Offloading distributed applications onto smartNICs using iPipe

https://homes.cs.washington.edu/~arvind/papers/ipipe.pdf

• Accelerate general purpose distributed applications?

• DMA: writing to and from the host memory

• Host communication

  • Use DMA

• Multi-core processors, low power

  • Extent of compute is limited (to offload)

• Network applications

  • RDMA, DPDK

• Packet arrive at the TX/RX ports

• Traffic manager

  • Abstractions between the ports and actual ports

• NIC cores handle all traffic on both the send and receive paths

  • Might not have computation to do

• Tight integration of computing and communication

• Computation requires additional overhead

• Shows how many cores are needed to fully utilize the available bandwidth

  • Max bandwidth = 10 Gbps

• 256B: not sufficient to process that many buckets

• Larger: saturated quickly

Take-away

• Quantifies the default forwarding tax of smartNics

• Dependent on packet size workload

Workload?

• No processing

• Actor: logic - workload has a lot of variants. That type of actor is going to move to DRR. But actor with more uniform distribution will be on FCFS cores.

  • Tail latency? How they measure that?

    • Window. So might not absolute, just relative.