Clusters Institute: Intro to Networking

Bill Abbott Sysadmin – Office of Advanced Research Computing Rutgers University

August 2019 1 Background Survey

• Familiarity with… • Ethernet • InfiniBand /OmniPath • Fibre Channel • TCP/IP • Firewalls and DMZ • Switches and routers • Tools: • Basic: ip/ifconfig, netstat, ethtool, • Advanced: tcpdump, iptraf-ng, ntop • Performance: iperf3, bbcp,

August 2019 2 What are We Trying to Accomplish?

• Speed • Latency vs Bandwidth • Reliability • Accessibility • Public • Campus • Inside/outside • VPN • DMZ

August 2019 3 TCP/IP

Necessary Background Needed: • Public vs Private (RFC 1918) Addressing • 10.xx.xx.xx • 192.168.xx.xx • 172.16-31.xx.xx • Subnetting • Gateway(s) • DNS • IPv6

August 2019 4 A TCP/IP example

You work at Company A. You have to contact Company B’s web department to acquire some data. • You have to call on the phone but don’t know the number, just the company name. • You call 411 and ask, and they give you the number. • You call that number and ask for extension 80. • The person who answers says hello, and you say hello. • The person tells you to call extension 443 instead. • The person who answers says hello, and you say hello. • You request the data, and the person reads it to you. • You both say goodbye.

August 2019 5 Ethernet vs InfiniBand (IB) and OmniPath (OPA)

• Most clusters have both Ethernet and IB or OPA • Advantages and disadvantages of each • (Relatively) slow Ethernet for external access, file copying, etc. • IB and OPA for MPI • RDMA vs IP • Management (PXE, Lights-Out, IPMI, infrastructure) • OPA 200G project now terminated

August 2019 6 Internal/External Interfaces to the Cluster • Who has access? • Who provides access? • What internal resources can they see • VPN considerations • DMZ considerations • Internal data movement

August 2019 7 Requirements • High-Speed Interconnect • Low latency, high bandwidth • Reliable • User support (MPI, etc) • Internet/External Access • Bandwidth • Access to/from external sites • Security • Management • Cheap and reliable • PXE • IPMI • Secure (Isolation from users)

August 2019 8 Questions?

August 2019 9 Extra Time: Physical Layer

August 2019 10