[LibreQoS] [Rpm] cloudflare on a roll

rjmcmahon rjmcmahon at rjmcmahon.com
Tue Apr 18 14:05:29 EDT 2023

> https://blog.cloudflare.com/making-home-internet-faster/

I wonder if we're all still missing it a bit. We're complaining that 
internet providers are using speed applied to a rated link capacity and 
then we say to use latency or responsiveness in its place. It's like 
saying a road has a speed and a latency. The road really has neither as 
direct attributes. It's stationary as are waveguides. Sure, we can come 
up with a rating of link capacity and link delay per what's attached to 
those waveguides and we also need to add the highly variable "working 
conditions" in order to take a synthetic measurement.

Us now saying "speed is the wrong metric" so use network latency can be 
equally confusing and equally wrong, e.g. if the app thread is CPU 

I think it's the travel times that matter to the end users. But the user 
doesn't know their destinations, A to B so-to-speak, so instead of 
execution times, we need to find a metric to hint at the users awaiting 
their devices (helping engineers to mitigate and eliminate the dreaded 
indeterminate progress indicators which is sad way to spend device 

An indirect way of measuring travel times may be to measure the thread 
write delays. A thread will run as fast as possible (AFAP) when its 
threads don't block, e.g. on network i/o.

I've added support for --tcp-write-times in iperf 2. This gives the 
amount of time the thread's select() blocks awaiting on the ability to 
write (or, with linux, the amount of time awaiting the syscall write() 
to complete. This along with --tcp-write-prefetch sets TCP_NOTSENT_LOWAT 
should give an idea of the amount of time awaiting network availability 
by a thread.

[rjmcmahon at ryzen3950 iperf2-code]$ src/iperf -c mail.rjmcmahon.com 
--tcp-write-times --histograms=1m --tcp-write-prefetch 16K -i 1 -t4
Client connecting to mail.rjmcmahon.com, TCP port 5001 with pid 212310 
(1 flows)
Write buffer size: 131072 Byte (writetimer-enabled)
TCP congestion control using cubic
TOS set to 0x0 (Nagle on)
TCP window size: 85.0 KByte (default)
Event based writes (pending queue watermark at 16384 bytes)
Enabled write histograms bin-width=1.000 ms, bins=100000
[  1] local port 38538 connected with 
port 5001 (prefetch=16384) (sock=3) (icwnd/mss/irtt=14/1448/12335) 
(ct=12.45 ms) on 2023-04-18 10:52:02.956 (PDT)
[ ID] Interval        Transfer    Bandwidth       Write/Err  Rtry     
Cwnd/RTT        NetPwr  write-times avg/min/max/stdev (cnt)
[  1] 0.00-1.00 sec  5.25 MBytes  44.0 Mbits/sec  42/1          0     
360K/65206 us  84.43  24.173/13.509/39.347/4.196 ms (42)
[  1] 0.00-1.00 sec W8-PDF: 
[  1] 1.00-2.00 sec  4.75 MBytes  39.8 Mbits/sec  38/0          6     
173K/35105 us  142  26.079/22.403/39.766/4.142 ms (38)
[  1] 1.00-2.00 sec W8-PDF: 
[  1] 2.00-3.00 sec  4.88 MBytes  40.9 Mbits/sec  39/0          4     
100K/19518 us  262  25.673/22.276/35.668/2.602 ms (39)
[  1] 2.00-3.00 sec W8-PDF: 
[  1] 3.00-4.00 sec  5.00 MBytes  41.9 Mbits/sec  40/0          1     
101K/19337 us  271  25.073/14.430/35.911/2.864 ms (40)
[  1] 3.00-4.00 sec W8-PDF: 
[  1] 0.00-4.06 sec  20.0 MBytes  41.3 Mbits/sec  160/2         11     
103K/20126 us  257  25.230/13.509/39.766/3.563 ms (160)
[  1] 0.00-4.06 sec W8(f)-PDF: 



Indeterminate Progress
Use indeterminate mode for the progress bar when you do not know how 
long an operation will take. Indeterminate mode is the default for 
progress bar and shows a cyclic animation without a specific amount of 
progress indicated. The following example shows an indeterminate 
progress bar.

More information about the LibreQoS mailing list