bitcoin / bitcoin

Bitcoin Core integration/staging tree
https://bitcoincore.org/en/download
MIT License
79.45k stars 36.35k forks source link

Raise maximum -dbcache setting #28249

Closed Sjors closed 1 month ago

Sjors commented 1 year ago

Please describe the feature you'd like to see added.

nMaxDbCache was set to 16 GB back in 2015: https://github.com/bitcoin/bitcoin/commit/b3ed4236beb7f68e1720ceb3da15e0c3682ef629#diff-d102b6032635ce90158c1e6e614f03b50e4449aa46ce23370da5387a658342fd

The UTXO set has been growing significantly over the past few months so we should probably raise it.

Is your feature related to a problem, if so please describe it.

Doing IBD without flushing dbcache is faster. That said, I'm not sure how much one single flush matters.

Describe the solution you'd like

Make nMaxDbCache bigger.

Describe any alternatives you've considered

No response

Please leave any additional context

Much growth recently: https://statoshi.info/d/000000009/unspent-transaction-output-set?orgId=1&from=now%2Fy&to=now&refresh=5s&viewPanel=8

Sjors commented 1 year ago

cc @sipa

jonatack commented 1 year ago

There is some related discussion and testing in https://github.com/bitcoin/bitcoin/pull/6102 containing the commit you linked to and more recently in https://github.com/bitcoin/bitcoin/pull/25325.

russeree commented 1 year ago

Would the eBPF USDT tracepoints for the UTXO cache be of use here as a way to measure the overall impact of these flush operations during an IBD or reindex operation?

https://github.com/bitcoin/bitcoin/tree/master/contrib/tracing#log_utxocache_flushpy

sipa commented 1 year ago

@Sjors When doing a full IBD from genesis with dbcache set to 16 GiB, is the limit ever reached (or gotten close to)? If not, it's probably unnecessary to raise it.

Sjors commented 1 year ago

@sipa I ran an IBD as part of testing #27596. Since I did a restart after the cache initially hit 1 GB, the data is not accurate, but the cache reached 13,334 MiB at block at block 798,000. A month earlier at 794,000 the log shows it as 12,766 MiB. So the memory footprint is growing ~7 GB per year (and that's not even the worst month, see chart I linked to above).

russeree commented 1 year ago

I am re running an -reindex=1 with tracepoints watching for flushes.

Here is my last log entry for cache size height 800673 size 14,045 MiB

2023-08-11T12:56:23Z UpdateTip: new best=00000000000000000002b6e22bae3b1f28339b6275c02ec9cbb6f19ffd9c7e9c height=800673 version=0x3017a000 log2_work=94.326952 tx=871155372 date='2023-07-28T20:42:50Z' progress=0.995044 cache=14045.6MiB(115054304txo)

Ran Again 2023-08-16T15:41:07Z UpdateTip: new best=000000000000000000000ff0f3d17f4d92a0534cb63ba8e7293d21ea09932997 height=803380 version=0x20006000 log2_work=94.361871 tx=879733788 date='2023-08-16T01:52:09Z' progress=0.999792 cache=14590.0MiB(119818570txo)

sipa commented 1 year ago

Seems reasonable to increase it to 32 GB in that case. I'm slightly worried about people setting it to the max "because they can, and higher is better" and then getting hit with an unexpectedly slow flush time at the end, but if we're close to hitting the maximum, then it's clearly not crazy to want to set it higher.

Sjors commented 1 year ago

Opened a PR #28358 to ~double~ quadruple it.