Chaintip performance analysis

[wmitsuda]

no concrete action planned atm, many ideas to be tested, this one is to track what have been attempted.

[wmitsuda]

ongoing experiments based on the assumption:

• reset DB state to latest of static files
• clean pagecache
• advance and profile a few blocks

build/bin/integration stage_exec --datadir $DATADIR --chain mainnet --reset
sync && sudo purge
build/bin/integration stage_exec --datadir $DATADIR --chain mainnet --sync.mode.chaintip --block=$STOPBLOCK --$PROFILEFLAGS

[wmitsuda]

initial CPU profiling shows decompressor.nextPos() as a hot path, but it looks more like a red herring.

attempted to do a couple of microoptimizations to it, but nothing significant. I'll look into usage patterns of callers to better understand what can be done.

[wmitsuda]

note to myself: use -focus "ExecBlockV3" into profiling to isolate the noise even on 1 block profiling.

[VBulikov]

yperbasis added Imp2 as Importance

[AskAlexSharov]

about nextPos(): it may be red because it’s actually where “page cache miss” happening. “.data” field usage.

You can read file twice: and see difference in profiler picture. Can use: “erigon snapshots uncompress file.kv > /dev/null”