LNet Router Config Guide2

From Lustre Wiki
Revision as of 15:41, 12 October 2017 by Sharmaso (talk | contribs)
Jump to navigation Jump to search

This document provides procedures to configure and tune an LNet router. It will also cover detailed instructions set on setting up connectivity of an Infiniband network to Intel OPA nodes using LNet router.

LNet

LNet supports different network types like Ethernet, InfiniBand, Intel Omni-Path and other proprietary network technologies such as the Cray’s Gemini. It routes LNet messages between different LNet networks using LNet routing. LNet’s routing capabilities provide an efficient protocol to enable bridging between different types of networks. LNet is part of the Linux kernel space and allows for full RDMA throughput and zero copy communications when supported by underlying network. Lustre can initiate a multi-OST read or write using a single Remote Procedure Call (RPC), which allows the client to access data using RDMA at near peak bandwidth rates. With Multi-Rail (MR) feature implemented in Lustre 2.10.X, it allows for multiple interfaces of same type on a node to be grouped together under the same LNet (ex tcp0, o2ib0, etc.). These interfaces can then be used simultaneously to carry LNet traffic. MR also has the ability to utilize multiple interfaces configured on different networks. For example, OPA and MLX interfaces can be grouped under their respective LNet and then can be utilized with MR feature to carry LNet traffic simultaneously.


LNet Configuration Example

An LNet router is a specialized Lustre client where Lustre file system is not mounted and only the LNet is running. A single LNet router can serve different file systems.



For the above example:

  • Servers are on LAN1, a Mellanox based InfiniBand network – 10.10.0.0/24
  • Clients are LAN2, an Intel OPA network – 10.20.0.0/24
  • Routers on LAN1 and LAN2 at 10.10.0.20, 10.10.0.21 and 10.20.0.29, 10.20.0.30 respectively

The network configuration on the nodes can be done either by adding the module parameters in lustre.conf /etc/modprobe.d/lustre.conf or dynamically by using the lnetctl command utility. Also, current configuration can be exported to a YAML format file and then the configuration can be set by importing that YAML file anytime needed.


Network Configuration by adding module parameters in lustre.conf

Servers:
options lnet networks="o2ib1(ib0)" routes="o2ib2 [email protected]"
 
Routers:
options lnet networks="o2ib1(ib0),o2ib2(ib1)" "forwarding=enabled"
 
Clients:
options lnet networks="o2ib2(ib0)" routes="o2ib1 [email protected]" 

NOTE: Restarting LNet is necessary to apply the new configuration. To do this, it is needed to unconfigure the LNet network and reconfigure again. Make sure that the Lustre network and Lustre file system are stopped prior to unloading the modules.

// To unload and load LNet module
modprobe -r lnet
modprobe lnet
  
// To unconfigure and reconfigure LNet
lnetctl lnet unconfigure
lnetctl lnet configure 


Dynamic Network Configuration using lnetctl command

Servers:
lnetctl net add --net o2ib1 --if ib0
lnetctl route add --net o2ib2 --gateway [email protected]
lnetctl peer add --nid [email protected]
 
Routers:
lnetctl net add --net o2ib1 --if ib0
lnetctl net add --net o2ib2 --if ib1
lnetctl peer add --nid [email protected]
lnetctl peer add --nid [email protected]
lnetctl set routing 1
   
Clients:
lnetctl net add --net o2ib2 --if ib0
lnetctl route add --net o2ib1 --gateway [email protected]
lnetctl peer add --nid [email protected] 


Importing/Exporting configuration using a YAML file format

// To export the current configuration to a YAML file
lnetctl export FILE.yaml
lnetctl export > FILE.yaml
  
  
// To import the configuration from a YAML file
lnetctl import FILE.yaml
lnetctl import < FILE.yaml 

