BOLT is a post-link optimizer developed to speed up large applications.
It achieves the improvements by optimizing application's code layout based on
execution profile gathered by sampling profiler, such as Linux perf tool.
An overview of the ideas implemented in BOLT along with a discussion of its
potential and current results is available in
CGO'19 paper.
VESPA is a static profiler based on ML designed to guide binary optimizations. VESPA removes the need for dynamic profiling to enable binary optimizations, thus simplifying and broadening the applicability of binary-optimization tools. Although VESPA is a general static profiler that can be used with different binary optimizers, we explored the use of VESPA with BOLT and you can see the discussion of this combined use here OOPSLA'21 paper.
BOLT operates on X86-64 and AArch64 ELF binaries. At the minimum, the binaries
should have an unstripped symbol table, and, to get maximum performance gains,
they should be linked with relocations (--emit-relocs or -q linker flag).
BOLT disassembles functions and reconstructs the control flow graph (CFG) before it runs optimizations. Since this is a nontrivial task, especially when indirect branches are present, we rely on certain heuristics to accomplish it. These heuristics have been tested on a code generated with Clang and GCC compilers. The main requirement for C/C++ code is not to rely on code layout properties, such as function pointer deltas. Assembly code can be processed too. Requirements for it include a clear separation of code and data, with data objects being placed into data sections/segments. If indirect jumps are used for intra-function control transfer (e.g., jump tables), the code patterns should be matching those generated by Clang/GCC.
NOTE: BOLT is currently incompatible with the -freorder-blocks-and-partition
compiler option. Since GCC8 enables this option by default, you have to
explicitly disable it by adding -fno-reorder-blocks-and-partition flag if
you are compiling with GCC8.
PIE and .so support has been added recently. Please report bugs if you encounter any issues.
You can build and use the docker image containing BOLT using our docker file. Alternatively, you can build BOLT manually using the steps below.
You can acces the docker image containing the artifact of the paper VESPA: Static Profiling for Binary Optimization at docker hub. If you have any doubts on how to run the scripts that reproduce the research questions please read the file published at zenodo.
BOLT heavily uses LLVM libraries, and by design, it is built as one of LLVM tools. The build process is not much different from a regular LLVM build. The following instructions are assuming that you are running under Linux.
Start with cloning LLVM and BOLT repos:
> git clone https://github.com/llvm-mirror/llvm llvm
> cd llvm/tools
> git checkout -b llvm-bolt f137ed238db11440f03083b1c88b7ffc0f4af65e
> git clone https://github.com/angelica-moreira/BOLT.git llvm-bolt
> cd ..
> patch -p 1 < tools/llvm-bolt/llvm.patch
Proceed to a normal LLVM build using a compiler with C++11 support (for GCC use version 4.9 or later):
> cd ..
> mkdir build
> cd build
> cmake -G Ninja ../llvm -DLLVM_TARGETS_TO_BUILD="X86;AArch64" -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=ON
> ninja
llvm-bolt will be available under bin/. Add this directory to your path to
ensure the rest of the commands in this tutorial work.
Note that we use a specific revision of LLVM as we currently rely on a set of patches that are not yet upstreamed.
BOLT runs many internal passes in parallel. If you foresee heavy usage of BOLT, you can improve the processing time by linking against one of memory allocation libraries with good support for concurrency. E.g. to use jemalloc:
> sudo yum install jemalloc-devel
> LD_PRELOAD=/usr/lib64/libjemalloc.so llvm-bolt ....
Or if you rather use tcmalloc:
> sudo yum install gperftools-devel
> LD_PRELOAD=/usr/lib64/libtcmalloc_minimal.so llvm-bolt ....
For a complete practical guide of using BOLT see Optimizing Clang with BOLT.
In order to allow BOLT to re-arrange functions (in addition to re-arranging
code within functions) in your program, it needs a little help from the linker.
Add --emit-relocs to the final link step of your application. You can verify
the presence of relocations by checking for .rela.text section in the binary.
BOLT will also report if it detects relocations while processing the binary.
This step is different for different kinds of executables. If you can invoke your program to run on a representative input from a command line, then check For Applications section below. If your program typically runs as a server/service, then skip to For Services section.
The version of perf command used for the following steps has to support
-F brstack option. We recommend using perf version 4.5 or later.
This assumes you can run your program from a command line with a typical input.
In this case, simply prepend the command line invocation with perf:
$ perf record -e cycles:u -j any,u -o perf.data -- <executable> <args> ...
