|
Product Description VirtualShare is an intelligent, low latency cluster resource virtualization engine and accelerator incorporating all the network interface and switching functions of HyperShare-Native Torus networks for scalable cluster systems. Thanks to its PCI Express Gen-2 interface, VirtualShare is a universal product that brings HyperShare capability to any off-the-shelf x86 server systems. VirtualShare incorporates unique Torus management and PCI Express interface acceleration functions which deliver outstanding latency performance and operational flexibility. With am high-performance FPGA platform at its core, VirtualShare carries powerful advantages over ASIC-based NICs, as it can intelligently reconfigure its cluster memory map addressing on the fly as part of its dynamic adaptability to evolving cluster resource reconfiguration need and node fault conditions. Therefore, VirtualShare is not only a cluster networking “handler”, but a performance “optimizer” and fault condition “compensator” in an off-the-shelf, commodity level product package with no market equal. Overall, VirtualShare enables data center and HPC managers to tap into the most complete cost and power containment opportunities along with top performance levels that the HyperShare standard offers to the industry. Special Cluster Resource Virtualization Capabilities Any cluster node to directly access and utilize any other node’s resources - be they memory, accelerators, storage subsystems, I/O or special functions - as if they were its own. Any PCI Express-based subsystem of any node - accelerator, mass-storage or other - can directly communicate with any other PCI Express subsystem of any other node without involving the CPUs or the system memory of their respective nodes. Product Features - Non-Coherent Globally Distributes Shared Memory Addressing
- Hardware Native System Resource Virtualization
- Switchless Cluster Interconnect
- Torus Network Topologies Support
- Built-In Acceleration Functions
- Torus Acceleration Engine
- Reduced Network Oversubscription
- Hardware-based Broadcast and Multi-Cast Synchronization Barrier
- Multi-Path Packet Accelerations
- Posted, Non-Posted, Completion
- Application Communications Offloading
- Concurrent Compute-Communication Operations
- Host Memory-to-Host Memory Hardware Active Latency of 900nS with 64-byte data packets
- TCP/IP Processing for Cloud Computing, Data Base and all IP-based Network Software, Delivering 3.5 uS of Application-to-Application Latency, Including OS Overhead, OS Jitter and TCP/IP Processing (see graphics rendition below) - 10x Latency Performance Improvement vs. Ethernet IP Processing
Extensive Software Stack Support - Complete PCIe driver Support with DMA S&G Engine Control
- Low Level API for SPE Engine with RDMA Capabilities
- Complete API for RMA and RDMA Engine
- Configuration and Optimization Console for Hardware Tuning
- Optimized MPI Support through OpenMPI Components
- Optimized PGAS Support through GasNET
2D Torus Product Version - 8-Bit, 5 GHz Clock Rate PCI Express Gen-2 Interface
- 4x 8-Bit Wide HT Network Channels
- 12.8 Gb/s (25.6 Gb/s bidirectional) Bandwidth per HT Network Link
- 2x Transceivers per HT Network Link - 8x Transceivers Total
- Each Transceiver Operates at Full HT 3.1 Clock Rate of 3.2 GHz (6.4 Gb/s)
- Multiple Network Engines Featuring 6x Virtual Channels per HT Network Link
- 1x Virtex 6, 34x34-pin FPGA
- Super-Low Profile Form Factor - 137mm x 75mm/50mm for Easy Fit in Compact Server Systems
3D Torus Product Version - 8-Bit, 5 GHz Clock Rate PCI Express Gen-2 System Interface
- 6x 8-Bit Wide HT Network Channels
- 12.8 Gb/s (25.6 Gb/s bidirectional) Bandwidth per HT Network Link
- 2x Transceivers per HT Network Link - 12x Transceivers Total
- Each Transceiver Operates at Full HT 3.1 Clock Rate of 3.2 GHz (6.4 Gb/s)
- Multiple Network Engines Featuring 6x Virtual Channels per HT Network Link
- 1x Xilinx Virtex 6, 34x34-pin FPGA
- Super-Low Profile Form Factor - 167mm x 104mm/68mm for Easy Fit in Compact Server Systems
|
