[Make-wifi-fast] [Ecn-sane] D* tcp looks pretty good, on paper

Rodney W. Grimes 4bone at gndrsh.dnsmgr.net
Fri Jan 8 09:35:15 EST 2021


> This is through one of the last remaining cerowrt boxes in the world,
> running fq_codel. tcp-davis takes about a 20% single stream throughput
> hit vs bbr.
> 
> I note, that I don't care one whit about throughput anymore. I care
> that nothing, NOTHING messes up my videoconference...

IMHO it is a fools errand to place all priorty on one of throughput
or delay, but it is reasonable to slightly bias the situation to
defer a bit over throughput if a stable lower delay is gained.  The
lost of some throughput can often be made up with additional capacity,
but nothing can increase the speed of light.

> and thus the tcp-rtt stats attached for davis are pleasing.

A questions/comment inline below about this data.

> 
> On Thu, Jan 7, 2021 at 12:26 PM Bob McMahon <bob.mcmahon at broadcom.com> wrote:
> >
> > FYI, one can try this out using iperf 2.1 with --trip-times.  This gives end/end delay at the application level. One can use --trip-times when clocks
> > are sync'd to get the write to read latencies which are the latencies at the application level.
> >
> > Note: I set up a Raspberry Pi 4 with a GPS hat from ubutronics for solderless pulse per second.  Then configured it as a PTP grandmaster.  This cost me around $200.

Ok, so that is the clock source node, could we get a better description of the network topology and end node hardware?

> > I also added support for a very crude --near-congestion option that paces the writes based upon the weight of the RTT.  The tcp_info struct is sampled and available
> > for other experiments though one would have to modify the source a bit. This current technique used by iperf 2.1 is designed for test networks only where all
> > traffic is under script control. We've had too many people measuring bloat as latency. We really need separate measurements between the two phenomena,
> > bloat vs latency, because they require different eng actions for a semiconductor supplier.
> >
> > Below are examples over a 10G link, first with no write pacing then with it. The server output, shown first, has the latency data (as well as the net power
> > and little's law calculation.)  (Note: use --histograms for to get full distributions.)

Is this network a 3 node physical dumb-bell, or is this in Netem or is this something more complicated?
What is the inherient delay path of this network?  Aka, what is D in the BDP, given the near saturation and the small window my guess is this is a simple 3 node dumb bell, but would like confirmation of that.