Once you get the service deployed and warmed-up, it is time to collect perf data with LBR (branch information). The exact perf command to use will depend on the service. E.g., to collect the data for all processes running on the server for the next 3 minutes use:
$ perf record -e cycles:u -j any,u -a -o perf.data -- sleep 180
Depending on the application, you may need more samples to be included with
your profile. It's hard to tell upfront what would be a sweet spot for your
application. We recommend the profile to cover 1B instructions as reported
by BOLT -dyno-stats option. If you need to increase the number of samples
in the profile, you can either run the sleep command for longer and use
-F<N> option with perf to increase sampling frequency.
Note that for profile collection we recommend using cycle events and not
BR_INST_RETIRED.*. Empirically we found it to produce better results.
If the collection of a profile with branches is not available, e.g., when you run on a VM or on hardware that does not support it, then you can use only sample events, such as cycles. In this case, the quality of the profile information would not be as good, and performance gains with BOLT are expected to be lower.
If perf record is not available to you, you may collect profile by first instrumenting the binary with BOLT and then running it.
$ llvm-bolt <executable> -instrument -o <instrumented-executable>
After you run instrumented-executable with the desired workload, its BOLT
profile should be ready for you in /tmp/prof.fdata and you can skip
Step 2.
Run BOLT with the -help option and check the category "BOLT instrumentation
options" for a quick reference on instrumentation knobs. Instrumentation is
experimental and currently does not work for PIEs/SOs.
NOTE: you can skip this step and feed perf.data directly to BOLT using
experimental -p perf.data option.
For this step, you will need perf.data file collected from the previous step and
a copy of the binary that was running. The binary has to be either
unstripped, or should have a symbol table intact (i.e., running strip -g is
okay).
Make sure perf is in your PATH, and execute perf2bolt:
$ perf2bolt -p perf.data -o perf.fdata <executable>
This command will aggregate branch data from perf.data and store it in a
format that is both more compact and more resilient to binary modifications.
If the profile was collected without LBRs, you will need to add -nl flag to
the command line above.
Once you have perf.fdata ready, you can use it for optimizations with
BOLT. Assuming your environment is setup to include the right path, execute
llvm-bolt:
$ llvm-bolt <executable> -o <executable>.bolt -data=perf.fdata -reorder-blocks=cache+ \
-reorder-functions=hfsort -split-functions=2 -split-all-cold -split-eh -dyno-stats
If you do need an updated debug info, then add -update-debug-sections option
to the command above. The processing time will be slightly longer.
For a full list of options see -help/-help-hidden output.
The input binary for this step does not have to 100% match the binary used for
profile collection in Step 1. This could happen when you are doing active
development, and the source code constantly changes, yet you want to benefit
from profile-guided optimizations. However, since the binary is not precisely the
same, the profile information could become invalid or stale, and BOLT will
report the number of functions with a stale profile. The higher the
number, the less performance improvement should be expected. Thus, it is
crucial to update .fdata for release branches.
Suppose your application can run in different modes, and you can generate
multiple profiles for each one of them. To generate a single binary that can
benefit all modes (assuming the profiles don't contradict each other) you can
use merge-fdata tool:
$ merge-fdata *.fdata > combined.fdata
Use combined.fdata for Step 3 above to generate a universally optimized
binary.
See Step 0 from Standard BOLT Usage section.
To generate the csv feature file (features.csv) you must do as follows:
$ llvm-bolt <executable> -gen-features
Before you go ahead make sure that you have python3 installed
in your machine and then create a virtual enviroment in your
desired directory ($VIRTUAL_ENV_DIR) as follows:
$ python3 -m venv ${VIRTUAL_ENV_DIR}/vespa-env
Now activate it and install the dependences needed for running
the scrpits of directory ${LLVM_BOLT_DIRECTORY}/utils/vespa_scripts.
The dependency files is at ${LLVM_BOLT_DIRECTORY}/utils/requirements.txt,
make sure to not forget to upgrate pip before installing the dependences, as
follows:
$ source ${VIRTUAL_ENV_DIR}/vespa-env/bin/activate
$ PYTHON_HOME=${VIRTUAL_ENV_DIR}/vespa-env/bin
$ ${PYTHON_HOME}/pip install --upgrade pip
$ ${PYTHON_HOME}/pip install -r ${LLVM_BOLT_DIRECTORY}/utils/requirements.txt
You can skip the installation if you had already done it before, but make sure to activate it
every time you want to run the scripts preprocessing.py and generate_probfiles.py.
