Designing accessible latency metrics Toke Hiland-Jrgensen 25th - - PowerPoint PPT Presentation
Designing accessible latency metrics Toke Hiland-Jrgensen 25th September 2013 1 / 6 Quantifying latency Bufferbloat is (getting) accepted in technical circles But mostly unknown to end-users Can be explained with some care
1 / 6
▶ Bufferbloat is (getting) accepted in technical circles ▶ But mostly unknown to end-users ▶ Can be explained with some care ▶ But how to quantify?
2 / 6
▶ The RRUL test
▶ Eight TCP streams to induce load ▶ Measure UDP and ICMP ping times ▶ Comparison using CDF plots
101 102 103 ms 0.0 0.2 0.4 0.6 0.8 1.0 Cumulative probability
Ping (ms) - fq_codel qdisc Ping (ms) - sfq qdisc Ping (ms) - codel qdisc Ping (ms) - pfifo_fast qdisc
You are here You could be here
3 / 6
▶ Minimum unloaded latency
▶ To where? ISP, nearest exchange, major site(s), E2E? ▶ From where? User device(s), CPE equipment? ▶ Uni/bidirectional? ICMP, UDP, load a full website?
▶ Latency under (saturated) load
▶ Needs reliable method to induce load ▶ Really hitting ”worst case” probably hard ▶ And what about outliers? ▶ Sampling frequency
▶ The ratio between the two
▶ A ”load degradation factor”? ▶ Logarithmic, linear, normalised? 4 / 6
▶ What is important for the user to know?
▶ An absolute measurement (ms) or a relative one (score)?
▶ Are bigger numbers better?
▶ Roundtrips/second rather than milliseconds of latency?
▶ Should minimum latency and degradation be combined
5 / 6
▶ Expose bufferbloat in the network ▶ Enable the consumer to influence latency/bandwidth
▶ Create incentives for improvement ▶ User (self-)information (like speedtest.net) ▶ ISP comparison charts ▶ Regulation requirements (e.g. bounds) ▶ QoS definition in contracts etc. ▶ Benchmarking in systems engineering
6 / 6