> >
> > No write pacing
> >
> > [rjmcmahon at localhost iperf2-code]$ src/iperf -s -i 1 -e
> > ------------------------------------------------------------
> > Server listening on TCP port 5001 with pid 24568
> > Read buffer size:  128 KByte (Dist bin width=16.0 KByte)
> > TCP window size: 85.3 KByte (default)
> > ------------------------------------------------------------
> > [  1] local 192.168.1.10%enp2s0 port 5001 connected with 192.168.1.62 port 50056 (MSS=1448) (trip-times) (sock=4) (peer 2.1.0-rc) on 2021-01-07 12:11:04 (PST)
> > [ ID] Interval        Transfer    Bandwidth    Burst Latency avg/min/max/stdev (cnt/size) inP NetPwr  Reads=Dist
> > [  1] 0.00-1.00 sec  1.09 GBytes  9.34 Gbits/sec  2.959/1.180/3.681/0.388 ms (8905/131072) 3.31 MByte 394522  18480=2459:2580:2475:2354:2203:2192:1974:2243
> > [  1] 1.00-2.00 sec  1.10 GBytes  9.41 Gbits/sec  2.993/2.302/3.703/0.359 ms (8978/131072) 3.36 MByte 393209  19482=2526:2850:3102:2622:2344:2297:1867:1874
> > [  1] 2.00-3.00 sec  1.10 GBytes  9.42 Gbits/sec  3.010/2.302/3.692/0.347 ms (8978/131085) 3.38 MByte 391047  19387=2563:2757:2928:2708:2432:2244:1829:1926
> > [  1] 3.00-4.00 sec  1.10 GBytes  9.41 Gbits/sec  3.009/2.301/3.668/0.348 ms (8979/131060) 3.38 MByte 391094  18821=2456:2585:2660:2545:2270:2239:1906:2160
> > [  1] 4.00-5.00 sec  1.10 GBytes  9.42 Gbits/sec  2.985/2.299/3.696/0.359 ms (8979/131070) 3.35 MByte 394295  19441=2509:2886:2959:2728:2336:2200:1971:1852
> > [  1] 5.00-6.00 sec  1.10 GBytes  9.41 Gbits/sec  2.977/2.258/3.671/0.363 ms (8978/131082) 3.34 MByte 395352  18509=2352:2602:2464:2380:2263:2142:2095:2211
> > [  1] 6.00-7.00 sec  1.10 GBytes  9.41 Gbits/sec  2.980/2.290/3.680/0.363 ms (8978/131072) 3.34 MByte 394873  18522=2407:2499:2565:2334:2213:2268:1999:2237
> > [  1] 7.00-8.00 sec  1.10 GBytes  9.42 Gbits/sec  2.980/2.253/3.702/0.362 ms (8979/131073) 3.35 MByte 394972  18615=2427:2592:2493:2460:2281:2057:2062:2243
> > [  1] 8.00-9.00 sec  1.10 GBytes  9.41 Gbits/sec  2.976/2.277/3.663/0.364 ms (8979/131065) 3.34 MByte 395443  18632=2338:2615:2647:2351:2192:2317:2063:2109
> > [  1] 9.00-10.00 sec  1.10 GBytes  9.41 Gbits/sec  2.976/2.293/3.690/0.366 ms (8978/131076) 3.34 MByte 395416  18428=2281:2622:2497:2275:2178:2253:2129:2193
> > [  1] 0.00-10.00 sec  11.0 GBytes  9.41 Gbits/sec  2.984/1.180/3.703/0.362 ms (89736/131072) 3.35 MByte 394014  188367=24320:26609:26793:24757:22712:22211:19916:21049
> >
> >
> > [rjmcmahon at localhost iperf2-code]src/iperf -c 192.168.1.10 --trip-times -i 1 -e
> > ------------------------------------------------------------
> > Client connecting to 192.168.1.10, TCP port 5001 with pid 18961 (1 flows)
> > Write buffer size: 131072 Byte
> > TCP window size: 85.0 KByte (default)
> > ------------------------------------------------------------
> > [  1] local 192.168.1.62%enp2s0 port 50056 connected with 192.168.1.10 port 5001 (MSS=1448) (trip-times) (sock=3) (ct=0.41 ms) on 2021-01-07 12:11:04 (PST)
> > [ ID] Interval        Transfer    Bandwidth       Write/Err  Rtry     Cwnd/RTT        NetPwr
> > [  1] 0.00-1.00 sec  1.09 GBytes  9.37 Gbits/sec  8937/0          0     1508K/1099 us  1065750
> > [  1] 1.00-2.00 sec  1.10 GBytes  9.41 Gbits/sec  8975/0          0     1508K/1087 us  1082218
> > [  1] 2.00-3.00 sec  1.10 GBytes  9.41 Gbits/sec  8975/0          0     1508K/1081 us  1088225
> > [  1] 3.00-4.00 sec  1.10 GBytes  9.42 Gbits/sec  8984/0          0     1508K/1085 us  1085300
> > [  1] 4.00-5.00 sec  1.10 GBytes  9.42 Gbits/sec  8980/0          0     1508K/1105 us  1065182
> > [  1] 5.00-6.00 sec  1.10 GBytes  9.41 Gbits/sec  8975/0          0     1582K/1100 us  1069428
> > [  1] 6.00-7.00 sec  1.10 GBytes  9.42 Gbits/sec  8979/0          0     1582K/1121 us  1049862
> > [  1] 7.00-8.00 sec  1.10 GBytes  9.41 Gbits/sec  8976/0          0     1582K/1133 us  1038396
> > [  1] 8.00-9.00 sec  1.10 GBytes  9.41 Gbits/sec  8978/0          0     1582K/1115 us  1055394
> > [  1] 9.00-10.00 sec  1.10 GBytes  9.42 Gbits/sec  8986/0          0     1582K/1122 us  1049744
> > [  1] 0.00-10.00 sec  11.0 GBytes  9.41 Gbits/sec  89748/0          0     1582K/1122 us  1048294
> >
> >
> > With write pacing