To generate the probabilities file that will have the prediction performed by VESPA and that will be used to guide BOLT optimizations, do as follows:
$ VESPA_SCRIPTS_HOME=${LLVM_BOLT_DIRECTORY}/utils/vespa_scripts/
$ source ${VIRTUAL_ENV_DIR}/vespa-env/bin/activate
$ PYTHON_HOME=${VIRTUAL_ENV_DIR}/vespa-env/bin
$ ${PYTHON_HOME}/python3 ${VESPA_SCRIPTS_HOME}/preprocessing.py <features_path> <stage_code> <result_path>
$ ${PYTHON_HOME}/python3 ${VESPA_SCRIPTS_HOME}/generate_probfiles.py <model_path> <data_base_path> <output_path> [bin_name]
$ deactivate
Note: The LLVM_BOLT_DIRECTORY variable represents llvm-bolt source directory
the ${HOME}/llvm/tools/llvm-bolt/.
The input variable features_path represents the path where the features.csv is, the stage_code can assume
values 0 (train) or 1 (prediction), and result_path is the path where the preprocessed csv file will be persisted.
The input variable model_path is the path where vespa ml model is, data_base_path represents the path where
the preprocessed csv file was persisted, the output_path is where you want to persist the probabilities file and
bin_name is optional, if you enter the binary name the output file will be named as bin_name_ml.pdata,
otherwise the file will be generated with the default name (prob.pdata).
Once you have prob.pdata ready, you can use it for optimizations with BOLT.
Assuming your environment is setup to include the right path, execute llvm-bolt:
$ llvm-bolt <executable> -o <executable>.bolt -infer-local-counts -ml-based -prob-file=prob.pdata \
-reorder-blocks=cache+ -split-functions=3 -split-all-cold -dyno-stats \
-icf=1 -use-gnu-stack
Be aware that the command above does not work well for rearrenging functions.
If you intend to enable this optimization, it is advisable to use the
flag -infer-global-counts instead of -infer-local-counts
when executing llvm-bolt:
$ llvm-bolt <executable> -o <executable>.bolt -infer-global-counts -ml-based \
-prob-file=prob.pdata -reorder-blocks=cache+ \
-split-functions=3 -split-all-cold -dyno-stats -icf=1 \
-reorder-functions=hfsort -split-eh -use-gnu-stack
Note: Here we are using the pre-trained model with our 243 programs, read the OOPSLA'21 paper for a deitaled information about it, as input for training, but feel free to explore tranning with other dataset following the steps presented in the Training VESPA with your own dataset tutorial.
We incorporated in BOLT some trivial static predictors and the Wu and Larus heuristics. Below are the steps of how to run BOLT with the Wu and Larus static heuristic-based profiler.
See Step 0 from Standard BOLT Usage section.
Since the trivial predictors and the Wu and Larus predictor are heuristic-based, and therefore perform branch probability inference by using heuristics that analysis the program's source code, you will not need to feed it with a file, you only need to specify the type of predictor you want to use. In the exemple below we are using Wu and Larus predictor with the intra-procedural strategy for converting probabilities into frequencies, mode details about this technique can be found at Wu and Larus's paper.
$ llvm-bolt <executable> -o <executable>.bolt -infer-local-counts -heuristic-based=wularus \
-reorder-blocks=cache+ -split-functions=3 -split-all-cold \
-dyno-stats -icf=1 -use-gnu-stack
We also implemented the inter-procedural algorithm described in the paper. Using this implementation, you may be able to perform function reordering better than using the previously inferred data. However, as mentioned earlier in the BOLT + VESPA section, it is not advisable to use this approach for large or highly optimized programs, since they usually have a large number of non-reducible graphs. This caused a combination of both intra-procedural and inter-procedural techiniques to yield worse results when compared against the binary produced using only the intra-procedural propagation.
$ llvm-bolt <executable> -o <executable>.bolt -infer-global-counts -heuristic-based=wularus \
-reorder-blocks=cache+ -split-functions=3 -split-all-cold -dyno-stats \
-icf=1 -reorder-functions=hfsort -split-eh -use-gnu-stack
For testing the trivial predictions please run BOLT with the -help option and check the
category "BOLT static infered profile options" for a quick reference on trivial heuristics
and its description, and replace the name of the trivial predictor in the field -heuristic-based.
BOLT and BOLT+VESPA are licensed under University of Illinois/NCSA Open Source License.