There is a default lnet.conf file installed at /etc/lnet.conf which has an example configuration in YAML format. Another example of a configuration in a YAML file is:

net:
    - net type: o2ib1
      local NI(s):
        - nid: [email protected]
          status: up
          interfaces:
              0: ib0
          tunables:
              peer_timeout: 180
              peer_credits: 8
              peer_buffer_credits: 0
              credits: 256
          lnd tunables:
              peercredits_hiw: 64
              map_on_demand: 32
              concurrent_sends: 256
              fmr_pool_size: 2048
              fmr_flush_trigger: 512
              fmr_cache: 1
          tcp bonding: 0
          dev cpt: -1
          CPT: "[0]"
route:
    - net: o2ib2
      gateway: [email protected]
      hop: 1
      priority: 0
      state: up
peer:
    - primary nid: [email protected]
      Multi-Rail: False
      peer ni:
        - nid: [email protected]
          state: up
          max_ni_tx_credits: 8
          available_tx_credits: 8
          min_tx_credits: 7
          tx_q_num_of_buf: 0
          available_rtr_credits: 8
          min_rtr_credits: 8
          refcount: 4
global:
    numa_range: 0
    max_intf: 200
    discovery: 1 


LNet provides a mechanism to monitor each route entry. LNet pings each gateway identified in the route entry on regular, configurable interval (live_router_check_interval) to ensure that it is alive. If sending over a specific route fails or if the router pinger determines that the gateway is down, then the route is marked as down and is not used. It is subsequently pinged on regular, configurable intervals (dead_router_check_interval) to determine when it becomes alive again.


Multi-Rail LNet Configuration Example

If the routers are MR enabled, we can add the routers as peers with multiple interfaces to the clients and the servers, the MR algorithm will ensure that both interfaces of the routers are used while sending traffic to the router. However, single interface failure will still cause the entire router to go down. With the network topology example in Figure 1 above, LNet MR can be configured like below:

Servers:
lnetctl net add --net o2ib1 --if ib0,ib1
lnetctl route add --net o2ib2 --gateway [email protected]
lnetctl peer add --nid [email protected],[email protected]
 
Routers:
lnetctl net add --net o2ib1 --if ib0,ib1
lnetctl net add --net o2ib2 --if ib2,ib3
lnetctl peer add --nid [email protected],[email protected]
lnetctl peer add --nid [email protected],[email protected]
lnetctl set routing 1
   
Clients:
lnetctl net add --net o2ib2 --if ib0,ib1
lnetctl route add --net o2ib1 --gateway [email protected]
lnetctl peer add --nid [email protected],[email protected] 


Fine-Grained Routing

The routes parameter, by identifying LNet routers in a Lustre configuration, is used to tell a node which route to use when forwarding traffic. It specifies a semi-colon-separated list of router definitions.

routes=dest_lnet [hop] [priority] [email protected]_lnet; \ dest_lnet [hop] [priority] [email protected]_lnet 

An alternative syntax consists of a colon-separated list of router definitions:

routes=dest_lnet: [hop] [priority] [email protected]_lnet \ [hop] [priority] [email protected]_lnet 

When there are two or more LNet routers, it is possible to give weighted priorities to each router using the priority parameter. Here are some possible reasons for using this parameter:

  • One of the routers is more capable than the other.
  • One router is a primary router and the other is a back-up.
  • One router is for one section of clients and the other is for another section.

Each router is moving traffic to a different physical location. The priority parameter is optional and need not be specified if no priority exists. The hop parameter specifies the number of hops to the destination. When a node forwards traffic, the route with the least number of hops is used. If multiple routes to the same destination network have the same number of hops, the traffic is distributed between these routes in a round-robin fashion. To reach/transmit to the LNet dest_lnet, the next hop for a given node is the LNet router with the NID router_NID in the LNet src_lnet. Given a sufficiently well-architected system, it is possible to map the flow to and from every client or server. This type of routing has also been called fine-grained routing.


