You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 15 Current »


The SBC can leverage Graphics Processing Units (GPUs) for increased transcoding capacity of T-SBC instances deployed in OpenStack cloud environments (Newton and above). GPU acceleration significantly increases transcoding capacity of virtual instances, which otherwise would have a limited scale. In many cases a GPU-accelerated solution performs better than specialized DSP hardware-based solutions. GPU-based solutions have the following benefits:

  • GPUs are ubiquitous, they are being offered as Commercial Off-the-Shelf (COTS) solutions by major hardware and cloud vendors.
  • GPU-based solutions leverage the steep rise in computing power delivered by increased industry investments in GPU technology.
  • GPUs have diverse applications. Unlike specialized DSP hardware, GPU devices can be reused for other applications.
  • GPUs can be leveraged by cloud-based virtual instances.

Note

G.722 Silence Suppression is not supported with GPU transcoding.

GPU Transcoding

Prior to SBC 8.0 release, for GPU-based solutions, the 

Unable to show "metadata-from": No such page "_space_variables"
could not offer more than one transcodable codec in the outgoing offer for following reasons:

  • Unlike CPU-based DSP that supports all the codecs, the GPU-based DSP supports specific codecs.
  • For GPU, if the configured codecs do not reside on the same DSP process, they cannot be used with a single DSP allocation.
  • Unaware of the cluster capabilities, the signaling SBC removes the codecs from the outgoing offer, which cannot be supported by the reserved DSP.


The

Unable to show "metadata-from": No such page "_space_variables"
 is enhanced to offer all the configured transcodable codecs that the 
Unable to show "metadata-from": No such page "_space_variables"
supports, in the outgoing offer and address the above GPU gaps.

The configured codecs include:

  • Codecs configured for transcoding
  • Codecs supported by same T-SBC instance that supports codecs received in the offer.

Supported Deployment Scenarios

GPU acceleration is currently supported on SBC SWe T-SBC instances on OpenStack (Newton and above). T-SBC is a component in Distributed SBC architecture that provides transcoding service.

GPU devices are attached to the SBC SWe instances through PCIe pass-through – a single GPU device can be used by only one instance at a time. The process of enabling PCIe pass-through in OpenStack is detailed later in this document. For performance considerations, NUMA locality of devices should be ensured.

 

NVIDIA GRID is not supported.

Supported GPU Devices

The following NVIDIA GPU devices are supported.

GPU Device NamePCI Vendor IDPCI Device IDReleases supportedRemarks
Tesla V100 PCIe 16GB10DE1DB47.0.0 onwardsrecommended
Tesla V100 SXM2 16GB10DE1DB17.2.0 onwards

Supported Codecs

  • AMR-NB
  • G729
  • G722
  • AMR-WB

In addition, G.711 is supported for GPU instances, but only when G.711 is being transcoded to a non-G.711 codec. You cannot currently configure transcoding from G.711 to G.711 on GPU instances. The coding rates and packetization times for the supported codecs are shown in the tables on the Audio Codecs page.


Note for Releases Prior to 8.0

The GPU transcoding solution currently does not support more than one non-G711 transcodable codec per leg on a trunk group. Therefore when configuring Packet Service Profiles, do not configure multiple non-G711 codecs on a single leg (This Leg/Other Leg parameters) when specifying the Codecs Allowed For Transcoding within Packet To Packet Control. Refer to Packet Service Profile - CLI or Packet To Packet Control - Codecs Allowed For Transcoding (EMA).

Note for Releases 8.x/9.x

Following issues exist in 8.x/9.x releases:

  • If  the first transcodable codec in the SDP offered by SBC is G.711, and the answer comes for non-G711 codec, then the SBC continues to use CPU resources for non-G711 transcoding.
  • If the first transcodable codec in the SDP offered by SBC is non-G711 and the answer comes for G711, then the SBC continues to use GPU resources for G.711 transcoding.


OPUS Encoding

OPUS on GPU supports the following:

  • Input sampling rates of 8Khz (NB) and 16Khz (WB) is supported on the IDP interface (encoder input and decoder output)
  • For the OPUS encoder, the only supported frame size is 20ms . 
  • However, the OPUS decoder supports decoding of 10ms, 20ms, 40ms and 60ms frames and packets of size 10ms through 60ms.
  • Encoding output bandwidths of 8Khz and 16Khz are supported for Opus .
  • Inband FEC encoding is supported.
  • Inband FEC decoding is supported except for 10ms frame size.
  • Variable Bit Rate (VBR) as well as Constant Bit Rate (CBR) is supported
  • DTX mode is supported.
  • Only single channel mode is supported
  • RFC2833 and OOB transmission and reception is supported on OPUS leg.


For the Opus Encoder, the following configs are supported.

  • Application   - VOIP mode
  • Complexity   - 3
  • Channels      - 1 ( single channel encoding)
  • Frame size   - 20ms 
  • cVbr              - 0 (disabled)
  • Mode            - SILK mode only


Feature Differences between CPU and GPU

Feature

CPU

GPU

Native Frame Size (for Encoder)

10ms

20ms

Packetization Times Supported

10ms, 20ms, 30ms, 40ms, 50ms, 60ms

20ms, 40ms and 60ms

Feature Comparison of CPU and GPU T-SBC Solutions

While GPU-based T-SBCs offer marked increase in scale when compared to CPU-based T-SBCs, there are some caveats with the GPU solution that are highlighted in the following table. 


Feature comparison of CPU and GPU T-SBC solution

Feature
CPU Solution
GPU Solution
ProvisioningCodecs are not provisioned during instantiation.

Codecs and their relative percentages are provisioned during instantiation using fields in the Heat template.

Inband Tone DetectionYesYes (except in G722 - AMRWB transcoding scenario)
RFC2833YesYes
Fax Tone DetectionYesNo
LRBT(TPAD)YesNo
G711 Silence SupressionYesYes

 



  • No labels