A packet-switched data link technology specifically designed to be a cluster interconnect. - http://www.myri.com/ - http://www.myri.com/myrinet/overview/index.html - ANSI/VITA 26-1998 -- http://www.myri.com/open-specs/myri-vme-d11.pdf (final draft) - http://www.myri.com/open-specs/link-history/index.html (original spec) "Myrinet is a cost-effective, high-performance, packet-communication and switching technology that is widely used to interconnect clusters of workstations, PCs, servers, or single-board computers" MyriNet provides a full-duplex link. Links have flow and error control. MyriNet supports both copper and fibre. The latest version, M3, supports a data rate of 2 Gb/s in either direction. Apparently, this is an industry-standard speed, based on 2.5 Gbaud with a 8b/10b-encoding. MyriNet is physically point-to-point, and depends on switches to connect multiple hosts. The standard switch is a 16-port cut-through crossbar switch, which provides low-latency switching. There is also a 'Network in a Box' superswitch which supplies 128 ports, interconnected by a network component (referred to as a 'full-bisection' network component). Apparently, these can be interconnected to scale to tens of thousands of hosts, and can provide redundant paths. 'Switch ports and interface ports'? A novel feature of MyriNet is that the host interface is engineered to interact directly with userspace processes, so allowing applications to bypass the OS. This is somehow related to the host adaptor running a control program. The host adaptor is connected via PCI. MyriNet packets have a 16-bit payload type identifier. $ Myrinet packets may be of any length, and thus can encapsulate other types of packets, including IP packets, without an adaptation layer. Each packet is identified by type, so that a Myrinet, like an Ethernet, may carry packets of many types or protocols concurrently. Thus, Myrinet supports several software interfaces. Specialized software packages developed by Myrinet customers for cluster-computing applications achieve short-message latencies between UNIX user processes less than 5µs, and sustained, one-way, data rates approaching 2 Gbits/s. Benchmarks of Myricom's GM message-passing system and application-programming interface (API) show sustained, one-way, data rates of ~1.96 Gbits/s between UNIX user processes in different hosts, short-message latencies as low as 7µs, and very low host-CPU utilization. The GM system provides protected user-level access to the Myrinet (secure in multi-user, multiprogramming environments); reliable, ordered delivery of messages; network mapping and route computation; and other features that support robust and error-free communication. Other software interfaces such as MPI, VI, Sockets, and TCP/IP are layered efficiently over GM, and are available from Myricom and from third parties. $ Performance Metric PCI64B 133MHz RISC & Memory PCI64C 200MHz RISC & Memory Sustained one-way data rate for large messages 236 MByte/s 243 MByte/s Sustained two-way data rate for large messages 312 MByte/s 429 MByte/s Latency for short messages 9.9 µs 7.6 µs Host-CPU utilization per message send 0.28 µs 0.28 µs Host-CPU utilization per message receive 0.56 µs 0.56 µs CategoryNetwork