3fcb545ab26be3e785b5e5654be0bdc77099d827 bench: benchmark transaction creation process (furszy)
a8a75346d7e7247596c8a580d65ceaad49c97b97 wallet: SelectCoins, return early if target is covered by preset-inputs (furszy)
f41712a734dc119f8a5e053a9cfa1f0411b5e8f1 wallet: simplify preset inputs selection target check (furszy)
5baedc33519661af9d19efcefd23dca8998d2547 wallet: remove fetch pre-selected-inputs responsibility from SelectCoins (furszy)
295852f61998a025b0b28a0671e6e1cf0dc08d0d wallet: encapsulate pre-selected-inputs lookup into its own function (furszy)
37e7887cb4bfd7db6eb462ed0741c45aea22a990 wallet: skip manually selected coins from 'AvailableCoins' result (furszy)
94c0766b0cd1990c1399a745c88c2ba4c685d8d1 wallet: skip available coins fetch if "other inputs" are disallowed (furszy)
Pull request description:
#### # Context (Current Flow on Master)
In the transaction creation process, in order to select which coins the new transaction will spend,
we first obtain all the available coins known by the wallet, which means walking-through the
wallet txes map, gathering the ones that fulfill certain spendability requirements in a vector.
This coins vector is then provided to the Coin Selection process, which first checks if the user
has manually selected any input (which could be internal, aka known by the wallet, or external),
and if it does, it fetches them by searching each of them inside the wallet and/or inside the
Coin Control external tx data.
Then, after finding the pre-selected-inputs and gathering them in a vector, the Coin Selection
process walks-through the entire available coins vector once more just to erase coins that are
in both vectors. So the Coin Selection process doesn’t pick them twice (duplicate inputs inside
the same transaction).
#### # Process Workflow Changes
Now, a new method, `FetchCoins` will be responsible for:
1) Lookup the user pre-selected-inputs (which can be internal or external).
2) And, fetch the available coins in the wallet (excluding the already fetched ones).
Which will occur prior to the Coin Selection process. Which allows us to never include the
pre-selected-inputs inside the available coins vector in the first place, as well as doing other
nice improvements (written below).
So, Coin Selection can perform its main responsibility without mixing it with having to fetch
internal/external coins nor any slow and unneeded duplicate coins verification.
#### # Summarizing the Improvements:
1) If any pre-selected-input lookup fail, the process will return the error right away.
(before, the wallet was fetching all the wallet available coins, walking through the
entire txes map, and then failing for an invalid pre-selected-input inside SelectCoins)
2) The pre-selected-inputs lookup failure causes are properly described on the return error.
(before, we were returning an "Insufficient Funds" error for everything, even if the failure
was due a not solvable external input)
3) **Faster Coin Selection**: no longer need to "remove the pre-set inputs from the available coins
vector so that Coin Selection doesn't pick them" (which meant to loop-over the entire
available coins vector at Coin Selection time, erasing duplicate coins that were pre-selected).
Now, the available coins vector, which is built after the pre-selected-inputs fetching,
doesn’t include the already selected inputs in the first place.
4) **Faster transaction creation** for transactions that only use manually selected inputs.
We now will return early, as soon as we finish fetching the pre-selected-inputs and
not perform the resources expensive calculation of walking-through the entire wallet
txes map to obtain the available coins (coins that we will not use).
---------------------------
Added a new bench (f6d0bb2) measuring the transaction creation process, for a wallet with ~250k UTXO, only using the pre-selected-inputs inside coin control. Setting `m_allow_other_inputs=false` to disallow the wallet to include coins automatically.
#### Result on this PR (tip f6d0bb2d):
| ns/op | op/s | err% | total | benchmark
|--------------------:|--------------------:|--------:|----------:|:----------
| 1,048,675.00 | 953.58 | 0.3% | 0.06 | `WalletCreateTransaction`
vs
#### Result on master (tip 4a4289e2):
| ns/op | op/s | err% | total | benchmark
|--------------------:|--------------------:|--------:|----------:|:----------
| 96,373,458.20 | 10.38 | 0.2% | 5.30 | `WalletCreateTransaction`
The benchmark took to run in master: **96.37 milliseconds**, while in this PR: **1 millisecond** 🚀 .
