With the introduction of layered frames, each database lookup may result
in hundreds of table lookups as the frame stack is traversed.
This change restores performance by introducing snapshots to limit the
lookup depth at the expense of slightly increased memory usage.
The snapshot contains the cumulative changes of all ancestors and itself
allowing the lookup recursion to stop whenever it is encountered.
The number of snapshots to keep in memory is a tradeoff between lookup
performance and memory usage - this change starts with a simple strategy
of keeping snapshots for head frames (approximately). T
he snapshot is created during checkpointing, ie after block validation,
to make sure that it's cheap to start verifying blocks - parent
snapshots are moved to the descendant as part of checkpointing which
effectively means that head frames hold snapshots in most cases.
The outcome of this tradeoff is that applying a block to a known head is
fast while creating a new branch of history remains expensive.
Another consequence is that when persisting changes to disk, we must
re-traverse the stack of changes to build a cumulative set of changes to
be persisted.
A future strategy might be to keep additional "keyframes" along the way,
ie one per epoch for example - this would bound the "branch creation"
cost to a constant factor, but memory overhead should first be
considered.
Another strategy might be to avoid keeping snapshots for non-canonical
branches, specially when they become older and thus less likely to be
branched from.
* `level` is updated to work like a temporary serial number to maintain
its relative position in the sorting order as frames are persisted
* a `snapshot` is added to some TxFrame instances - the snapshot
collects all ancestor changes up to and including the given frame.
`level` is used as a marker to prune the snapshot of changes that have
been persisted already.
* stack traversals for the purpose of lookup stop when they encounter a
snapshot - this bounds the lookup depth to the first encountered
snapshot
After this PR, sync performance lands at about 2-3 blocks per second
(~10x improvement) - this is quite reasonable when comparing with block
import which skips the expensive state root verification and thus
achieves ~20 blk/s on the same hardware. Additional work to bring live
syncing performance in line with disk-based block import would focus on
reducing state root verification cost.
* Simplify txFrame protocol, improve persist performance
To prepare forked-layers for further surgery to avoid the nesting tax,
the commit/rollback style of interacting must first be adjusted, since
it does not provide a point in time where the frame is "done" and goes
from being actively written to, to simply waiting to be persisted or
discarded.
A collateral benefit of this change is that the scheme removes some
complexity from the process by moving the "last saved block number" into
txframe along with the actual state changes thus reducing the risk that
they go "out of sync" and removing the "commit" consolidation
responsibility from ForkedChain.
* commit/rollback become checkpoint/dispose - since these are pure
in-memory constructs, there's less error handling and there's no real
"rollback" involved - dispose better implies that the instance cannot be
used and we can more aggressively clear the memory it uses
* simplified block number handling that moves to become part of txFrame
just like the data that the block number references
* avoid reparenting step by replacing the base instead of keeping a
singleton instance
* persist builds the set of changes from the bottom which helps avoid
moving changes in the top layers through each ancestor level of the
frame stack
* when using an in-memory database in tests, allow the instance to be
passed around to enable testing persist and reload logic
* Use unittest2 test runner
Since upgrading to unittest2, the test runner prints the command line to
re-run a failed test - this however relies on actually using the
unittest2 command line runner.
Previously, test files were assigned numbers - with the unittest2
runner, tests are run using suite/category names instead, like so:
```
# run the Genesis suite
build/all_tests "Genesis::``
# run all tests with "blsMapG1" in the name
build/all_tests "blsMapG1*"
# run tests verbosely
build/all_tests -v
```
A reasonable follow-up here would be to review the suite names to make
them easier to run :)
* lint
* easier-to-compare test order
* bump unittest2 (also the repo)
By introducing the "shared rocksdb instance" concept to the backend, we
can remove the "piggybacking" mode , thus reducing the complexity of
database initialisation and opening the possibility of extending how
write batching works across kvt/aristo.
The change makes explicit the hidden shared state that was previously
hiding in closures and provides the first step towards simplifying the
"commit/persist" interface of coredb, preparing it for optimizations to
reduce the "layering tax" that `forked-layers` introduced.
* Update Nimbus EVM code to use the latest nim-evmc which is now on EVMC v12.1.0
* Fix copyright.
* Fix tests.
* Update to use FkLatest.
* Fix copyright and update test helper.
* renamed nimbus folder to execution_chain
* Renamed "nimbus" references to "execution_chain"
* fixed wrongly changed http reference
* delete snap types file given that it was deleted before this PR merge
* missing 'execution_chain' replacement
---------
Co-authored-by: pmmiranda <pedro.miranda@nimbus.team>
* aristo: fork support via layers/txframes
This change reorganises how the database is accessed: instead holding a
"current frame" in the database object, a dag of frames is created based
on the "base frame" held in `AristoDbRef` and all database access
happens through this frame, which can be thought of as a consistent
point-in-time snapshot of the database based on a particular fork of the
chain.
In the code, "frame", "transaction" and "layer" is used to denote more
or less the same thing: a dag of stacked changes backed by the on-disk
database.
