osdi20-panic.pdf

Background

• Pipeline-of-Offloads NICs

  • place multiple offloads in a pipeline to enable packets to be processed by a chain of functions
  • usually static chains of functions

    • need buffer or recirculation to bypass chains

      

• Manycore NICs

  • many tile-based cores that are repsonsible for different tasks

  • can’t send data between cores so must go through DMA memory before being processed by another core

  • difficult to isolate competing tenants

    • can statically partition core resources → inefficient

    

• RMT NICs

  • Match+Action table used

    • packet header is analyzed, then matched with specific service (offload)

    • can handle multiple tenants with good scheduling system

      • each offload runs at line-rate with extremely low latency

      

Design of PANIC

1. RMT pipeline parses packet header and matches them to chain of offloads
   1. also generates a PANIC descriptor for multi-tenant scenario and for priority scheduling
2. packet is injected to switching fabric (crowbar method)
   1. if does not need to be processed by offload, can just go to output port
   2. if need, send to central scheduler
3. if no load, packets are chained with a source route that takes them from offload to offload without stopping at packet scheduler
   1. push scheduling is used where scheduler push work to the offloads directly
4. When load is high, loaded unit detours a packet that was pushed to it off to the buffer in the central scheduler
   1. pulling scheduling is used where service units pull work once their pulled work is complete

RMT Pipeline:

• Service Chain: a list of services to send the packet to along with per-service metadata from the RMT
  • enables scheduler to perform dynamic load balancing across multiple computation units implementing the same service in parallel

High Performance Interconnect

• PANIC utilizes non-blocking, low latency, and high throughput crossbar interconnect network
  • crossbar interconnects do not scale well with more cores
  • could use flattened butterfly topology

Centralized Scheduler