udt
udt copied to clipboard
securesocketfunneling
Performance question.
I'm looking for sth like UDT, the original one has great performance based on my test. And I'm very glad to find this project base on asio. I try to test the performance, to see if it has similar performance as the original implementation. I think maybe it's not faire enough, just post what I did.
client/server don't have performance data, but the original udt has appclient/appserver which seems similar with client/server. So what I did is run server and appserver from my vps, and run appclient from local, all are Ubuntu 15.10, build by default, here is the result:
This result is appclient => server
SendRate(Mb/s) RTT(ms) CWnd PktSndPeriod(us) RecvACK RecvNAK
24.1335 99.676 1324 137.306 3 0
90.6163 98.639 2243 136.911 3 0
90.558 97.731 2679 136.537 2 0
91.0242 96.272 2816 135.98 3 0
91.3385 95.342 3082 135.612 2 0
91.688 94.133 3073 135.062 3 0
91.6764 93.274 3961 171 3 2
52.8253 93.17 4247 395 1 10
15.5245 93.17 4247 805 0 9
14.5115 93.17 4247 805 0 0
14.4419 93.17 4247 805 0 0
14.4531 93.17 4247 805 0 0
9.29345 93.17 4247 805 0 0
14.4525 93.17 4247 805 0 0
14.4409 93.17 4247 805 0 0
14.4525 93.17 4247 805 0 0
7.73286 93.17 4247 805 0 0
14.4525 93.17 4247 805 0 0
14.4409 93.17 4247 805 0 0
14.4526 93.17 4247 805 0 0
8.024 93.17 4247 805 0 0
14.4409 93.17 4247 805 0 0
14.4525 93.17 4247 805 0 0
13.2181 93.17 4247 805 0 0
6.76626 93.17 4247 805 0 0
14.4525 93.17 4247 805 0 0
14.4525 93.17 4247 805 0 0
13.7887 93.17 4247 805 0 0
9.47973 93.17 4247 805 0 0
14.4525 93.17 4247 805 0 0
14.4525 93.17 4247 805 0 0
11.0756 93.17 4247 805 0 0
12.6015 93.17 4247 805 0 0
14.453 93.17 4247 805 0 0
14.4417 93.17 4247 805 0 0
9.42143 93.17 4247 805 0 0
14.4409 93.17 4247 805 0 0
14.4644 93.17 4247 805 0 0
14.4293 93.17 4247 805 0 0
8.89791 93.17 4247 805 0 0
14.4411 93.17 4247 805 0 0
14.4525 93.17 4247 805 0 0
14.4525 93.17 4247 805 0 0
This is appclient => appserver
SendRate(Mb/s) RTT(ms) CWnd PktSndPeriod(us) RecvACK RecvNAK
34.4614 90.939 2101 1 42 0
312.943 104.957 105764 13 107 17
566.696 98.397 105163 14.5492 298 3
336.159 103.437 108948 30 216 9
515.817 104.418 112849 17.9674 495 10
532.556 93.949 103530 23.4924 445 5
453.818 108.273 115857 14 491 32
553.036 97.286 104946 15.2761 243 9
527.182 96.964 106919 21 325 8
370.735 95.995 105952 20.0181 235 1
572.541 104.304 112683 20 525 11
378.541 95.557 95754 19.3174 184 3
512.555 97.808 106675 15.4918 447 6
557.422 97.282 106693 20.2108 370 9
606.808 98.572 104215 17 496 11
644.919 104.724 108863 16 570 8
536.526 96.48 106222 16.1085 466 7
609.325 96.104 95083 18 511 7
654.918 109.585 118807 14 585 8
512.403 98.308 101933 19.1569 256 6
636.51 106.445 104975 15 591 5
572.702 105.457 114654 18 387 11
319.632 94.019 96389 30.0678 100 5
252.778 92.01 101011 31.8469 137 4
233.593 96.406 98043 23 210 3
431.089 96.021 105243 17.4637 251 4
528.022 101.97 109578 16.2454 330 7
612.842 108.271 117648 13 502 6
427.218 92.732 102354 37 199 9
509.328 96.64 95778 16.8195 390 3
502.551 95.952 105901 19.3726 373 11
616.464 105.089 103671 12.2732 504 7
598.754 108.013 113997 13 435 4
517.938 95.275 98832 21 255 8
465.443 95.221 101348 27 361 7
221.902 92.388 102005 49 67 9
229.576 89.751 95617 53.6216 193 5
478.182 99.121 108034 20 367 4
549.114 101.459 112186 16.175 465 4
405.513 92.486 101180 46 258 10
439.703 92.998 102394 26.4527 404 36
718.899 109.115 117012 9.02486 821 0
653.231 105.689 112807 9.433 581 6
571.427 108.307 114919 18 365 6
556.381 96.216 103107 18.9554 534 3
328.657 96.151 106017 22.0423 165 5
725.314 104.714 115571 9.10074 765 1
776.852 109.384 118583 7.22473 842 1
The original implementation performance is much better.
