xpra icon
Bug tracker and wiki

This bug tracker and wiki are being discontinued
please use https://github.com/Xpra-org/xpra instead.


Changes between Initial Version and Version 1 of Ticket #2022, comment 15


Ignore:
Timestamp:
11/04/18 18:17:27 (2 years ago)
Author:
Antoine Martin
Comment:

Legend:

Unmodified
Added
Removed
Modified
  • Ticket #2022, comment 15

    initial v1  
    88
    99Thanks again!
    10 
    11 Replying to [comment:14 Antoine Martin]:
    12 > As per previous comments, '''do not''' disable the other video encoders.
    13 >
    14 > What I found in the logs despite that:
    15 > * {{{DPI set to 45 x 25 (wanted 96 x 96)}}}
    16 > You should be using the patched dummy driver from the repo.
    17 > * we don't find a video encoder for small areas (because you disabled the other video encoders):
    18 > {{{
    19 > video size 484x92 out of range for nvenc, min 128x128
    20 > (..)
    21 > Error: no video pipeline options found for BGRX at 485x93
    22 > }}}
    23 > * your card is limited to 2 contexts since you're using a gamer card without a license key:
    24 > {{{
    25 > CUDA 9.2.0 / PyCUDA 2018.1, found 1 device:
    26 >   + GeForce GTX 1060 6GB @ 0000:AF:00.0 (memory: 98% free, compute: 6.1)
    27 > (..)
    28 > setup_pipeline failed for (84, (1, 1), None, 0, 0, None, 'BGRX', (1, 1), 520, 188, video_spec(nvenc)): cannot open encoding session, context is NULL
    29 > }}}
    30 > And that causes failures when we reach the limit. 2 contexts should be enough if you only have one "video" window though.
    31 > When we do hit this limit, things go downhill fast:
    32 > {{{
    33 > setup_pipeline(..) failed! took 2197.59ms
    34 > }}}
    35 > That's the encoding pipeline blocked for over 2 seconds! ouch.
    36 > * the batch delay is already high because of this:
    37 > {{{
    38 > update_batch_delay: congestion                   : 1.87,8.74  {}
    39 > }}}
    40 > You may want to turn off bandwidth detection to see if that helps with {{{--bandwidth-detection=no}}}. (#1912)
    41 > Or post the {{{-d bandwidth}}} log output so we can fix this.
    42 > But then it also goes up because of this:
    43 > {{{
    44 > update_batch_delay: compression-work-queue       : 2.64,5.12  {u'target': 1000, u'weight_multiplier': 3648, u'smoothing': u'logp', u'aim': 250, u'aimed_avg': 5247, u'div': 1000, u'avg': 12057, u'recent': 13314}
    45 > }}}
    46 > That's nvenc hogging the compresion work queue because it is slow to setup, and even worse when it hits the 2 context limit and fails (2 seconds as per above)... not much we can do about this.
    47 > * the video context seems to get recycled too quickly, then we see:
    48 > {{{
    49 > check_pipeline('auto', 451, 134, 'BGRX') setting up a new pipeline as check failed - encodings=['h264', 'h265']
    50 > }}}
    51 > why it failed is not logged, so I can't be sure why, but here are some possibilities:
    52 > * r20914: a refresh issue fixed recently (ticket:1816#comment:14)
    53 > * r20926: a bug fixed today causing unnecessary video context churn
    54 >
    55 > r20927 also adds an env var you can use to turn on video context cleanup debugging:
    56 > {{{
    57 > XPRA_DEBUG_VIDEO_CLEAN=1 xpra start ...
    58 > }}}
    59 > This should tell us why the video context is being destroyed so quickly.
    60 >
    61 > There are new centos 7.5 builds here: [https://xpra.org/beta].