> >
> > [rjmcmahon at localhost iperf2-code]$ src/iperf -s -i 1 -e
> > ------------------------------------------------------------
> > Server listening on TCP port 5001 with pid 24702
> > Read buffer size:  128 KByte (Dist bin width=16.0 KByte)
> > TCP window size: 85.3 KByte (default)
> > ------------------------------------------------------------
> > [  1] local 192.168.1.10%enp2s0 port 5001 connected with 192.168.1.62 port 50072 (MSS=1448) (trip-times) (sock=4) (peer 2.1.0-rc) on 2021-01-07 12:14:59 (PST)
> > [ ID] Interval        Transfer    Bandwidth    Burst Latency avg/min/max/stdev (cnt/size) inP NetPwr  Reads=Dist
> > [  1] 0.00-1.00 sec  1.08 GBytes  9.31 Gbits/sec  0.401/0.193/2.682/0.168 ms (8876/131084)  456 KByte 2904347  19868=3296:2404:2508:2797:3559:1778:1551:1975
> > [  1] 1.00-2.00 sec  1.10 GBytes  9.41 Gbits/sec  0.400/0.219/0.627/0.053 ms (8971/131071)  460 KByte 2937822  19117=3069:2267:2307:2510:3029:1824:1683:2428
> > [  1] 2.00-3.00 sec  1.09 GBytes  9.39 Gbits/sec  0.374/0.193/0.541/0.055 ms (8958/131060)  428 KByte 3143030  18942=2846:2423:2304:2417:2927:1831:1856:2338
> > [  1] 3.00-4.00 sec  1.09 GBytes  9.39 Gbits/sec  0.385/0.190/0.664/0.070 ms (8952/131072)  441 KByte 3050401  19248=3041:2175:2343:2749:3320:1805:1526:2289
> > [  1] 4.00-5.00 sec  1.09 GBytes  9.40 Gbits/sec  0.380/0.197/0.546/0.057 ms (8965/131075)  436 KByte 3095915  19959=3321:2398:2551:2738:3500:1840:1532:2079
> > [  1] 5.00-6.00 sec  1.09 GBytes  9.39 Gbits/sec  0.369/0.198/0.536/0.051 ms (8956/131072)  423 KByte 3177431  21060=3627:2456:2886:3189:4246:1813:1190:1653
> > [  1] 6.00-7.00 sec  1.09 GBytes  9.39 Gbits/sec  0.380/0.202/0.562/0.054 ms (8959/131077)  436 KByte 3086914  19263=3044:2338:2424:2505:3155:1809:1636:2352
> > [  1] 7.00-8.00 sec  1.09 GBytes  9.40 Gbits/sec  0.376/0.198/0.541/0.053 ms (8965/131061)  432 KByte 3122495  19137=3079:2303:2340:2455:3017:1822:1683:2438
> > [  1] 8.00-9.00 sec  1.10 GBytes  9.41 Gbits/sec  0.381/0.208/0.576/0.054 ms (8974/131073)  438 KByte 3083767  19162=3050:2269:2392:2486:3019:1891:1667:2388
> > [  1] 9.00-10.00 sec  1.09 GBytes  9.40 Gbits/sec  0.371/0.194/0.582/0.057 ms (8964/131070)  425 KByte 3169244  19143=3006:2411:2303:2462:3067:1744:1760:2390
> > [  1] 0.00-10.00 sec  10.9 GBytes  9.39 Gbits/sec  0.382/0.190/2.682/0.076 ms (89544/131072)  437 KByte 3074913  194908=31380:23444:24362:26308:32839:18161:16084:22330
> >
> >
> > [rjmcmahon at localhost iperf2-code]$ src/iperf -c 192.168.1.10 --near-congestion=0.05 --trip-times -i 1 -e
> > ------------------------------------------------------------
> > Client connecting to 192.168.1.10, TCP port 5001 with pid 19320 (1 flows)
> > Write buffer size: 131072 Byte
> > TCP near-congestion delay weight set to 0.0500
> > TCP window size: 85.0 KByte (default)
> > ------------------------------------------------------------
> > [  1] local 192.168.1.62%enp2s0 port 50072 connected with 192.168.1.10 port 5001 (MSS=1448) (trip-times) (sock=3) (ct=0.40 ms) on 2021-01-07 12:14:59 (PST)
> > [ ID] Interval        Transfer    Bandwidth       Write/Err  Rtry     Cwnd/RTT        NetPwr
> > [  1] 0.00-1.00 sec  1.08 GBytes  9.31 Gbits/sec  8881/0          0     1135K/373 us  3120427
> > [  1] 1.00-2.00 sec  1.10 GBytes  9.41 Gbits/sec  8971/0          0     1135K/391 us  3007281
> > [  1] 2.00-3.00 sec  1.09 GBytes  9.39 Gbits/sec  8958/0          0     1135K/331 us  3547260
> > [  1] 3.00-4.00 sec  1.09 GBytes  9.39 Gbits/sec  8952/0          0     1135K/288 us  4074155
> > [  1] 4.00-5.00 sec  1.09 GBytes  9.40 Gbits/sec  8965/0          0     1135K/301 us  3903855
> > [  1] 5.00-6.00 sec  1.09 GBytes  9.39 Gbits/sec  8955/0          0     1135K/414 us  2835144
> > [  1] 6.00-7.00 sec  1.09 GBytes  9.40 Gbits/sec  8961/0          0     1135K/470 us  2499013
> > [  1] 7.00-8.00 sec  1.09 GBytes  9.40 Gbits/sec  8964/0          0     1135K/350 us  3356941
> > [  1] 8.00-9.00 sec  1.10 GBytes  9.41 Gbits/sec  8973/0          0     1135K/472 us  2491756
> > [  1] 9.00-10.00 sec  1.09 GBytes  9.40 Gbits/sec  8964/0          0     1135K/402 us  2922710
> > [  1] 0.00-10.00 sec  10.9 GBytes  9.39 Gbits/sec  89547/0          0     1135K/402 us  2919642