Although this is not a requirement, in practice each frame holds the
change set of a single block - as such, the frame and its ancestors
leading up to the on-disk state represents the state of the database
after that block has been applied.
"committing" means merging the changes to its parent frame so that the
difference between them is lost and only the cumulative changes remain -
this facility enables frames to be combined arbitrarily wherever they
are in the dag.
In particular, it becomes possible to consolidate a set of changes near
the base of the dag and commit those to disk without having to re-do the
in-memory frames built on top of them - this is useful for "flattening"
a set of changes during a base update and sending those to storage
without having to perform a block replay on top.
Looking at abstractions, a side effect of this change is that the KVT
and Aristo are brought closer together by considering them to be part of
the "same" atomic transaction set - the way the code gets organised,
applying a block and saving it to the kvt happens in the same "logical"
frame - therefore, discarding the frame discards both the aristo and kvt
changes at the same time - likewise, they are persisted to disk together
- this makes reasoning about the database somewhat easier but has the
downside of increased memory usage, something that perhaps will need
addressing in the future.
Because the code reasons more strictly about frames and the state of the
persisted database, it also makes it more visible where ForkedChain
should be used and where it is still missing - in particular, frames
represent a single branch of history while forkedchain manages multiple
parallel forks - user-facing services such as the RPC should use the
latter, ie until it has been finalized, a getBlock request should
consider all forks and not just the blocks in the canonical head branch.
Another advantage of this approach is that `AristoDbRef` conceptually
becomes more simple - removing its tracking of the "current" transaction
stack simplifies reasoning about what can go wrong since this state now
has to be passed around in the form of `AristoTxRef` - as such, many of
the tests and facilities in the code that were dealing with "stack
inconsistency" are now structurally prevented from happening. The test
suite will need significant refactoring after this change.
Once this change has been merged, there are several follow-ups to do:
* there's no mechanism for keeping frames up to date as they get
committed or rolled back - TODO
* naming is confused - many names for the same thing for legacy reason
* forkedchain support is still missing in lots of code
* clean up redundant logic based on previous designs - in particular the
debug and introspection code no longer makes sense
* the way change sets are stored will probably need revisiting - because
it's a stack of changes where each frame must be interrogated to find an
on-disk value, with a base distance of 128 we'll at minimum have to
perform 128 frame lookups for *every* database interaction - regardless,
the "dag-like" nature will stay
* dispose and commit are poorly defined and perhaps redundant - in
theory, one could simply let the GC collect abandoned frames etc, though
it's likely an explicit mechanism will remain useful, so they stay for
now
More about the changes:
* `AristoDbRef` gains a `txRef` field (todo: rename) that "more or less"
corresponds to the old `balancer` field
* `AristoDbRef.stack` is gone - instead, there's a chain of
`AristoTxRef` objects that hold their respective "layer" which has the
actual changes
* No more reasoning about "top" and "stack" - instead, each
`AristoTxRef` can be a "head" that "more or less" corresponds to the old
single-history `top` notion and its stack
* `level` still represents "distance to base" - it's computed from the
parent chain instead of being stored
* one has to be careful not to use frames where forkedchain was intended
- layers are only for a single branch of history!
* fix layer vtop after rollback
* engine fix
* Fix test_txpool
* Fix test_rpc
* Fix copyright year
* fix simulator
* Fix copyright year
* Fix copyright year
* Fix tracer
* Fix infinite recursion bug
* Remove aristo and kvt empty files
* Fic copyright year
* Fix fc chain_kvt
* ForkedChain refactoring
* Fix merge master conflict
* Fix copyright year
* Reparent txFrame
* Fix test
* Fix txFrame reparent again
* Cleanup and fix test
* UpdateBase bugfix and fix test
* Fixe newPayload bug discovered by hive
* Fix engine api fcu
* Clean up call template, chain_kvt, andn txguid
* Fix copyright year
* work around base block loading issue
* Add test
* Fix updateHead bug
* Fix updateBase bug
* Change func commitBase to proc commitBase
* Touch up and fix debug mode crash
---------
Co-authored-by: jangko <jangko128@gmail.com>
* Move EIP-7702 Authorization validation to authority func
If the authorization is invalid the transaction itself is still valid,
the invalid authorization will be skipped.
* Fix copyright year
* Refactor TxPool: leaner and simpler
* Rewrite test_txpool
Reduce number of tables used, from 5 to 2. Reduce number of files.
If need to modify the price rule or other filters, now is far more easier because only one table to work with(sender/nonce).
And the other table is just a map from txHash to TxItemRef.
Removing transactions from txPool either because of producing new block or syncing became much easier.
Removing expired transactions also simple.
Explicit Tx Pending, Staged, or Packed status is removed. The status of the transactions can be inferred implicitly.
Developer new to TxPool can easily follow the logic.
But the most important is we can revive the test_txpool without dirty trick and remove usage of getCanonicalHead furthermore to prepare for better integration with ForkedChain.
Instead of using ancient/dirty code to setup the rpc test, now using newest method from TxPool and ForkedChain.
Also fix some bugs in server_api discovered when using this new setup.
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.
