discrepancy between max sessions on a transcoder node and NVENC sessions on GPU

[kyriediculous]

There is a discrepancy between what a session means for a transcoder node and what a session means on the GPU.

• A session on the node is a transcoding job and is for all renditions requested within that job

• A session on the GPU is per-rendition

This creates a discrepancy whereby a stream that has multiple requested renditions only takes up 1 session on the node, but could cap or go over the session limit on the GPU if more than 3 (for consumer grade HW) renditions are requested.

For the current jobs sent by the Livepeer API this isn't as problematic as it requests 3 renditions, but if there's ever jobs that require more renditions then the experience could quickly deteriorate.

Are there possibilities here for hardware that isn't patched? Transcoding in serial isn't an option because decoding would have to happen for each rendition requested which creates a bottleneck because most GPU's only have one single concurrent decoding session anyway.

Would it be possible however to "batch" renditions together and then do a second batch (which would include another decode) per set of 3 requested renditions.

From discord: