubuntu_supports_internet_repository Updatenode_Performance_Enhancements Updatenode_Sync_Service_Nodes_for_install Updatenode_enhancement_for_security_keys_and_certificates Updatenode_support_to_run_mypostscript_sections Updatenode_to_the_Management_Node Updates_to_xcatsetup_and_Cluster_Config_File Windows_Multiple_Disks_Partitions_Configuration Windows_Multiple_Nics_Configuration Windows_Multiple_Winpes Windows_Support_to_Run_Postscripts_Postbootscripts_in_Windows_Deployment Wish_List_for_xCAT_2 Wish_List_for_xCAT_FVT XCAT 2 Mini Designs for New Features XCAT_2.10_Release_Notes XCAT_2.11.1_Release_Notes XCAT_2.11_ESP_Release_Notes XCAT_2.11_Release_Notes XCAT_2.12.1_Release_Notes XCAT_2.12.2_Release_Notes XCAT_2.12.3_Release_Notes XCAT_2.12.4_Release_Notes XCAT_2.12.5_Release_Notes XCAT_2.12_Release_Notes XCAT_2.13.10_Release_Notes XCAT_2.13.11_Release_Notes XCAT_2.13.1_Release_Notes XCAT_2.13.2_Release_Notes XCAT_2.13.3_Release_Notes XCAT_2.13.4_Release_Notes XCAT_2.13.5_Release_Notes XCAT_2.13.6_Release_Notes XCAT_2.13.7_Release_Notes XCAT_2.13.8_Release_Notes XCAT_2.13.9_Release_Notes XCAT_2.13_Release_Notes XCAT_2.14.1_Release_Notes XCAT_2.14.2_Release_Notes XCAT_2.14.3_Release_Notes XCAT_2.14.4_Release_Notes XCAT_2.14.5_Release_Notes XCAT_2.14.6_Release_Notes XCAT_2.14_Release_Notes XCAT_2.15.1_Release_Notes XCAT_2.15_Release_Notes XCAT_2.16.1_Release_Notes XCAT_2.16.2_Release_Notes XCAT_2.16.3_Release_Notes XCAT_2.16.4_Release_Notes XCAT_2.16.5_Release_Notes XCAT_2.16_Release_Notes XCAT_2.17_Release_Notes XCAT_2.9.2_Release_Notes XCAT_2.9.2_aix_Release_Notes XCAT_2.9.3_Release_Notes XCAT_2.9.4_Release_Notes XCAT_Cluster_in_SoftLayer XCAT_HPC_Integration_Files_per_OS,_Architecture,_Product XCAT_Health_Check XCAT_IPv6_support_on_AIX XCAT_IPv6_support_on_Linux XCAT_Mini Design_Template XCAT_Monitoring_Failure_Analysis_Conditions_Batched_Events XCAT_Postscripts_framework_enhancements XCAT_Provisioning_OS_on_consistent_disk XCAT_REST_API_enhancements XCAT_Support_for_Kits_and_Kit_Components XCAT_probe XCAT_support_for_multiple_network_domains Xcatexport_and_xcatimport_commands Xcatmon_and_appstatus_for_HPC_products Xeon_Phi_support cofignetwork_for_RH7 confignetwork integrate confignics_configeth_configib functions usage determine_install_disk diskdiscover_and_configraid docker_instances_lifecycle_management enhancement_for_rinstall_command logging pre script to make debug easy makedhcp_append_statement_automatically_rather_than_get_statement_from_nodeset mini design of ZTP based onie switch discovery and configuration minidesign Switch based switch discovery minidesign_xcatprobe node status enhancement ntp_enhancement openbmc_support post.xcat_restructure rflash_streamline_procedure rscan command for hypervisor rspconfig network rspconfig ntpservers rspconfig password scaleup_basedon_xcat_in_openshift_cluster supporting_partitionfile_for_Ubuntu supporting_persistent_kernel_cmdline_after_diskful_installation test_sidebar the mini design of xcat inventory export and import for osimage use fifo pipe to enhance getadapter xCAT Build Schedule xCAT_2.12_ReleaseInformation xCAT_2.13_ReleaseInformation xCAT_2.14_ReleaseInformation xCAT_2.15_ReleaseInformation xCAT_Docker_Image xcatprobe_framework xcatprobe_requests
16 Hierarchical_Design
ligc edited this page 2015-07-30 04:22:49 -04:00
Table of Contents

Table of Contents

Design_Warning

A few rules about how xCAT should behave in a hierarchical cluster (i.e. a cluster w/service nodes).

