NetworkInfrastructure

From OACISS Systems Wiki
Revision as of 02:38, 28 November 2020 by Nic-wiki (talk | contribs)
Jump to navigation Jump to search

OACISS hostnames and speeds

OACISS hardware may have several network connections. If the fully qualified public hostname is 'foo.nic.uoregon.edu' for the host named "foo", then relevant DNS entries are

  • foo.nic.uoregon.edu - the public (128.223.202.x) Ethernet interface
  • foo.stor - the private (172.17.202.x) Ethernet interface
  • foo.ib - the IPoIB interface, if the host has Infiniband

Obviously .stor and .ib are not publicly recognized TLDs. They can only be resolved by OACISS' authoritative DNS server for 128.223.202.0/24, which is only accessible by the same subnet.

The following table describes at a glance the network line rates available host by host.

OACISS Network Speeds
Host Ethernet speed IB speed NFS mount source
Orthus 10G x Ethernet
Delphi 10G x Ethernet
Instinct 10G x Ethernet
Aurora 10G EDR IPoIB
Centaur 10G x Ethernet
Minotaur 10G x Ethernet
Cyclops 10G 2xEDR Ethernet
Gorgon 10G 2xEDR Ethernet
Eagle 10G 2xEDR [64G / pcie] Ethernet
Illyad 10G 2xEDR IPoIB
Gilgamesh 10G 2xEDR IPoIB
Godzilla 10G EDR Ethernet
Voltar 10G EDR IPoIB
Pegasus 10G EDR IPoIB
Vina 10G x Ethernet
Pike 10G x Ethernet
Compute: KNL Grover 1G x Ethernet

The NFS has a 40GbE network adapter to the private network and (if data is in RAM or on NVME scratch) can fill any 4 hosts' 10Gbit pipes. Infiniband hosts reliably achieve 12.5/25GBps in the osu_bw microbenchmark. Single-threaded speeds exceeding 25-30Gbps have been difficult to reach using IPoIB, but parallel iperf3 instances have reached speeds of around 75-80Gb/s.

Network Hardware

Ethernet

The NIC Ethernet setup has four switches, two per cabinet, one each for private/storage and public networks.

  • Dell N4032 x 2: Private network
    • Switches are stacked. R35 unit is master (#1) and R34 is failover/member (#2)
    • 40Gbps trunk link between racks (Fo1/1/1 and Fo2/1/1)
    • 40GbE to Mnemosyne (Fo1/1/2)
  • Dell PowerConnect 8024 + N4032: Public network
    • PC8024 in R34
      • 10Gbps UOnet uplink (Te1/0/1)
      • 2 ports LAG trunk to R35 (Te1/0/21 & 22)

Ethernet interfaces are associated with DNS records:

  • The public interface (e.g. FOO.nic.uoregon.edu)
  • The private interface with local-only DNS record (e.g. FOO.stor)

Oaciss switches.png

Infiniband network

OACISS has a 36 port SB7700 EDR Infiniband switch.

Hosts generally have either one or two EDR links to the switch, providing either 100 or 200Gbps connectivity.

Hosts with 200Gbps:

  • Cyclops
  • Gorgon
  • Eagle

Hosts with 100Gbps:

  • Aurora (sx-aurora)
  • Illyad
  • Gilgamesh
  • Pegasus
  • Godzilla
  • Mnemosyne

Hosts within the NIC network will resolve the domain '.ib' in reference to the IPoIB interfaces which provide ~80Gbps connectivity after the encapsulation penalty. Please note, there is no need to specify host.ib in order to have MPI utilize the IB interface; MPI will automatically detect and prefer the ultra-high-speed link unless it is explicitly disabled.

Planning is underway to connect all hosts to the Infiniband network at at least FDR speed.