cfcf0f152c
Fix Conflicts: fdbclient/NativeAPI.actor.cpp fdbserver/ClusterRecovery.actor.cpp fdbserver/MasterInterface.h fdbserver/masterserver.actor.cpp flow/error_definitions.h |
||
---|---|---|
.. | ||
BenchCallback.actor.cpp | ||
BenchEncrypt.cpp | ||
BenchHash.cpp | ||
BenchIONet2.actor.cpp | ||
BenchIterate.cpp | ||
BenchMem.cpp | ||
BenchMetadataCheck.cpp | ||
BenchNet2.actor.cpp | ||
BenchPopulate.cpp | ||
BenchRandom.cpp | ||
BenchRef.cpp | ||
BenchStream.actor.cpp | ||
BenchTimer.cpp | ||
BenchVersionVector.cpp | ||
CMakeLists.txt | ||
GlobalData.cpp | ||
GlobalData.h | ||
README.md | ||
benchmark.cmake | ||
flowbench.actor.cpp |
README.md
Summary
flowbench
is an executable that can be used to microbenchmark parts of the FoundationDB code. The goal is to make it easy to test the performance of various sub-millisecond operations using flow
and fdbrpc
. Specifically, this tool can be used to:
- Test the performance effects of changes to the actor compiler or to the
flow
andfdbrpc
libraries - Test the performance of various uses of the
flow
andfdbrpc
libraries - Find areas for improvement in the
flow
andfdbrpc
libraries - Compare
flow
/fdbrpc
primitives to alternatives provided by the standard library or other third-party libraries.
Usage
- To build the
flowbench
executable, runninja flowbench
ormake flowbench
depending on which build system you're using. - Then you can run
bin/flowbench --help
to see possible uses offlowbench
. - Running
bin/flowbench
directly will run all registered benchmarks, but you may want to limit your run to a subset of benchmarks. This can be done by runningbin/flowbench --benchmark_filter=<regex>
- All benchmark names can be listed with
bin/flowbench --benchmark_list_tests
- Example output:
$ bin/flowbench --benchmark_filter=bench_ref
2020-08-04 21:49:40
Running bin/flowbench
Run on (7 X 2904 MHz CPU s)
CPU Caches:
L1 Data 32 KiB (x7)
L1 Instruction 32 KiB (x7)
L2 Unified 256 KiB (x7)
L3 Unified 12288 KiB (x1)
Load Average: 0.15, 0.15, 0.72
---------------------------------------------------------------------------------------------------------------
Benchmark Time CPU Iterations UserCounters...
---------------------------------------------------------------------------------------------------------------
bench_ref_create_and_destroy<RefType::RawPointer> 4.90 ns 4.90 ns 116822124 items_per_second=203.88M/s
bench_ref_create_and_destroy<RefType::UniquePointer> 4.94 ns 4.94 ns 141101924 items_per_second=202.555M/s
bench_ref_create_and_destroy<RefType::SharedPointer> 42.5 ns 42.5 ns 13802909 items_per_second=23.531M/s
bench_ref_create_and_destroy<RefType::FlowReference> 5.05 ns 5.05 ns 100000000 items_per_second=197.955M/s
bench_ref_copy<RefType::RawPointer> 1.15 ns 1.15 ns 612121585 items_per_second=871.218M/s
bench_ref_copy<RefType::SharedPointer> 10.0 ns 10.0 ns 67553102 items_per_second=99.8113M/s
bench_ref_copy<RefType::FlowReference> 2.33 ns 2.33 ns 292317474 items_per_second=428.507M/s
- More detailed documentation can be found at https://github.com/google/benchmark
Existing Benchmarks
bench_populate
measures the population of a vector of mutationsbench_ref
compares the performance of theflow
Reference
type to other pointer typesbench_iterate
measures iteration over a list of mutationsbench_stream
measures the performance of writing to and reading from aPromiseStream
bench_random
measures the performance ofDeterministicRandom
.bench_timer
measures the perforamnce of FoundationDB timers.
Future use cases
- Benchmark the overhead of sending and receiving messages through
FlowTransport
- Benchmark the performance of serializing/deserializing various types