| 1 |
Fixing Performance |
| 2 |
------------------ |
| 3 |
|
| 4 |
1) Observations |
| 5 |
|
| 6 |
The following observations are made when considering the current (17/11/2004) |
| 7 |
problems in gstreamer. |
| 8 |
|
| 9 |
- startup time. |
| 10 |
Most of the time is spend in reading the xml registry. Several methods exist |
| 11 |
for not using a registry (see testsuite/plugin/). |
| 12 |
Since registries are pluggable, one can also write a binary registry which |
| 13 |
should significantly reduce the size and the load time. |
| 14 |
|
| 15 |
- Performance in state changes |
| 16 |
State changes are pretty heavy currently because negotiation happens at |
| 17 |
each state change. The PAUSED->PLAYING state change in particular is |
| 18 |
too heavy and should be made more 'snappy'. |
| 19 |
|
| 20 |
- Performance in data passing. |
| 21 |
Too much checking is done in the push/pull functions. Most of these checks |
| 22 |
can be performed more efficiently in the plugins, like checking if the |
| 23 |
element is sufficiently negotiated. The discont event 'invent' hack used |
| 24 |
for fixing the synchronisation has to go away. |
| 25 |
|
| 26 |
We also propose a get_range based method for pads so that random-access |
| 27 |
in file-based sources can happen more efficiently (typical for muxers). |
| 28 |
|
| 29 |
- Scheduling overhead. |
| 30 |
The current default opt scheduler has extensive data-structures and does |
| 31 |
uses fairly complicated algorithms for grouping and scheduling elements. |
| 32 |
This has performance impact on pipeline constructions and pipeline |
| 33 |
execution. |
| 34 |
|
| 35 |
- Events in dataflow. |
| 36 |
Because events are put inside the dataflow, a typical chain function has |
| 37 |
to check if the GstData object is an event of a buffer. |
| 38 |
|
| 39 |
|
| 40 |
2) Proposed solutions |
| 41 |
|
| 42 |
- startup time. |
| 43 |
Volunteers to implement a binary registry? |
| 44 |
|
| 45 |
- The changes listed in the separate document 'threading' will greatly |
| 46 |
reduce the number of negotiation steps during state changes. It will |
| 47 |
also reduce the latency between PAUSED<->PLAYING since it basically |
| 48 |
involves unlocking a mutex in the sinks. |
| 49 |
|
| 50 |
- adding return values to push/pull will offload some checks to the plugins |
| 51 |
that can more efficiently and more accuratly handle most error cases. |
| 52 |
With the new sync model in 'threading' it is also not needed to have |
| 53 |
the 'invented' events in the pull function. |
| 54 |
|
| 55 |
- The following scheduling policies should be made possible for pads: |
| 56 |
|
| 57 |
1) get_range <-> pull-range-based |
| 58 |
2) get <-> pull based |
| 59 |
3) push <-> chain-based |
| 60 |
|
| 61 |
The scheduling policies are listed in order of preference. A pad should |
| 62 |
export what scheduling it supports and the core will select what |
| 63 |
scheduling method to use. It is possible for a pad to support more than |
| 64 |
one scheduling method. |
| 65 |
|
| 66 |
It is possible that two elements cannot connect when they do not support |
| 67 |
compatible scheduling policies. We require that pull-based pads also |
| 68 |
support the chain based methods, at minimal, using a helper function to |
| 69 |
queue a buffer and that get-based pads also support a push based implementation. |
| 70 |
|
| 71 |
- With the changes listed in 'threading' the scheduler will be a lot more |
| 72 |
simple since it does not have to keep track of groups and connected |
| 73 |
elements. Starting and scheduling the pipeline will be a matter of |
| 74 |
starting the source/loop GstTasks and they will from there on take |
| 75 |
over. |
| 76 |
|
| 77 |
- move events out of the dataflow. Different methods will be used to |
| 78 |
send/receive events. This will also remove some of the checks in |
| 79 |
the push/pull functions. Note that plugins are still able to serialize |
| 80 |
buffers and events because they hold the streaming lock. |
| 81 |
|
| 82 |
3) detailed explanation |
| 83 |
|
| 84 |
TO BE WRITTEN |
| 85 |
|