Advanced Routing Parameters

In a Lustre configuration where different types of LNet networks are connected by routers, several kernel module parameters can be set to monitor and improve routing performance. These parameters are set in /etc/modprobe.d/lustre.conf file. The routing related parameters are:

  • auto_down - Enable/disable (1/0) the automatic marking of router state as up or down. The default value is 1. To disable router marking, set:
options lnet auto_down=0 
  • avoid_asym_router_failure - Specifies that if even one interface of a router is down for some reason, the entire router is marked as down. This is important because if nodes are not aware that the interface on one side is down, they will still keep pushing data to the other side presuming that the router is healthy, when it really is not. To turn it on:
options lnet avoid_asym_router_failure=1 
  • live_router_check_interval - Specifies a time interval in seconds after which the router checker will ping the live routers. The default value is 60. To set the value to 50, use:
options lnet live_router_check_interval=50 
  • dead_router_check_interval - Specifies a time interval in seconds after which the router checker will check the dead routers. The default value is 60. To set the value to 50:
options lnet dead_router_check_interval=50 
  • router_ping_timeout - Specifies a timeout for the router checker when it checks live or dead routers. The router checker sends a ping message to each dead or live router once every dead_router_check_interval or live_router_check_interval respectively. The default value is 50. To set the value to 60:
options lnet router_ping_timeout=60 
  • check_routers_before_use - Specifies that routers are to be checked before use. Set to off by default. If this parameter is set to on, the dead_router_check_interval parameter must be given a positive integer value.
options lnet check_routers_before_use=1 


The router_checker obtains the following information from each router:

  • time the router was disabled
  • elapsed disable time

If the router_checker does not get a reply message from the router within router_ping_timeout seconds, it considers the router to be down. If a router that is marked “up” responds to a ping, the timeout is reset. If 100 packets have been sent successfully through a router, the sent-packets counter for that router will have a value of 100. The statistics data of an LNet router can be found from /proc/sys/lnet/stats. If no interval is specified, then statistics are sampled and printed only one time. Otherwise, statistics are sampled and printed at the specified interval (in seconds). These statistics can be displayed using lnetctl utility as well like below:

# lnetctl stats show
statistics:
    msgs_alloc: 0
    msgs_max: 2
    errors: 0
    send_count: 887
    recv_count: 887
    route_count: 0
    drop_count: 0
    send_length: 656
    recv_length: 70048
    route_length: 0
    drop_length: 0 

The ping response also provides the status of the NIDs of the node being pinged. In this way, the pinging node knows whether to keep using this node as a next-hop or not. If one of the NIDs of the router is down and the avoid_asym_router_failure is set, then that router is no longer used.


LNet Dynamic Configuration

LNet can be configured dynamically using the lnetctl utility. The lnetctl utility can be used to initialize LNet without bringing up any network interfaces. This gives flexibility to the user to add interfaces after LNet has been loaded. In general the lnetctl format is as follows: lnetctl cmd subcmd [options] The following configuration items are managed by the tool:

  • Configuring/Unconfiguring LNet
  • Adding/Removing/Showing Networks
  • Adding/Removing/Showing Peers
  • Adding/Removing/Showing Routes
  • Enabling/Disabling routing
  • Configuring Router Buffer Pools


Configuring/Unconfiguring LNet

After LNet has been loaded via modprobe (modprobe lnet), the lnetctl utility can be used to configure LNet without bringing up networks that are specified in the module parameters. It can also be used to configure network interfaces specified in the module parameters by providing the --all option.

// To configure LNet
lnetctl lnet configure [--all]
  
// To unconfigure LNet
lnectl lnet unconfigure 


Adding/Removing/Showing Networks

Now LNet is ready to be configured with networks to be added. To add an o2ib1 LNet network on ib0 and ib1 interfaces:

lnetctl net add --net o2ib1 --if ib0,ib1 

Using the show subcommands, it is possible to review the configuration:

