When running the import, currently blocks are loaded in batches into a
`seq` then passed to the importer as such.
In reality, blocks are still processed one by one, so the batching does
not offer any performance advantage. It does however require that the
client wastes memory, up to several GB, on the block sequence while
they're waiting to be processed.
This PR introduces a persister that accepts these potentially large
blocks one by one and at the same time removes a number of redundant /
unnecessary copies, assignments and resets that were slowing down the
import process in general.
The forking facility has been replaced by ForkedChain - frames and
layers are two other mechanisms that mostly do the same thing at the
aristo level, without quite providing the functionality FC needs - this
cleanup will make that integration easier.
The current getCanonicalHead of core db should not be confused with ForkedChain.latestHeader.
Therefore we need to use getCanonicalHead to restricted case only, e.g. initializing ForkedChain.
In block processing, depending on the complexity of a transaction and
hotness of caches etc, signature checking can actually make up the
majority of time needed to process a transaction (60% observed in some
randomly sampled block ranges).
Fortunately, this is a task that trivially can be offloaded to a task
pool similar to how nimbus-eth2 does it.
This PR introduces taskpools in the most simple way possible, by
performing signature checking concurrently with other TX processing,
assigning a taskpool task per TX effectively.
With this little trick, we're in gigagas land 🎉 on my laptop!
```
INF 2024-12-10 21:05:35.170+01:00 Imported blocks
blockNumber=3874817 b... mgps=1222.707 ...
```
Tests don't use the taskpool for now because it needs manual cleanup and
we don't have a good mechanism in place. Future PR:s should address this
by creating a common shutdown sequence that also closes and cleans up
other resources like the DB.
Co-authored-by: andri lim <jangko128@gmail.com>
A bit unexpectedly, nibble handling shows up in the profiler mainly
because the current impl is tuned towards slicing while the most common
operation is prefix comparison - since the code is simple, might has
well get rid of some of the excess fat by always aliging the nibbles to
the byte buffer.
* `shouldPrepareTracer` always true
* simple `pop` should not copy value (reading the memory shows up in a
profiler)
* continuation code simplified
* remove some unnecessary EH
* Re-org internal descriptor `CanonicalDesc` as `PivotArc`
why:
Despite its name, `CanonicalDesc` contained a cursor arc (or leg) from
the base tree with a designated block (or Header) on its arc members
(aka blocks.) The type is used more generally than only for s block on
the canonical cursor.
Also, the `PivotArc` provides some more fields for caching intermediate
data. This simplifies managing extra arguments for some functions.
* Remove cruft
details:
No need to find cursor arc if it is given as function argument.
* Rename prototype variables `head: PivotArc` to `pvarc`
why:
Better reading
* Function and code massage, adjust names
details:
Avoid the syllable `canonical` in function names that do not strictly
apply to the canonical chain. So renaming
* findCanonicalHead() => findCursorArc()
* canonicalChain() => findHeader()
* trimCanonicalChain() => trimCursorArc()
* Combine `updateBase()` function-args into single `PivotArgs` object
why:
Will generalise action for more complex scenarios in future.
* update `calculateNewBase()` return code type => `PivotArc`
why:
So it can directly be used as argument into `updateBase()`
* Update `calculateNewBase()` for target on parent arc
* Update unit tests
Each branch node may have up to 16 sub-items - currently, these are
given VertexID based when they are first needed leading to a
mostly-random order of vertexid for each subitem.
Here, we pre-allocate all 16 vertex ids such that when a branch subitem
is filled, it already has a vertexid waiting for it. This brings several
important benefits:
* subitems are sorted and "close" in their id sequencing - this means
that when rocksdb stores them, they are likely to end up in the same
data block thus improving read efficiency
* because the ids are consequtive, we can store just the starting id and
a bitmap representing which subitems are in use - this reduces disk
space usage for branches allowing more of them fit into a single disk
read, further improving disk read and caching performance - disk usage
at block 18M is down from 84 to 78gb!
