e25e4dc7e7
This improves performance with sandboxing. One can have Bazel output performance data [1][2], which can be viewed via chrome://tracing or https://ui.perfetto.dev/. With Perfetto, the following SQL queries [3][4] can summarize the sandboxing metrics, as well as the actual compilation time. http://screen/5TxbeZTso4gNDfD Note that the dur column is in nanoseconds, so we convert to seconds. These numbers could further be divided by the number of processes (in my case 48) to get a scaled output. SELECT SUM(dur) / 1000000000.0 FROM slice WHERE name = "sandbox.createFileSystem"; SELECT SUM(dur) / 1000000000.0 FROM slice WHERE name = "sandbox.delete"; SELECT SUM(dur) / 1000000000.0 FROM slice WHERE name = "subprocess.run"; I benchmarked the local compilation of //:skia_public using --config=clang_linux (our custom Linux toolchain). I was sure to clear the Bazel cache before each run and not count the time to download/update a toolchain for the first time. The control measurements (without this CL) are: Wall Time = 272.2s sandbox.createFileSystem = 5466.9s subprocess.run = 2961.0s sandbox.delete = 4112.3s With this CL: Wall Time = 53.9s (5.05x faster) sandbox.createFileSystem = 403.4s subprocess.run = 1610.3s sandbox.delete = 348.8s The control measurement is a touch misleading. Due to it being so slow, I had recommended developers use a ramdisk when building on machines with sufficient RAM via the Bazel flag --sandbox_base=/dev/shm Here is the control measurement when using a RAM disk: Wall Time = 21.2s sandbox.createFileSystem = 58.9s subprocess.run = 705.1s sandbox.delete = 46.6s With this CL and a RAM disk: Wall Time = 19.2s (10% faster) sandbox.createFileSystem = 21.8s subprocess.run = 722.9s sandbox.delete = 16.2s For devs who cannot or are not using a RAM disk, this is a huge win. With a RAM disk, it's a modest improvement. On an RBE build, this had minimal impact. I did my best to bust the action cache with a fake define and the before and after times were about the same. This was inspired by [5] and [6]. [1] Add --profile=/tmp/profile.gz to any command [2] https://bazel.build/rules/performance#performance-profiling [3] https://perfetto.dev/docs/quickstart/trace-analysis#sample-queries [4] https://perfetto.dev/docs/analysis/sql-tables#slice [5]93f21c9ef3[6]311acff345Change-Id: Iceb2606e86111159141a286d01fc002d09fe3fe4 Reviewed-on: https://skia-review.googlesource.com/c/skia/+/547822 Reviewed-by: Leandro Lovisolo <lovisolo@google.com>
107 lines
3.2 KiB
Python
107 lines
3.2 KiB
Python
load("//bazel:macros.bzl", "exports_files_legacy")
|
|
load("@rules_cc//cc:defs.bzl", "cc_toolchain", "cc_toolchain_suite")
|
|
load(":linux_amd64_toolchain_config.bzl", "provide_linux_amd64_toolchain_config")
|
|
load(":mac_m1_toolchain_config.bzl", "provide_mac_m1_toolchain_config")
|
|
|
|
licenses(["notice"])
|
|
|
|
exports_files_legacy()
|
|
|
|
# https://bazel.build/reference/be/c-cpp#cc_toolchain_suite
|
|
# cc_toolchain_suite will fetch deps for toolchains it will not use, which
|
|
# is why we split up the suites by OS. When attempting to fetch "@clang_linux_amd64//:all_files",
|
|
# a script compiled for linux to extract ar archives will break on other platforms.
|
|
cc_toolchain_suite(
|
|
name = "clang_suite_linux",
|
|
toolchains = {
|
|
# The key is target_cpu|compiler
|
|
# compiler appears to be a string we can choose arbitrarily
|
|
# https://bazel.build/reference/command-line-reference?hl=en#flag--compiler
|
|
"k8|host_is_linux_amd64": ":linux_amd64_host",
|
|
"x86_64|host_is_linux_amd64": ":linux_amd64_host",
|
|
"k8": ":linux_amd64_host",
|
|
},
|
|
)
|
|
|
|
cc_toolchain_suite(
|
|
name = "clang_suite_mac",
|
|
toolchains = {
|
|
"arm64|host_is_mac_m1": ":mac_m1_host",
|
|
"darwin|host_is_mac_m1": ":mac_m1_host",
|
|
"darwin_arm64|host_is_mac_m1": ":mac_m1_host",
|
|
},
|
|
)
|
|
|
|
filegroup(name = "not_implemented")
|
|
|
|
filegroup(
|
|
name = "archive_linux_amd64_files",
|
|
srcs = [
|
|
"linux_trampolines/ar_trampoline_linux.sh",
|
|
"@clang_linux_amd64//:archive_files",
|
|
],
|
|
)
|
|
|
|
filegroup(
|
|
name = "compile_linux_amd64_files",
|
|
srcs = [
|
|
"linux_trampolines/IWYU_mapping.imp",
|
|
"linux_trampolines/clang_trampoline_linux.sh",
|
|
"@clang_linux_amd64//:compile_files",
|
|
],
|
|
)
|
|
|
|
filegroup(
|
|
name = "link_linux_amd64_files",
|
|
srcs = [
|
|
# Bazel assumes it is talking to Clang when linking.
|
|
"linux_trampolines/clang_trampoline_linux.sh",
|
|
"@clang_linux_amd64//:link_files",
|
|
],
|
|
)
|
|
|
|
filegroup(
|
|
name = "all_mac_m1_files",
|
|
srcs = [
|
|
"mac_trampolines/ar_trampoline_mac.sh",
|
|
"mac_trampolines/clang_trampoline_mac.sh",
|
|
"mac_trampolines/lld_trampoline_mac.sh",
|
|
"@clang_mac_m1//:all_files",
|
|
],
|
|
)
|
|
|
|
provide_linux_amd64_toolchain_config(
|
|
name = "linux_amd64_toolchain_config",
|
|
)
|
|
|
|
provide_mac_m1_toolchain_config(
|
|
name = "mac_m1_toolchain_config",
|
|
)
|
|
|
|
# https://bazel.build/reference/be/c-cpp#cc_toolchain
|
|
cc_toolchain(
|
|
name = "linux_amd64_host",
|
|
all_files = ":not_implemented",
|
|
ar_files = ":archive_linux_amd64_files",
|
|
compiler_files = ":compile_linux_amd64_files",
|
|
dwp_files = ":not_implemented",
|
|
linker_files = ":link_linux_amd64_files",
|
|
objcopy_files = ":not_implemented",
|
|
strip_files = ":not_implemented",
|
|
supports_param_files = False,
|
|
toolchain_config = ":linux_amd64_toolchain_config",
|
|
)
|
|
|
|
cc_toolchain(
|
|
name = "mac_m1_host",
|
|
all_files = ":all_mac_m1_files",
|
|
ar_files = ":all_mac_m1_files",
|
|
compiler_files = ":all_mac_m1_files",
|
|
dwp_files = ":not_implemented",
|
|
linker_files = ":all_mac_m1_files",
|
|
objcopy_files = ":not_implemented",
|
|
strip_files = ":not_implemented",
|
|
supports_param_files = False,
|
|
toolchain_config = ":mac_m1_toolchain_config",
|
|
)
|