Attribute Summary

  • noderes.servicenode: the service node for this node as known by the mn. Used normally for distributing out cmds from the mn.
  • noderes.xcatmaster: the service node for this node as known by the node. The default for several of the attrs below.
  • noderes.tftpserver: the tftp service node for this node as known by the node (defaults back to noderes.xcatmaster)
  • noderes.nfsserver: the service node that this node should get files from via http, ftp, etc. (as known by the node).
  • noderes.monserver: the monitoring event collection point. Two hostnames separated by a comma: as known by mn, and as known by node
  • nodehm.conserver: the service node that should provide console access to this node (as known by the mn)
  • networks.dhcpserver: specifies the dhcpserver on each network.
  • If the noderes/nodehm specific attrs are not specified for a node:
    • if the servicenode table has at least one entry, try to guess which service node(s) should service this node based on networks
    • if the servicenode table is empty, assume the node is serviced by the mn

New servicenode Table

  • Lists all the service nodes in your cluster and what services they should be running. You can set servicenode-related attributes in nodehm or noderes to explicitly state that either a specific sn or the mn should be used.
  • Attributes:
    • node - hostname of service node (as known by mn)
    • nameserver - yes or 1 to set up dns caching (default is yes)
    • dhcpserver - yes or 1 (default is yes) (may support including NIC names, dynamic range, noderange, but this is still being discussed)
    • tftpserver - yes or 1 (default is yes)
    • nfsserver - yes or 1 (default is yes) Note that this doesn't necessarily mean nfs. It is a general attr for any file access services (e.g. http, ftp)
    • conserver - yes or 1 (default is yes)
    • monserver - yes or 1 (default is yes)
    • ntpserver - yes or 1 (default is yes)
    • ldapserver - yes or 1 (ldap proxy) (default is yes)
    • (can't have syslog in this table because the sn needs to set up syslog before the db access is established)

Environments to Support

  • Subnetted Service Nodes - Each service node is on its own subnet with the nodes it is responsible for.
    • Every node/nodegroup has explict noderes.servicenode, noderes.xcatmaster, etc., entries. The user must ensure that each servicenode listed in one of those attributes is also added to the servicenode table for correct servicenode install/configuration.
  • Pooled Service Nodes - The cluster network is logically flat and any service node can service any node.
    • All service nodes are listed in the service node table. (The nodename key in the servicenode table is "as known by the management node".)
    • noderes.xcatmaster is null, indicating that the SN that responds to the node boot should be its SN.
    • noderes.servicenode can contain a list of SNs, so that top down cmds like xdsh can be dispatched to any one of them (trying them in order)
  • Combination of Subnetted and Pooled Service Nodes - Compute nodes are segregated into subnets, but within each subnet there is a pool of service nodes.
    • noderes.xcatmaster is null, indicating that the SN that responds to the node boot should be its SN.
    • noderes.servicenode can contain a list of SNs in that pool, so that top down cmds like xdsh can be dispatched to any one of them (trying them in order)
  • site.disjointdhcps should be set to 1 so that the dhcpd.conf on the SN will be filled in with only the nodes that have that SN listed in their noderes.servicenode attribute

Makedhcp

  • Command will be sent to all service nodes in servicenode table that have dhcpserver attr filled in.
  • On each service node:
    • the NICs on the SN will be compared to the xcat networks table and all nodes that are on a subnet with this SN will be put in its dhcpd.conf

Nodeset

  • For now, service nodes that have servicenode.tftpserver filled in will always mount /tftpboot read/write.
  • This means the nodeset cmd can be done completely from the mn.
  • This also means there is no need for guessing, because all sn's that have the tftp service set up will be able to serve any node.

Makedns

  • Every sn in the servicenode table that has the nameserver attr filled in will have dns set up on it.
  • The DNS on the SN is just a forwarding/caching DNS set up to forward all requests to the MN, and then remember the answer

For more information about name resolution in an xCAT Cluster, refer to the following:

[Cluster_Name_Resolution]

Makenetworks

  • The cmd should be sent to every sn in the servicenode table, so that it can discover its networks.

Conserver

  • Conserver will always use nodehm.conserver to determine who should have the console for each node.
  • The cmd makeconservercf should be distributed to the service nodes based on nodehm.conserver

Distributing Commands Run by the Admin on the MN

  • Use noderes.servicenode to determine who should run the cmd for each node
  • For nodes that don't have a noderes.servicenode value, randomly distribute the cmds to the sn's in the servicenode table

HW Ctrl Cmds

  • If noderes.servicenode is set for a hw ctrl point (mm, hmc, fsp, etc.), then hw ctrl cmds should be distributed to those service nodes
    • I.e. if "rpower node1 on" is run, xcat 1st looks up the hcp (e.g. ppc.hcp) of node1. Assume it is called fsp1. Then it looks up noderes.servicenode for fsp1. If that is set, for example, to sn1, then the rpower cmd will be dispatch to sn1 and then sn1 will contact fsp1 to power on node1.

Scaling and HA Considerations

The goal for the service nodes is to configure them such that we can have more than one in a building block and any node in that BB can be serviced by any service node in the BB.  This way if we need better scaling within the BB, we can simply add more service nodes.  And to provide HA we can simply have multiple service nodes and if one service node goes down, the other service nodes within the BB can pick up the load.  For example, all service nodes within a BB have DHCP and/or NIM configured for all the compute nodes in the BB, and have all the same OS images on them.  Thus when a compute node boots, it will broadcast PXE or bootp for a server, and whichever service node responds first can deploy the node.