> >
> > Bob
> >
> >
> > On Thu, Jan 7, 2021 at 11:22 AM Taran Lynn via Make-wifi-fast <make-wifi-fast at lists.bufferbloat.net> wrote:
> >>
> >> The source can be found at https://github.com/lambda-11235/tcp_davis .
> >>
> >> The code mentioned in the paper can be found under the tag "arxiv_2020". The current master branch has an additional stable mode that I was testing out.
> >>
> >>
> >> On Thu, Jan 7, 2021 at 10:35 AM Dave Taht <dave.taht at gmail.com> wrote:
> >>>
> >>> See: https://arxiv.org/pdf/2012.14996.pdf
> >>>
> >>> Things I really like:
> >>>
> >>> * they used flent
> >>> * Using "variance" as the principal signal. This is essentially one of
> >>> the great unpublished and unanalyzed improvements on the minstrel
> >>> algorithm as well
> >>> * Conventional ecn response
> >>> * outperforms bbr on variable links
> >>>
> >>> Only negative so far is I haven't found any published source to it. :(
> >>>
> >>> Otherwise a very promising start to a year.
> >>>
> >>> "The choice of feedback mechanism between delay and packet loss has
> >>> long been a point of contention in TCP congestion control. This has
> >>> partly been resolved, as it has become increasingly evident that delay
> >>> based methods are needed to facilitate modern interactive web
> >>> applications. However, what has not been resolved is what control
> >>> should be used, with the two candidates being the congestion window
> >>> and the pacing rate. BBR is a new delay based congestion control
> >>> algorithm that uses a pacing rate as its primary control and the
> >>> congestion window as a secondary control. We propose that a congestion
> >>> window first algorithm might give more desirable performance
> >>> characteristics in situations where latency must be minimized even at
> >>> the expense of some loss in throughput. To evaluate this hypothesis we
> >>> introduce a new congestion control algorithm called TCP D*, which is a
> >>> congestion window first algorithm that adopts BBR's approach of
> >>> maximizing delivery rate while minimizing latency. In this paper, we
> >>> discuss the key features of this algorithm, discuss the differences
> >>> and similarity to BBR, and present some preliminary results based on a
> >>> real implementation."
> >>>
> >>>
> >>>
> >>>
> >>> --
> >>> "For a successful technology, reality must take precedence over public
> >>> relations, for Mother Nature cannot be fooled" - Richard Feynman
> >>>
> >>> dave at taht.net <Dave T?ht> CTO, TekLibre, LLC Tel: 1-831-435-0729
> >>
> >> _______________________________________________
> >> Make-wifi-fast mailing list
> >> Make-wifi-fast at lists.bufferbloat.net
> >> https://lists.bufferbloat.net/listinfo/make-wifi-fast
> >
> >
> > This electronic communication and the information and any files transmitted with it, or attached to it, are confidential and are intended solely for the use of the individual or entity to whom it is addressed and may contain information that is confidential, legally privileged, protected by privacy laws, or otherwise restricted from disclosure to anyone else. If you are not the intended recipient or the person responsible for delivering the e-mail to the intended recipient, you are hereby notified that any use, copying, distributing, dissemination, forwarding, printing, or copying of this e-mail is strictly prohibited. If you received this e-mail in error, please return the e-mail to the sender, delete it from your computer, and destroy any printed copy of it.
> 
> 
> 
> -- 
> "For a successful technology, reality must take precedence over public
> relations, for Mother Nature cannot be fooled" - Richard Feynman
> 
> dave at taht.net <Dave T?ht> CTO, TekLibre, LLC Tel: 1-831-435-0729

[ Attachment, skipping... ]

[ Attachment, skipping... ]

[ Attachment, skipping... ]

[ Attachment, skipping... ]

> _______________________________________________
> Ecn-sane mailing list
> Ecn-sane at lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/ecn-sane
> 
-- 
Rod Grimes                                                 rgrimes at freebsd.org


More information about the Make-wifi-fast mailing list