Consensys/teku

investigate performance in p2p-subscribe-all-subnets

Closed this issue · 1 comments

it looks like throughput reduced post deneb, which is curious, and I'd like to understand if we're missing anything when subscribed to all subnets in deneb world.

on holesky on an 8 cpu node i had to adjust a few parameters and increase the heap size significantly when enabling p2p-subscribe-all-subnets. was -Xmx16g before it got to 2% GC . and even then we were basically at 100% cpu.

--Xp2p-batch-verify-signatures-max-threads=8 \
--Xp2p-batch-verify-signatures-queue-capacity=30000

Need more investigation on a 16 cpu instance, but i think we'll need to specify some minimum specs for this mode.
ref #8387