* the in-memory footprint of VertexRef reduced allowing more instances
to fit into caches and less memory to be used overall.
Because of the increased locality of reference, it turns out that we no
longer need to iterate over the entire database to efficiently generate
the hash key database because the normal computation is now faster -
this significantly benefits "live" chain processing as well where each
dirtied key must be accompanied by a read of all branch subitems next to
it - most of the performance benefit in this branch comes from this
locality-of-reference improvement.
On a sample resync, there's already ~20% improvement with later blocks
seeing increasing benefit (because the trie is deeper in later blocks
leading to more benefit from branch read perf improvements)
```
blocks: 18729664, baseline: 190h43m49s, contender: 153h59m0s
Time (total): -36h44m48s, -19.27%
```
Note: clients need to be resynced as the PR changes the on-disk format
R.I.P. little bloom filter - your life in the repo was short but
valuable
* Kludge: fix `eip4844` import in `validate`
why:
Importing `validate` needs `blscurve` here or with the importing module.
* Separate out `FC` descriptor iinto separate file
why:
Needed for external descriptor access (e.g. for debugging)
* Debugging toolkit for `FC`
* Verify chain descriptor after changing state
The EVM stack is a hot spot in EVM execution and we end up paying a nim
seq tax in several ways, adding up to ~5% of execution time:
* on initial allocation, all bytes get zeroed - this means we have to
choose between allocating a full stack or just a partial one and then
growing it
* pushing and popping introduce additional zeroing
* reallocations on growth copy + zero - expensive again!
* redundant range checking on every operation reducing inlining etc
Here a custom stack using C memory is instroduced:
* no zeroing on allocation
* full stack allocated on EVM startup -> no reallocation during
execution
* fast push/pop - no zeroing again
* 32-byte alignment - this makes it easier for the compiler to use
vector instructions
* no stack allocated for precompiles (these never use it anyway)
Of course, this change also means we have to manage memory manually -
for the EVM, this turns out to be not too bad because we already manage
database transactions the same way (they have to be freed "manually") so
we can simply latch on to this mechanism.
While we're at it, this PR also skips database lookup for known
precompiles by resolving such addresses earlier.
`updateOk` is obsolete and always set to true - callers should not have
to care about this detail
also take the opportunity to clean up storage root naming
When walking AriVtx, parsing integers and nibbles actually becomes a
hotspot - these trivial changes reduces CPU usage during initial key
cache computation by ~15%.
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.
This kind of data is not used except in tests where it is used only to
create databases that don't match actual usage of aristo.
Removing simplifies future optimizations that can focus on processing
specific leaf types more efficiently.
A casualty of this removal is some test code as well as some proof
generation code that is unused - on the surface, it looks like it should
be possible to port both of these to the more specific data types -
doing so would ensure that a database written by one part of the
codebase can interact with the other - as it stands, there is confusion
on this point since using the proof generation code will result in a
database of a shape that is incompatible with the rest of eth1.
* switch to Nim v2.0.12
* fix LruCache capitalization for styleCheck
* KzgProof/KzgCommitment for styleCheck
* TxEip4844 for styleCheck
* styleCheck issues in nimbus/beacon/payload_conv.nim
* ENode for styleCheck
* isOk for styleCheck
* some more styleCheck fixes
* more styleCheck fixes
---------
Co-authored-by: jangko <jangko128@gmail.com>
* prefer the spec-derived name where possible
* don't pass stateRoot to LedgerRef and friends (it doesn't do anything)
* add deprecation warning in graphql - it needs updating to use
forkedchain instead
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.
* partial commit
* fixes
* remove converters too
* revert changes on nimbus_verified_proxy
* revert changes in converter
* revert changes(re-xport) in rpc_types
* update copyright year
* replace types in other binaries
* chain config bug
* fix rebase conflict imcomplete buffer
* fix more rebase buffers
* remove ditto types and converters
* fix the tests
* update copyright year
Jacek Sieka
andri lim
Jordan Hrycaj
tersec
Advaita Saha
Chirag Parmar