Test Details Overview

Verify block signatures before taking the ledger writer (#5246)

* Benchmark block processing and the ledger writer contention it causes

A dev node with two thousand opened accounts processes rounds of
one-raw transfers between neighbouring accounts through the block
processor from the bootstrap source, with elections, backlog scans and
the bounded backlog switched off. A competing thread takes and
releases the ledger writer once a millisecond, as confirmation
processing would, and records how long each acquisition waited.

Reported per run: blocks per second, the competing writer's wait
percentiles and the number of times it got the lock.

* Verify block signatures before taking the ledger writer

The block processor used to check every signature inside the ledger
write transaction. It now resolves each block's signer under a read
transaction, verifies the batch in parallel and hands the verdicts to
the ledger, so the writer is held for the ledger updates alone and
other writers such as confirmation processing get it sooner.

nano::signed_message carries a signer, message and signature with
nothing verified. nano::signature_verification is the verdict on one
and mints itself through its static verify, the single producer; a
dummy exists for tests behind a dev-network assert, like vote_permit.

nano::signature_verifier shares a batch between the calling thread and
a pool of extra threads, sized by verification_threads under
[node.block_processor]. The unused signature_checker_threads key is
removed.

ledger::block_signer resolves the account the ledger will check a
block against, nullopt when the predecessor is missing. ledger::process
takes an optional verdict and reuses it only when it covers the exact
signer, hash and signature it would check; a verdict about another
block is a caller bug and asserts. The signer may legitimately differ
when an epoch block's predecessor is rolled back between the two
transactions, in which case the ledger verifies itself.

The verdict rides on block_context, so the batch loop stays a plain
range-for.

The block processing benchmark gains a verification_threads sweep.

__________

Type: commit

Hash: 6b49a6c44d3187ff8b590c7eb29b4ede72bf19ce

Created: today

__________

gr0v1ty/nano-node:6b49a6c44d3187ff8b590c7eb29b4ede72bf19ce

Test Results

Test Case Status Duration Median Deviation Log Graph
5n4pr_conf_10k_bintree PASS 112 s -1.0 s GitHub job logs Show
5n4pr_conf_10k_change PASS 128 s -6.0 s GitHub job logs Show
5n4pr_conf_change_dependant PASS 109 s -6.0 s GitHub job logs -
5n4pr_conf_change_independant FAIL 3 s -109.0 s Test log -
5n4pr_conf_send_dependant PASS 105 s -9.0 s GitHub job logs -
5n4pr_conf_send_independant PASS 129 s 15.0 s GitHub job logs -
5n4pr_rocks_10k_bintree PASS 113 s 0.0 s GitHub job logs -
5n4pr_rocks_10k_change FAIL 3 s -155.0 s Test log -