The paragraph above states the ultimate goal, but there may be some services that can't be set up this way.  In these cases, we may have to divide up responsibility of the compute nodes between the service nodes.  But we want to limit the number of services we have to do this for, so we limit how much we have to fail over to another service node when one service node dies.

Also, not all services for a compute node have to be provided by the same service node.  For example, TFTP can be served from one machine, DHCP from another, and NFS from another.

Database Access

Information about all the nodes in the whole cluster is kept in the xCAT DB on the management node (EMS).  This is the master source of the data.  Specific information that is needed by a service node for a specific operation must be accessed somehow from the management node.  There are several possible ways to do this:

  • Every plugin is structured such that it is split into 2 functions:  preprocess_request() and process_request().  The preprocess_request() function queries the DB for whatever data it needs (for that operation for the nodes specified) and stores it temporarily in internal hash tables.  The hash tables are passed to process_request() to actually perform the operation.  When a plugin is executed directly on the MN, both are run in succession.  When the operation needs to be distributed out to the service nodes, preprocess_request() is run on the MN, then the data is passed to the appropriate SNs and process_request() is run on the SNs. The is the preferred solution.
  • Have the SN query the MN for the data it needs, using the client/server protocol.  For example, when a service node needs to be configured to respond correctly to compute node PXE requests, the makedhcp command is run on the SN and it is passed a list of nodes that this service node is responsible for.  This command queries the particular node attributes it needs, using the client/server connection to xcatd on the management node.  Then it writes out dhcpd.conf. This will cause a scaling bottleneck on the management node for large clusters, because all the service nodes will be asking the management node for data at the same time?

Setting Up the Service Nodes

The management node installs (or diskless boots) the service nodes and uses the servicenode postinstall script to set up the node as a service node.

Details for Specific Commands

Because the scaling challenges for each operation vary greatly, we can't use the same approach for every operation.  This section enumerates many of the commands and operations and describes what approach will be used for each to overcome the scaling issues. (This info may be out of date):

  • PXE Booting Compute Nodes:  DHCP will be configured on each service node to respond to the PXE requests for its nodes.  The /tftpboot file system will be mounted from the management node so it will have all the proper hex files.  The /install file system will also be mounted from the MN so the OS image files will be available if the compute node is being installed or diskless booted.
  • For system p support, each service node is a NIM master.  The MN is a NIM master also and originally all the SPOT and lpp_source resources on the MN should be put in /install.  Since this file system will be mounted on the service nodes, the appropriate NIM resources can be defined on each service node and refer (when necessary) to the software files in /install.  (Configuring NIM automatically configures the bootp daemon.)
  • The rpower command for system x (BMC) will be passed along to the service nodes (along with the node list and attributes) via the xcat client/server mechanism.  The service node will communicate with each node's BMC to perform the operation.  Note, that the BMC user/pw info must also be sent, so the client/server XML communication may need to be encrypted.
  • Should the "simple" rpower commands (query, on, off) for system p be done straight from the MN (w/o involving the service node), since it is really the HMC that has to implement it anyway??
  • The boot-to-openfirmware rpower command will be distributed to the service nodes.
  • The compute nodes will forward their syslog entries to the service node.  Will the service nodes forward all syslog entries to the MN?
  • Each service node will run an intermediate DNS daemon.
  • Each service node will run an intermediate NTP daemon.
  • One designated service node in each BB will run conserver and the compute nodes will have their consoles directed to that.  The rconsole command on the MN will connect to the conserver on the relevant SN.  A specific conserver.cf for each SN needs to be generated with just its CNs.
  • Each node will have a MonitorService attribute to designate which service node will be the collection point for events and performance monitoring info.

News

History

  • Oct 22, 2010: xCAT 2.5 released.
  • Apr 30, 2010: xCAT 2.4 is released.
  • Oct 31, 2009: xCAT 2.3 released.
    xCAT's 10 year anniversary!
  • Apr 16, 2009: xCAT 2.2 released.
  • Oct 31, 2008: xCAT 2.1 released.
  • Sep 12, 2008: Support for xCAT 2
    can now be purchased!
  • June 9, 2008: xCAT breaths life into
    (at the time) the fastest
    supercomputer on the planet
  • May 30, 2008: xCAT 2.0 for Linux
    officially released!
  • Oct 31, 2007: IBM open sources
    xCAT 2.0 to allow collaboration
    among all of the xCAT users.
  • Oct 31, 1999: xCAT 1.0 is born!
    xCAT started out as a project in
    IBM developed by Egan Ford. It
    was quickly adopted by customers
    and IBM manufacturing sites to
    rapidly deploy clusters.