TCPIP timeout handling improvement

[skrchnavy]

Improvement of #107

Code is simplified in calculation when timeout elapsed. (removed variable now)

poll interval calculation is improved.

Note: Poll interval is necessary on windows, because select is not 'broken' by signal. so waiting in select for example 10 seconds would block program for 10 seconds until signal (CTRL+C) processed. \ On linux there is no such issue with .

poll_ interval is calculated as 50% of previous poll interval (min is 50ms). \ Previous constant 10ms poll interval could increase CPU utilization.

When merged, #61 can be closed

[MatthieuDartiailh]

LGTM. Can I assume you checked this works as expected ?

[skrchnavy]

yes, i tested it on windows 7 and works as described.

With 10s timeout, it waited 2secs then ~1 sec, then ~0.5 sec, etc. and finished after 123 iterations, table shows time missing to timeout (measured after select unblocked)

8.0
6.99799990654
6.4960000515
6.24300003052
6.11599993706
6.05099987984
5.99900007248
5.94799995422
5.89700007439
==== CUT ====
0.148000001907
0.095999956131
0.0439999103546
-0.00799989700317

Last value shows that function finshed 8ms later then planed.

With 3s timeout, measurement is:

1.5
0.748000144958
0.371000051498
0.180999994278
0.085000038147
0.0329999923706
-0.0190000534058

finished 19ms after planed timeout.

[MatthieuDartiailh]

Thinking a bit more about this, I would like to see a small modification to provide a better support of short timeout (sometimes useful in debugging). Could you add a min_select_timeout = max(min(select_timeout/10, 50), 1) , and use it instead of 50, such that we get something closer to the expected timeout for timeout in the range 1 to 50 ms.

To fix the conflicts please rebase on master rather than merging it makes for a cleaner history.

[MatthieuDartiailh]

There are still some conflicts that needs to be solved, once it is done I will merge.

[skrchnavy]

I was not able locally rebase changes so I did the merge. I will do final change in calculation of select_timeout and min_select_timeout and then you can review / merge

[skrchnavy]

ready for final review. measurments (diff is difference between real timeout and expected timeout, it shall be <1% from expeted timeout):

• timeout=1s

  select_timout=0.25s
  select_timout=0.125s
  select_timout=0.0625s
  select_timout=0.03125s
  select_timout=0.015625s
  select_timout=0.01s
  diff=7.999897003173828ms

• timeout=2s

  select_timout=0.5s
  select_timout=0.25s
  select_timout=0.125s
  select_timout=0.0625s
  select_timout=0.03125s
  select_timout=0.02s
  diff=3.000020980834961ms

• timeout 10s

  select_timout=1.0s
  select_timout=0.5s
  select_timout=0.25s
  select_timout=0.125s
  select_timout=0.1s
  select_timout=0.1s
  select_timout=0.1s
      ...
  select_timout=0.1
  diff=95.99995613098145ms

[MatthieuDartiailh]

I just added a tiny bit more docs (in case somebody in the future think about reverting to only use select timeout). If you are happy with the change, you can trigger bors

bors delegate+

[bors[bot]]

:v: skrchnavy can now approve this pull request

[skrchnavy]

Agreed and approved

[MatthieuDartiailh]

bors r+

[bors[bot]]

Build succeeded

• continuous-integration/travis-ci/push