lnetctl net show -v
net:
    - net type: lo
      local NI(s):
        - nid: [email protected]
          status: up
          statistics:
              send_count: 0
              recv_count: 0
              drop_count: 0
          tunables:
              peer_timeout: 0
              peer_credits: 0
              peer_buffer_credits: 0
              credits: 0
          lnd tunables:
          tcp bonding: 0
          dev cpt: 0
          CPT: "[0]"
    - net type: o2ib1
      local NI(s):
        - nid: [email protected]
          status: up
          interfaces:
              0: ib0
          statistics:
              send_count: 0
              recv_count: 0
              drop_count: 0
          tunables:
              peer_timeout: 180
              peer_credits: 8
              peer_buffer_credits: 0
              credits: 256
          lnd tunables:
              peercredits_hiw: 64
              map_on_demand: 32
              concurrent_sends: 256
              fmr_pool_size: 2048
              fmr_flush_trigger: 512
              fmr_cache: 1
          tcp bonding: 0
          dev cpt: -1
          CPT: "[0]"
        - nid: [email protected]
          status: up
          interfaces:
              0: ib1
          statistics:
              send_count: 0
              recv_count: 0
              drop_count: 0
          tunables:
              peer_timeout: 180
              peer_credits: 8
              peer_buffer_credits: 0
              credits: 256
          lnd tunables:
              peercredits_hiw: 64
              map_on_demand: 32
              concurrent_sends: 256
              fmr_pool_size: 2048
              fmr_flush_trigger: 512
              fmr_cache: 1           
          tcp bonding: 0
          dev cpt: -1
          CPT: "[0]" 

To delete a net:

lnetctl net del --net o2ib1 --if ib0 

Resulatnt configuration would be like:

lnetctl net show -v
net:
    - net type: lo
      local NI(s):
        - nid: [email protected]
          status: up
          statistics:
              send_count: 0
              recv_count: 0
              drop_count: 0
          tunables:
              peer_timeout: 0
              peer_credits: 0
              peer_buffer_credits: 0
              credits: 0
          lnd tunables:
          tcp bonding: 0
          dev cpt: 0
          CPT: "[0]"
    - net type: o2ib1
      local NI(s):
        - nid: [email protected]
          status: up
          interfaces:
              0: ib0
          statistics:
              send_count: 0
              recv_count: 0
              drop_count: 0
          tunables:
              peer_timeout: 180
              peer_credits: 8
              peer_buffer_credits: 0
              credits: 256
          lnd tunables:
              peercredits_hiw: 64
              map_on_demand: 32
              concurrent_sends: 256
              fmr_pool_size: 2048
              fmr_flush_trigger: 512
              fmr_cache: 1
          tcp bonding: 0
          dev cpt: -1
          CPT: "[0]" 


Adding/Removing/Showing Peers

lnetctl command to add remote peers:

lnetctl peer add --prim_nid [email protected] --nid [email protected] 

Verify the added peers configuration:

lnetctl peer show -v
peer:
    - primary nid: [email protected]
      Multi-Rail: True
      peer ni:
        - nid: [email protected]
          state: NA
          max_ni_tx_credits: 8
          available_tx_credits: 8
          min_tx_credits: 7
          tx_q_num_of_buf: 0
          available_rtr_credits: 8
          min_rtr_credits: 8
          refcount: 1
          statistics:
              send_count: 2
              recv_count: 2
              drop_count: 0
        - nid: [email protected]
          state: NA
          max_ni_tx_credits: 8
          available_tx_credits: 8
          min_tx_credits: 7
          tx_q_num_of_buf: 0
          available_rtr_credits: 8
          min_rtr_credits: 8
          refcount: 1
          statistics:
              send_count: 1
              recv_count: 1
              drop_count: 0 

To delete a peer:

lnetctl peer del --prim_nid [email protected] --nid [email protected] 


Adding/Removing/Showing Routes

To add a routing path:

lnetctl route add --net o2ib2 --gateway [email protected] --hop 2 --prio 1 

Verify the added route information:

lnetctl route show --verbose
route:
 - net: o2ib2
   gateway: [email protected]
   hop: 2
   priority: 1
   state: down 

To make this configuration permanent it is necessary to create a YAML file under /etc/sysconfig/lnet.conf. We can then import/export the live configuration file:

 
lnetctl export > /etc/sysconfig/lnet.conf 

The lnet script in /etc/init.d/ is compatible with DLC and should be enabled to be started at boot time:

systemctl enable lnet 


Enabling/Disabling Routing

To enable/disable routing on a node:

