Currently, computed hash keys are stored in a separate column family
with respect to the MPT data they're generated from - this has several
disadvantages:
* A lot of space is wasted because the lookup key (`RootedVertexID`) is
repeated in both tables - this is 30% of the `AriKey` content!
* rocksdb must maintain in-memory bloom filters and LRU caches for said
keys, doubling its "minimal efficient cache size"
* An extra disk traversal must be made to check for existence of cached
hash key
* Doubles the amount of files on disk due to each column family being
its own set of files
Here, the two CFs are joined such that both key and data is stored in
`AriVtx`. This means:
* we save ~30% disk space on repeated lookup keys
* we save ~2gb of memory overhead that can be used to cache data instead
of indices
* we can skip storing hash keys for MPT leaf nodes - these are trivial
to compute and waste a lot of space - previously they had to present in
the `AriKey` CF to avoid having to look in two tables on the happy path.
* There is a small increase in write amplification because when a hash
value is updated for a branch node, we must write both key and branch
data - previously we would write only the key
* There's a small shift in CPU usage - instead of performing lookups in
the database, hashes for leaf nodes are (re)-computed on the fly
* We can return to slightly smaller on-disk SST files since there's
fewer of them, which should reduce disk traffic a bit
Internally, there are also other advantages:
* when clearing keys, we no longer have to store a zero hash in memory -
instead, we deduce staleness of the cached key from the presence of an
updated VertexRef - this saves ~1gb of mem overhead during import
* hash key cache becomes dedicated to branch keys since leaf keys are no
longer stored in memory, reducing churn
* key computation is a lot faster thanks to the skipped second disk
traversal - a key computation for mainnet can be completed in 11 hours
instead of ~2 days (!) thanks to better cache usage and less read
amplification - with additional improvements to the on-disk format, we
can probably get rid of the initial full traversal method of seeding the
key cache on first start after import
All in all, this PR reduces the size of a mainnet database from 160gb to
110gb and the peak memory footprint during import by ~1-2gb.
When `nimbus import` runs, we end up with a database without MPT roots
leading to long startup times the first time one is needed.
Computing the state root is slow because the on-disk order based on
VertexID sorting does not match the trie traversal order and therefore
makes lookups inefficent.
Here we introduce a helper that speeds up this computation by traversing
the trie in on-disk order and computing the trie hashes bottom up
instead - even though this leads to some redundant reads of nodes that
we cannot yet compute, it's still a net win as leaves and "bottom"
branches make up the majority of the database.
This PR also addresses a few other sources of inefficiency largely due
to the separation of AriKey and AriVtx into their own column families.
Each column family is its own LSM tree that produces hundreds of SST
filtes - with a limit of 512 open files, rocksdb must keep closing and
opening files which leads to expensive metadata reads during random
access.
When rocksdb makes a lookup, it has to read several layers of files for
each lookup. Ribbon filters to skip over files that don't have the
requested data but when these filters are not in memory, reading them is
slow - this happens in two cases: when opening a file and when the
filter has been evicted from the LRU cache. Addressing the open file
limit solves one source of inefficiency, but we must also increase the
block cache size to deal with this problem.
* rocksdb.max_open_files increased to 2048
* per-file size limits increased so that fewer files are created
* WAL size increased to avoid partial flushes which lead to small files
* rocksdb block cache increased
All these increases of course lead to increased memory usage, but at
least performance is acceptable - in the future, we'll need to explore
options such as joining AriVtx and AriKey and/or reducing the row count
(by grouping branch layers under a single vertexid).
With this PR, the mainnet state root can be computed in ~8 hours (down
from 2-3 days) - not great, but still better.
Further, we write all keys to the database, also those that are less
than 32 bytes - because the mpt path is part of the input, it is very
rare that we actually hit a key like this (about 200k such entries on
mainnet), so the code complexity is not worth the benefit really, in the
current database layout / design.
* replace rocksdb row cache with larger rdb lru caches - these serve the
same purpose but are more efficient because they skips serialization,
locking and rocksdb layering
* don't append fresh items to cache - this has the effect of evicting
the existing items and replacing them with low-value entries that might
never be read - during write-heavy periods of processing, the
newly-added entries were evicted during the store loop
* allow tuning rdb lru size at runtime
* add (hidden) option to print lru stats at exit (replacing the
compile-time flag)
pre:
```
INF 2024-09-03 15:07:01.136+02:00 Imported blocks
blockNumber=20012001 blocks=12000 importedSlot=9216851 txs=1837042
mgas=181911.265 bps=11.675 tps=1870.397 mgps=176.819 avgBps=10.288
avgTps=1574.889 avgMGps=155.952 elapsed=19m26s458ms
```
post:
```
INF 2024-09-03 13:54:26.730+02:00 Imported blocks
blockNumber=20012001 blocks=12000 importedSlot=9216851 txs=1837042
mgas=181911.265 bps=11.637 tps=1864.384 mgps=176.250 avgBps=11.202
avgTps=1714.920 avgMGps=169.818 elapsed=17m51s211ms
```
9%:ish import perf improvement on similar mem usage :)
Based on some simple testing done with a few combinations of cache
sizes, it seems that the block cache has grown in importance compared to
the where we were before changing on-disk format and adding a lot of
other point caches.
With these settings, there's roughly a 15% performance increase when
processing blocks in the 18M range over the status quo while memory
usage decreases by more than 1gb!
Only a few values were tested so there's certainly more to do here but
this change sets up a better baseline for any future optimizations.
In particular, since the initial defaults were chosen root vertex id:s
were introduced as key prefixes meaning that storage for each account
will be grouped together and thus it becomes more likely that a block
loaded from disk will be hit multiple times - this seems to give the
block cache an edge over the row cache, specially when traversing the
storage trie.
* creating a seq from a table that holds lots of changes means copying
all data into the table - this can be several GB of data while syncing
blocks
* nim fails to optimize the moving of the `WidthFirstForest` - the real
solution is to not construct a `wff` to begin with, but this PR provides
relief while that is being worked on
This spike fix allows us to bump the rocksdb cache by another 2 GB and
still have a significantly lower peak memory usage during sync.
When processing long ranges of blocks, the account cache grows unbounded
which cause huge memory spikes.
Here, we move the cache to a second-level cache after each block - the
second-level cache is cleared on the next block after that which creates
a simple LRU effect.
There's a small performance cost of course, though overall the freed-up
memory can now be reassigned to the rocksdb row cache which not only
makes up for the loss but overall leads to a performance increase.
The bump to 2gb of rocksdb row cache here needs more testing but is
slightly less and loosely basedy on the savings from this PR and the
circular ref fix in #2408 - another way to phrase this is that it's
better to give rocksdb more breathing room than let the memory sit
unused until circular ref collection happens ;)
These options are there mainly to drive experiments, and are therefore
hidden.
One thing that this PR brings in is an initial set of caches and buffers for rocksdb - the set that I've been using during various performance tests to get to a viable baseline performance level.
Jacek Sieka