Infiniband Bart Taylor
What it is InfiniBand™ Architecture defines a new interconnect technology for servers that changes the way data centers will be built, deployed and managed. By creating a centralized I/O fabric, InfiniBand Architecture enables greater server performance and design density while creating data center solutions that offer greater reliability and performance scalability. InfiniBand technology is based upon a channel-based switched fabric point-to- point architecture. --
History Infiniband is the result of a merger of two competing designs for an inexpensive high-speed network. Future I/O combined with Next Generation I/O form what we know as Infiniband. Future I/O was being developed by Compaq, IBM, and HP Next Generation I/O was being developed by Intel, Microsoft, and Sun Microsystems Infiniband Trade Association maintains the specification
The Basic Idea High speed, low latency data transport Bidirectional serial bus Switched fabric topology Several devices communicate at once Data transferred in packets that together form messages Messages are direct memory access, channel send/receive, or mulitcast Host Channnel Adapters (HCAs) are deployed on PCI cards
Main Features Low Latency Messaging: < 6 microseconds Highly Scalable: Tens of thousands of nodes Bandwidth: 3 levels of link performance 2.5 Gbps 10 Gbps 30 Gbps Allows multiple fabrics on a single cable Up to 8 virtual lanes per link No interdependency between different traffic flows
Physical Devices Standard copper cabling Max distance of 17 meters Fiber-optic cabling Max distance of 10 kilometers Host Channnel Adapters on PCI cards PCI, PCI-X, PCI-Express InfiniBand Switches 10Gbps non-blocking, per port Easily cascadable
Host Channel Adapters Standard PCI 133 MBps PCI MBps PCI-X 1066 MBps PCI-X MBps PCI-Express x1 5Gbps x4 20Gbps x8 40Gbps x16 80Gbps
DAFS Direct Access File System Protocol for file storage and access Data transferred as logical files, not physical storage blocks Transferred directly from storage to client Bypasses CPU and Kernel Provides RDMA functionality Uses the Virtual Interface (VI) architecture Developed by Microsoft, Intel, and Compaq in 1996
RDMA
TCP/IP Packet Overhead
Latency Comparison Standard Ethernet TCP/IP Driver –80 to 100 microseconds latency Standard Ethernet Dell NIC with MPICH over TCP/IP –65 microseconds latency Infiniband 4X with MPI Driver –6 microseconds Myrinet –6 microseconds Quadrics –3 microseconds
Latency Comparison
References Infiniband Trade Association - OpenIB Alliance - TopSpin - Wikipedia - O’Reilly - Sourceforge - infiniband.sourceforge.net Performance Comparison of MPI Implementations over InfiniBand, Myrinet and Quadrics. Computer and Information Science. Ohio State University. - nowlab.cis.ohio-state.edu/projects/mpi-iba/publication/sc03.pdf