And when I run the same test second time without restart server side, the result is unbelievable. appclient => server
SendRate(Mb/s) RTT(ms) CWnd PktSndPeriod(us) RecvACK RecvNAK
0.28588 97.857 646 55555.6 2 0
0.244557 95.544 429 47619 3 0
0.291147 93.994 289 41666.7 3 0
0.326084 93.24 224 38461.5 2 0
0.337731 92.351 155 34482.8 3 0
0.384314 91.786 109 31250 3 0
0.419252 91.522 88 29411.8 2 0
0.454189 91.209 65 27027 3 0
0.465835 90.956 50 25000 3 0
0.512433 90.817 43 23809.5 2 0
0.547393 90.661 35 22222.2 3 0
0.5823 90.584 30 20833.3 3 0
0.59394 90.559 28 20000 2 0
0.652193 90.558 25 18867.9 3 0
0.675463 90.519 24 17857.1 3 0
0.687105 90.479 23 17241.4 2 0
0.722044 90.322 22 16393.4 3 0
0.698753 90.254 22 15625 3 0
appclient => appserver
SendRate(Mb/s) RTT(ms) CWnd PktSndPeriod(us) RecvACK RecvNAK
16.768 91.043 975 1 27 0
246.607 92.939 93489 6.99483 73 7
462.314 90.552 99320 12.8825 119 4
381.681 90.52 94368 22.7385 217 10
485.299 90.202 99519 38.5489 302 13
464.906 91.307 100365 15.3132 443 3
419.472 90.379 85311 57 150 81
202.348 89.526 92842 30.2625 133 0
481.028 89.825 89443 26.1014 378 7
649.977 92.387 96653 30 607 24
530.592 92.673 102103 26 415 7
439.689 89.936 95234 28 313 4
498.335 91.492 100308 20 481 8
146.068 91.828 101302 25 18 9
499.547 90.614 90035 22 415 4
507.244 91.391 96509 17 479 6
328.391 91.317 99491 20.1534 140 12
Should I suppose the performance will be better if we change the appclient to client? (As I said I can't test this since client doesn't have similar performance logs). Is it difficult to add performance logs to client?
Actually I modified client/server to send a file, the performance is not good too, say 1/10 of the original one.
Do we have any plan on performance tuning?
Hello,
Thanks for your feedback. It is really appreciated. Improving performance is indeed a priority for this project.
We had a workaround with timers since boost asio timers are not reliable with short duration. And we are still trying to figure it out what causes this performance drop but it may take some times since we are doing this on our free time. There may be a protocol implementation difference which causes this behaviour. Our library is still young and the protocol documentation is not as clear as expected on some element (ack, light ack for example). We had to dig in the original source code to determine some behaviour.
client/server don't have performance data
You can activate some logging on client or server by changing a template parameter on the protocol.
using udt_protocol = ip::udt<connected_protocol::logger::FileLog<1000>>;
This will log the internal protocol variables each seconds in a file name session_log*.log and the result can be parsed with a python script available in the tools directory.
@securesocketfunneling @thomas-yuan is this still an outstanding issue?