// To enable routing
lnetctl set routing 1
  
// To disable routing
lnetctl set routing 0 

To check the routing information:

lnetctl routing show 


Configure Router Buffers

The router buffers are used to buffer messages which are being routed to other nodes. These buffers are shared by all CPU partitions, so their values should be decided by keeping in consideration the number of CPTs in the the system. The minimum value of these buffers per CPT are:

#define LNET_NRB_TINY_MIN     512
#define LNET_NRB_SMALL_MIN    4096
#define LNET_NRB_LARGE_MIN    256 
A value greater than 0 for router buffers will set the number of buffers accordingly. A value of 0 will reset the router buffers to system defaults. To configure router buffer pools using lnetctl:

// VALUE must be greater than or equal to 0
  
lnetctl set tiny_buffers 4096
lnetctl set small_buffers 8192
lnetctl set large_buffers 2048 


ARP flux issue for MR node

Due to Linux routing quirks, if there are two network interfaces on the same node, the HW address returned in the ARP for a specific IP might not necessarily be the one for the exact interface being ARPed. This causes problems for o2iblnd, because it resolves the address using IPoIB, and gets the wrong Infiniband address. To get around this problem, it is required to setup routing entries and rules to tell the linux Kernel to respond with the correct HW address. Below is an example setup on a node with 2 Ethernet, 2 MLX and 2 OPA interfaces.

  • OPA interfaces:
    • ib0: 192.168.1.1
    • ib1: 192.168.2.1
  • MLX interfaces:
    • ib2: 172.16.1.1
    • ib3: 172.16.2.1
#Setting ARP so it doesn't broadcast
sysctl -w net.ipv4.conf.all.rp_filter=0
sysctl -w net.ipv4.conf.ib0.arp_ignore=1
sysctl -w net.ipv4.conf.ib0.arp_filter=0
sysctl -w net.ipv4.conf.ib0.arp_announce=2
sysctl -w net.ipv4.conf.ib0.rp_filter=0
  
sysctl -w net.ipv4.conf.ib1.arp_ignore=1
sysctl -w net.ipv4.conf.ib1.arp_filter=0
sysctl -w net.ipv4.conf.ib1.arp_announce=2
sysctl -w net.ipv4.conf.ib1.rp_filter=0
  
sysctl -w net.ipv4.conf.ib2.arp_ignore=1
sysctl -w net.ipv4.conf.ib2.arp_filter=0
sysctl -w net.ipv4.conf.ib2.arp_announce=2
sysctl -w net.ipv4.conf.ib2.rp_filter=0
  
sysctl -w net.ipv4.conf.ib3.arp_ignore=1
sysctl -w net.ipv4.conf.ib3.arp_filter=0
sysctl -w net.ipv4.conf.ib3.arp_announce=2
sysctl -w net.ipv4.conf.ib3.rp_filter=0
  
ip neigh flush dev ib0
ip neigh flush dev ib1
ip neigh flush dev ib2
ip neigh flush dev ib3
   
echo 200 ib0 >> /etc/iproute2/rt_tables
echo 201 ib1 >> /etc/iproute2/rt_tables
echo 202 ib2 >> /etc/iproute2/rt_tables
echo 203 ib3 >> /etc/iproute2/rt_tables

ip route add 192.168.0.0/16 dev ib0 proto kernel scope link src 192.168.1.1 table ib0
ip route add 192.168.0.0/16 dev ib1 proto kernel scope link src 192.168.2.1 table ib1
ip rule add from 192.168.1.1 table ib0
ip rule add from 192.168.2.1 table ib1
   
ip route add 172.16.0.0/16 dev ib2 proto kernel scope link src 172.16.1.1 table ib2
ip route add 172.16.0.0/16 dev ib3 proto kernel scope link src 172.16.2.1 table ib3
ip rule add from 172.16.1.1 table ib2
ip rule add from 172.16.2.1 table ib3
ip route flush cache 


