[gvisor] notable network performance degradation

[arno01]

Some providers have notable network performance degradation when using gVisor.

52 Mbit/s Down & 5.19 Mbit/s Up (measured with speedtest.py) while normally (with “runc” - default container runtime) we were getting 1058.92 Mbit/s Down & 924.30 Mbit/s Up.

@hydrogen18 found that

https://gvisor.dev/performance/applications.csv it seems their own benchmarks indicate performance is very poor (gVisor networking) however

$ TZ=UTC date; curl -s https://gvisor.dev/performance/applications.csv
Thu  5 May 14:04:57 UTC 2022
runtime,method,metric,result
runc,http.node,transfer_rate,3814.85
runc,http.node,latency,11.0
runc,http.node,requests_per_second,885.81
runc,http.ruby,transfer_rate,2874.38
runc,http.ruby,latency,18.0
runc,http.ruby,requests_per_second,539.97
runsc,http.node,transfer_rate,1615.54
runsc,http.node,latency,27.0
runsc,http.node,requests_per_second,375.13
runsc,http.ruby,transfer_rate,1382.71
runsc,http.ruby,latency,38.0
runsc,http.ruby,requests_per_second,259.75

Workaround: provider can disable gVisor by specifying --deployment-runtime-class none argument for akash provider run.