ACKs for top commit:
S3RK:
Code Review ACK 3fcb545ab26be3e785b5e5654be0bdc77099d827
achow101:
ACK 3fcb545ab26be3e785b5e5654be0bdc77099d827
aureleoules:
reACK 3fcb545ab26be3e785b5e5654be0bdc77099d827
Tree-SHA512: 42f833e92f40c348007ca565a4c98039e6f1ff25d8322bc2b27115824744779baf0b0a38452e4e2cdcba45076473f1028079bbd0f670020481ec5d3db42e4731
This directory contains integration tests that test bitcoind and its utilities in their entirety. It does not contain unit tests, which can be found in /src/test, /src/wallet/test, etc.
This directory contains the following sets of tests:
- fuzz A runner to execute all fuzz targets from /src/test/fuzz.
- functional which test the functionality of bitcoind and bitcoin-qt by interacting with them through the RPC and P2P interfaces.
- util which tests the utilities (bitcoin-util, bitcoin-tx, ...).
- lint which perform various static analysis checks.
The util tests are run as part of make check
target. The fuzz tests, functional
tests and lint scripts can be run as explained in the sections below.
Running tests locally
Before tests can be run locally, Bitcoin Core must be built. See the building instructions for help.
Fuzz tests
See /doc/fuzzing.md
Functional tests
Dependencies and prerequisites
The ZMQ functional test requires a python ZMQ library. To install it:
- on Unix, run
sudo apt-get install python3-zmq
- on mac OS, run
pip3 install pyzmq
On Windows the PYTHONUTF8
environment variable must be set to 1:
set PYTHONUTF8=1
Running the tests
Individual tests can be run by directly calling the test script, e.g.:
test/functional/feature_rbf.py
or can be run through the test_runner harness, eg:
test/functional/test_runner.py feature_rbf.py
You can run any combination (incl. duplicates) of tests by calling:
test/functional/test_runner.py <testname1> <testname2> <testname3> ...
Wildcard test names can be passed, if the paths are coherent and the test runner
is called from a bash
shell or similar that does the globbing. For example,
to run all the wallet tests:
test/functional/test_runner.py test/functional/wallet*
functional/test_runner.py functional/wallet* (called from the test/ directory)
test_runner.py wallet* (called from the test/functional/ directory)
but not
test/functional/test_runner.py wallet*
Combinations of wildcards can be passed:
test/functional/test_runner.py ./test/functional/tool* test/functional/mempool*
test_runner.py tool* mempool*
Run the regression test suite with:
test/functional/test_runner.py
Run all possible tests with
test/functional/test_runner.py --extended
In order to run backwards compatibility tests, first run:
test/get_previous_releases.py -b
to download the necessary previous release binaries.
By default, up to 4 tests will be run in parallel by test_runner. To specify
how many jobs to run, append --jobs=n
The individual tests and the test_runner harness have many command-line
options. Run test/functional/test_runner.py -h
to see them all.
Speed up test runs with a ramdisk
If you have available RAM on your system you can create a ramdisk to use as the cache
and tmp
directories for the functional tests in order to speed them up.
Speed-up amount varies on each system (and according to your ram speed and other variables), but a 2-3x speed-up is not uncommon.
To create a 4GB ramdisk on Linux at /mnt/tmp/
:
sudo mkdir -p /mnt/tmp
sudo mount -t tmpfs -o size=4g tmpfs /mnt/tmp/
Configure the size of the ramdisk using the size=
option.
The size of the ramdisk needed is relative to the number of concurrent jobs the test suite runs.
For example running the test suite with --jobs=100
might need a 4GB ramdisk, but running with --jobs=32
will only need a 2.5GB ramdisk.
To use, run the test suite specifying the ramdisk as the cachedir
and tmpdir
:
test/functional/test_runner.py --cachedir=/mnt/tmp/cache --tmpdir=/mnt/tmp
Once finished with the tests and the disk, and to free the ram, simply unmount the disk:
sudo umount /mnt/tmp
Troubleshooting and debugging test failures
Resource contention
The P2P and RPC ports used by the bitcoind nodes-under-test are chosen to make conflicts with other processes unlikely. However, if there is another bitcoind process running on the system (perhaps from a previous test which hasn't successfully killed all its bitcoind nodes), then there may be a port conflict which will cause the test to fail. It is recommended that you run the tests on a system where no other bitcoind processes are running.
On linux, the test framework will warn if there is another bitcoind process running when the tests are started.