LNet Tuning

LNet tuning is possible by passing parameters to the Lustre Network Driver (LND). The Lustre Network Driver for RDMA is the ko2iblnd kernel module. This driver is used both for Intel OPA cards and InfiniBand cards. With Lustre 2.10.X, network card can be detected and, using the /usr/sbin/ko2iblnd-probe, tunable parameters are set for supported cards. Intel OPA and Intel True Scale cards are automatically detected and configured by the script to achieve optimal performance with Lustre. The script can be modified to detect other network cards and set optimal parameters. The following is an example parameter settings in the configuration file (/etc/modprobe.d/ko2iblnd.conf) for a Lustre peer.

alias ko2iblnd-opa ko2iblnd
options ko2iblnd-opa peer_credits=128 peer_credits_hiw=64 credits=1024 concurrent_sends=256 ntx=2048 map_on_demand=32 fmr_pool_size=2048 fmr_flush_trigger=512 fmr_cache=1 

Parameters from /etc/modprobe.d/ko2iblnd.conf are global and cannot be different for different fabrics. It is thus recommended to use lnetctl command utility to specify it correctly for each fabric and then store the configuration into /etc/sysconfig/lnet.conf to make it persist across reboots. Previously all peers (compute nodes, LNet router, servers) on the network required identical tunable parameters for LNet to work independently from the hardware technology used (Intel OPA or InfiniBand). So, If you are routing into a fabric with older Lustre nodes, these parameters must be updated to apply identical options to the ko2iblnd module. LNet uses peer credits and network interface credits to send data through the network in the chunks of a fixed size of 1MB. The peer_credits tunable parameter manages the number of concurrent sends to a single peer and can be monitored using the /proc/sys/lnet/peers interface. The number for peer_credits can be increased using a module parameter for the specific Lustre Network Driver (LND):

//The default value is 8
ko2iblnd-opa peer_credits=128 

It is not always mandatory to increase peer_credits to obtain good performance, because in very large installation, an increased value can overload the network and increase the memory utilization of the OFED stack. The tunable network interface credits (credits) limits the number of concurrent sends to a single network, and can be monitored using the proc/sys/lnet/nis interface. The number of network interface credits can be increased using a module parameter for the specific Lustre Network Driver (LND):

// The default value is 64 and is shared across all the CPU partitions (CPTs).
ko2iblnd-opa credits=1024 

Fast Memory Registration (FMR) is a technique to reduce memory allocation costs. In FMR, memory registration is divided in two phases:

  1. allocating resources needed by the registration and
  2. registering using resources obtained from the first step.

The resource allocation and de-allocation can be managed in batch mode, and as result, FMR can achieve a much faster memory registration. To enable FMR in LNet, the value for map_on_demand should be more than zero.

 // The default value is 0
ko2iblnd-opa map_on_demand=32 

Earlier, Fast Memory Registration was supported only by Intel OPA and Mellanox FDR cards based on the mlx4 driver, but not by Mellanox FDR/EDR cards based on the mlx5 driver. With Lustre 2.10.X release, Fast Reg memory registration support for MLX5 (v2.7.16) is added.

Table 1: Lustre Suggested Tunables for Intel OPA
Tunable Suggested Value Default Value
peer_credits 128 8
peer_credits_hiw 64 0
credits 1024 64
concurrent_sends 256 0
ntx 2048 512
map_on_demand 32 0
fmr_pool_size 2048 512
fmr_flush_trigger 512 384
fmr_cache 1 1

From the above table:

  • peer_credits_hiw sets high water mark to start to retrieve credits.
  • concurrent_sends is the number of concurrent hardware sends to a single peer.
  • ntx is the number of message descriptors allocated for each pool.
  • fmr_pool_size is the size of the FMR pool on each CPT.
  • fmr_flush_trigger is the number of dirty FMRs that triggers a pool flush.
  • fmr_cache should be set to non-zero to enable FMR caching.