Tracing is costly

[deepfire]

@mrBliss reports a relatively high cost incurred by logging.

Description of his setup:

1. chain syncing between two local processes:

   one fully-synced "server" node (nothing to do but serve blocks) and one empty "client" node (nothing to do but download blocks)

2. log configuration from configuration/configuration-mainnet.yaml
3. two cases:
   • normal cardano-node, 50k blocks in 1 minute
   • cardano-node with all tracers set to nullTracer: 75k blocks in 1 minute

[deepfire]

@mrBliss, @dcoutts, how do you propose we should account for the non-framework costs of forcing the lazily-evaluated thunks of the traces?

It's pretty obvious that profiling needs to be done before any conclusions can be made, though..

[mrBliss]

@mrBliss, @dcoutts, how do you propose we should account for the non-framework costs of forcing the lazily-evaluated thunks of the traces?

It's pretty obvious that profiling needs to be done before any conclusions can be made, though..

Most of the things that are being traced are already evaluated to (almost) normal form, at least on consensus side. Related: IntersectMBO/ouroboros-network#1386

[CodiePP]

this PR #494 defines the set of tracers which are on per default.
the amount of logged information can be controlled by either turning on/off a tracer and setting the minSeverity filter.

[CarlosLopezDeLara]

This issue should have been closed after IntersectMBO/ouroboros-network#1386 and IntersectMBO/ouroboros-consensus#739