If there are zombie bitcoind processes after test failure, you can kill them by running the following commands. Note that these commands will kill all bitcoind processes running on the system, so should not be used if any non-test bitcoind processes are being run.
killall bitcoind
or
pkill -9 bitcoind
Data directory cache
A pre-mined blockchain with 200 blocks is generated the first time a functional test is run and is stored in test/cache. This speeds up test startup times since new blockchains don't need to be generated for each test. However, the cache may get into a bad state, in which case tests will fail. If this happens, remove the cache directory (and make sure bitcoind processes are stopped as above):
rm -rf test/cache
killall bitcoind
Test logging
The tests contain logging at five different levels (DEBUG, INFO, WARNING, ERROR
and CRITICAL). From within your functional tests you can log to these different
levels using the logger included in the test_framework, e.g.
self.log.debug(object)
. By default:
- when run through the test_runner harness, all logs are written to
test_framework.log
and no logs are output to the console. - when run directly, all logs are written to
test_framework.log
and INFO level and above are output to the console. - when run by our CI (Continuous Integration), no logs are output to the console. However, if a test
fails, the
test_framework.log
and bitcoinddebug.log
s will all be dumped to the console to help troubleshooting.
These log files can be located under the test data directory (which is always printed in the first line of test output):
<test data directory>/test_framework.log
<test data directory>/node<node number>/regtest/debug.log
.
The node number identifies the relevant test node, starting from node0
, which
corresponds to its position in the nodes list of the specific test,
e.g. self.nodes[0]
.
To change the level of logs output to the console, use the -l
command line
argument.
test_framework.log
and bitcoind debug.log
s can be combined into a single
aggregate log by running the combine_logs.py
script. The output can be plain
text, colorized text or html. For example:
test/functional/combine_logs.py -c <test data directory> | less -r
will pipe the colorized logs from the test into less.
Use --tracerpc
to trace out all the RPC calls and responses to the console. For
some tests (eg any that use submitblock
to submit a full block over RPC),
this can result in a lot of screen output.
By default, the test data directory will be deleted after a successful run.
Use --nocleanup
to leave the test data directory intact. The test data
directory is never deleted after a failed test.
Attaching a debugger
A python debugger can be attached to tests at any point. Just add the line:
import pdb; pdb.set_trace()
anywhere in the test. You will then be able to inspect variables, as well as call methods that interact with the bitcoind nodes-under-test.
If further introspection of the bitcoind instances themselves becomes
necessary, this can be accomplished by first setting a pdb breakpoint
at an appropriate location, running the test to that point, then using
gdb
(or lldb
on macOS) to attach to the process and debug.
For instance, to attach to self.node[1]
during a run you can get
the pid of the node within pdb
.
(pdb) self.node[1].process.pid
Alternatively, you can find the pid by inspecting the temp folder for the specific test you are running. The path to that folder is printed at the beginning of every test run:
2017-06-27 14:13:56.686000 TestFramework (INFO): Initializing test directory /tmp/user/1000/testo9vsdjo3
Use the path to find the pid file in the temp folder:
cat /tmp/user/1000/testo9vsdjo3/node1/regtest/bitcoind.pid
Then you can use the pid to start gdb
:
gdb /home/example/bitcoind <pid>
Note: gdb attach step may require ptrace_scope to be modified, or sudo
preceding the gdb
.
See this link for considerations: https://www.kernel.org/doc/Documentation/security/Yama.txt
Often while debugging RPC calls in functional tests, the test might time out before the
process can return a response. Use --timeout-factor 0
to disable all RPC timeouts for that particular
functional test. Ex: test/functional/wallet_hd.py --timeout-factor 0
.
Profiling
An easy way to profile node performance during functional tests is provided
for Linux platforms using perf
.
Perf will sample the running node and will generate profile data in the node's
datadir. The profile data can then be presented using perf report
or a graphical
tool like hotspot.
To generate a profile during test suite runs, use the --perf
flag.
To see render the output to text, run
perf report -i /path/to/datadir/send-big-msgs.perf.data.xxxx --stdio | c++filt | less
For ways to generate more granular profiles, see the README in test/functional.
Util tests
Util tests can be run locally by running test/util/test_runner.py
.
Use the -v
option for verbose output.
Lint tests
Dependencies
Lint test | Dependency |
---|---|
lint-python.py |
flake8 |
lint-python.py |
mypy |
lint-python.py |
pyzmq |
lint-python-dead-code.py |
vulture |
lint-shell.py |
ShellCheck |
lint-spelling.py |
codespell |
In use versions and install instructions are available in the CI setup.
Please be aware that on Linux distributions all dependencies are usually available as packages, but could be outdated.
Running the tests
Individual tests can be run by directly calling the test script, e.g.:
test/lint/lint-files.py
You can run all the shell-based lint tests by running:
test/lint/all-lint.py
Writing functional tests
You are encouraged to write functional tests for new or existing features. Further information about the functional test framework and individual tests is found in test/functional.