What does TNet make fast?
Scalable Systems Lab / The University of New Mexico
© Summer 2000 by Adrian Riedo
- Address Translation on NIC
- Table mapped support on NIC
- Burst Optimized Transmit and Receive Path on NIC
- Remote interrupt capability
- Cut-through routing in switch
- Ring support in switch doubles bisectional bandwidth
- MPI Broadcast is Multicast on network (Multicast support in switch)
- Micropackets - Variable sized packets (4 - 112 Byte payload)
- TNet + FCI + MPI designed by one team
- Bandwidth can be doubled by using 2.5 Gbit/s chips
- Latency can be halved by making FCI software more efficient
Current MPI Results on 32 Bit NIC: 100 MB/s Bandwidth & 14 ?s Latency
Expected MPI Performance on 64 Bit NIC: 180 MB/s & 10 ?s Latency