# Improve SCION router performance to meet target

[jiceatscion]

This item comprises the following tasks:

• ## Create necessary profiling, additional metrics, and other instrumentation to reveal router performance bottlenecks. #4412
• ## Design and implement router performance improvements. #4413
• ## Calculate end-to-end target platform router performance. #4414

[jiceatscion]

The deliverable of this work item is a router with satisfactory predicted performance. This is an open-ended task.

Profiling and other instrumentation have to be created ad-hoc to reveal opportunities for optimization, followed or interleaved with a cycle of implementing and testing improvements. The cycle stops when we agree that performance improvement is no-longer a priority.

[jiceatscion]

Observations from initial profiling and other measurements:

• The socket API had a constant and significant packet overhead. iperf3 reveals a minimum packet time of about 1 microsecond no-mater what the network device is: veth+bridge, Gig-e NIC, 10 Gig-e NIC, etc.
• Packet processing itself represents 11% of realtime execution in the router (excluding go runtime cost).
• with small (172 bytes) packets, 25% of the router time is spent in the kernel sending packets (excluding go runtime cost).
• with small (172 bytes) packets, 27% of the router time is spent in the kernel receiving packets (excluding go runtime cost).
• Another 30.81% is spent in the go runtime, mostly as:
  • Allocating/GCing: 2.88% (MsgHdrUnpack->ParseInetAddr, resolveLocalDst)
  • Chanrecv: 3.6%
  • Chansend: 5.62%
  • computeProcID: 2.37%
  • runtime.Schedule (excluding channel ops): 6.8%
  • mmsghdrs unpack: 2.45%
  • mmsghdrsPacker pack: 2.39%
  • runtime mapaccess1/2: 3.48%
  • metrics: 1.22%
  • time.Now: 1.54%
• The remaining 4.65% is small bits and pieces.