Analysis of binary size

In order to be able to tackle binary size analysis, the Envoy Mobile team has standardized the process in this document.

Object files analysis

Getting a binary

In order to have consistency of results this is the toolchain used to build the binary for analysis:

  1. clang-8

  2. lld (installed with clang)

  3. arm64 machine

The arm64 machine used to build the test binary was an arm64 ec2 a1.2xlarge instance with Ubuntu 18.04. The following script was run to set up the necessary tools:

# basic toolchain
sudo apt-get install build-essential openjdk-8-jdk python zip unzip \
  software-properties-common make cmake bc libtool ninja-build automake \
  time apt-transport-https
# clang-8
sudo tar -C /usr/local/ -xvf clang+llvm-8.0.0-aarch64-linux-gnu.tar.xz --strip 1
rm -rf clang+llvm-8.0.0-aarch64-linux-gnu.tar.xz
# bazel 0.26.1
mkdir -p /tmp/bazel_build
unzip -o -d /tmp/bazel_build
rm -rf
cd /tmp/bazel_build
env EXTRA_BAZEL_ARGS="--host_javabase=@local_jdk//:jdk" bash ./
sudo cp output/bazel /usr/local/bin/bazel
rm -rf /tmp/bazel_build
# bloaty
git clone
cd bloaty
mkdir build
cd build
cmake ..
make -j6
cp bloaty /usr/local/bin/bloaty

The binary being compiled is //test/performance:test_binary_size. The binary is getting built with the following build command:

./bazelw build //test/performance:test_binary_size --config=sizeopt --copt=-ggdb3 --linkopt=-fuse-ld=lld

Thus the binary is compiled with the following flags pertinent to reducing binary size:

  1. -c opt: bazel compilation option for size optimization. Due to sizeopt.

  2. --copt -Os: optimize for size. Due to sizeopt.

  3. --copt=-ggdb3: keep debug symbols. Later stripped with strip. Explicitly added.

  4. --linkopt=-fuse-ld=lld: use the lld linker. Explicitly added.

  5. --define=google_grpc=disabled: more info in the envoy docs. Due to the project’s .bazelrc.

  6. --define=signal_trace=disabled: more info in the envoy docs. Due to the project’s .bazelrc.

  7. --define=tcmalloc=disabled: more info in the envoy docs. Due to the project’s .bazelrc.

  8. --define=hot_restart=disabled: more info in the envoy docs. Due to the project’s .bazelrc.

After compiling, the binary can be stripped of all symbols by using strip:

strip -s bazel-bin/test/performance/test_binary_size

The unstripped and stripped binary can then be used for analysis.


While there are a lot of tools out there that can be used for binary size analysis (otool, objdump, jtool), Bloaty has been the tool of choice to run object file analysis.

Bloaty’s layering of data sources is extremely helpful in being able to explode the binary in all sorts of different ways. For example, one can look at the composition of each compile unit in terms of sections:

$ bloaty --debug-file=bin/test_binary_size -c envoy.bloaty -d sections,bloaty_package,compileunits bin/test_binary_size.stripped
7.7%   109Ki   7.7%   110Ki bazel-out/aarch64-opt/bin/external/envoy_api/envoy/api/v2/route/
  81.9%  89.6Ki  81.4%  89.6Ki .text
  13.6%  14.9Ki  13.5%  14.9Ki .eh_frame
  3.1%  3.45Ki   3.1%  3.45Ki .eh_frame_hdr
  1.3%  1.48Ki   1.3%  1.48Ki .data
  0.0%       0   0.6%     672 .bss
  0.1%      72   0.1%      72 .rodata

Or one might want to see how sections of the binary map to compilation units:

$ bloaty --debug-file=bin/test_binary_size -c envoy.bloaty -d bloaty_package,compileunits,sections bin/test_binary_size.stripped
13.2%   929Ki  13.0%   929Ki .rodata
    81.2%   755Ki  81.2%   755Ki [section .rodata]
    15.4%   143Ki  15.4%   143Ki boringssl/
        37.9%  54.4Ki  37.9%  54.4Ki external/boringssl/src/crypto/obj/obj.c
        21.7%  31.1Ki  21.7%  31.1Ki external/boringssl/src/third_party/fiat/curve25519.c
        17.3%  24.9Ki  17.3%  24.9Ki external/boringssl/src/crypto/fipsmodule/bcm.c
        11.3%  16.2Ki  11.3%  16.2Ki external/boringssl/err_data.c
         4.5%  6.39Ki   4.5%  6.39Ki [55 Others]
         0.9%  1.26Ki   0.9%  1.26Ki external/boringssl/src/crypto/x509v3/v3_crld.c
         0.7%  1.06Ki   0.7%  1.06Ki external/boringssl/src/crypto/asn1/tasn_typ.c

These different representations will give you perspective about how different changes in the binary will affect size. Note that the envoy.bloaty config refers to a bloaty config that has regexes to capture output. The example config used in this type of analysis is:

custom_data_source: {
  name: "bloaty_package"
  base_data_source: "compileunits"

  #envoy source code.
  rewrite: {
    pattern: "^(external/envoy/source/)(\\w+/)(\\w+)"
    replacement: "envoy \\2"

  #envoy third party libraries.
  rewrite: {
      pattern: "^(external/)(\\w+/)"
      replacement: "\\2"

  #all compiled protos.
  rewrite: {
      pattern: "([ |])$"
      replacement: "compiled protos"

Open issues regarding size

perf/size is a label tagging all current open issues that can improve binary size. Check out the issues here. After performing any change that tries to address these issues you should run through the analysis pipeline described above, and make sure your changes match expectations.

Current status


When compiling Envoy Mobile for arm64 only, we found the final size to be approximately 4.6 MB as of v0.2.3.03062020.

This analysis was done by:

  • Compiling the analysis variant example app for release

  • Exporing the app for Ad Hoc distribution using Xcode

  • Enabling app thinning for arm64 only

  • Investigating the .ipa file and/or App Thinning Size Report.txt

  • Doing the same for the analysis control example app in that repository, and comparing the size differences

App + On Demand Resources size: 27 KB compressed, 113 KB uncompressed
App size: 27 KB compressed, 113 KB uncompressed

App + On Demand Resources size: 4.6 MB compressed, 13.8 MB uncompressed
App size: 4.6 MB compressed, 13.8 MB uncompressed

Net: 4.6 MB compressed, 13.8 MB uncompressed


This is being done in #742.

CI integration

CI validates that no PR increases the binary size of the library above a specific threshold specified in the perf.yml configuration.

The status of this job is reported on PRs in the perf / size_compare task.