-
Notifications
You must be signed in to change notification settings - Fork 29.6k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
src: use SIMD for snapshot checksum #47145
Conversation
Snapshot checksum verification uses `v8::Internal::Checksum`, which uses zlib's adler32 checksum. However it was using the slow version of adler32 for machines with no SIMD. Let's change it to use the SIMD version when it is available. Fortunately the necessary incantations were already included in our _other_ copy of Chromium's zlib in `deps/zlib/zlib.gyp`, so this is mostly just copying from there. On a very recent Intel machine, this speeds up the snapshot checksumming from 870us to 300us (a 570us startup saving), according to: ```bash for i in {1..100}; do ./node --profile-deserialization -e 0 done \ |& grep 'Verifying snapshot checksum' \ | awk '{s+=$5} END{print s/NR}' ``` P.S.: maybe we should just disable this entirely via `--no-verify-snapshot-checksum`? It doesn't feel like a super necessary protection.
Review requested:
|
Perhaps you can just use deps/zlib and get rid of v8's fork? Code duplication is bad enough, I'd rather not duplicate the magic build incantations as well. :-)
It's off by default in release builds. If you're seeing a performance improvement, it's either because you're testing with a debug build or because of something else (snapshot compression? but that was disabled recently in #45716.) |
I'll give it a go and report back. It's a little weird because v8 expects these the symbols to be defined with the prefix
Yeah, I was wrong: it turns out the flag does not effect it either way. It is controlled by the generated output in diff --git a/deps/v8/src/snapshot/mksnapshot.cc b/deps/v8/src/snapshot/mksnapshot.cc
index dd67969694..23fb00aab9 100644
--- a/deps/v8/src/snapshot/mksnapshot.cc
+++ b/deps/v8/src/snapshot/mksnapshot.cc
@@ -87,7 +87,7 @@ class SnapshotFileWriter {
fprintf(
fp,
"bool Snapshot::ShouldVerifyChecksum(const v8::StartupData* data) {\n");
- fprintf(fp, " return true;\n");
+ fprintf(fp, " return false;\n");
fprintf(fp, "}\n");
fprintf(fp, "} // namespace internal\n");
fprintf(fp, "} // namespace v8\n"); But other than that, this is happening on release builds AFAICT: $ gdb out/Release/node
(gdb) b Cr_z_adler32
Breakpoint 1 at 0x1f7aa20
(gdb) r
Starting program: /home/ubuntu/node/out/Release/node
Thread 1 "node" hit Breakpoint 1, 0x000055555737aa20 in Cr_z_adler32 ()
(gdb) bt
#0 0x000055555737aa20 in Cr_z_adler32 ()
#1 0x000055555674e3dc in v8::internal::Checksum(v8::base::Vector<unsigned char const>) ()
#2 0x000055555674e6d5 in v8::internal::Snapshot::Initialize(v8::internal::Isolate*) [clone .part.0] ()
#3 0x000055555608954b in v8::Isolate::Initialize(v8::Isolate*, v8::Isolate::CreateParams const&) ()
... |
We currently have two copies of Chromium's zlib: one in deps/zlib and another in deps/v8/third_party/zlib. This has a couple of disadvantages: 1. There is an additional cost to keeping both dependencies up-to-date, and in fact they were already out-of-sync (see the refs). 2. People who compile with --shared-zlib (i.e. distro maintainers) will probably not be thrilled to learn that there is still a copy of zlib inside. 3. It's aesthetically unpleasing. Centralize on the version in V8 rather than the one in deps, and delete the one in deps. Basically I just copied deps/zlib/zlib.gyp to tools/v8_gypfiles/zlib.gyp, since the former had a better build configuration (see the refs). This approach seemed better than the opposite approach of centralizing on deps/zlib because: 1. We would need to occasionally bump deps/zlib manually after bumping deps/v8, which seemed annoying. 2. The maintenance steps for bumping zlib seemed more onerous than the maintenance steps for bumping V8. (If we would prefer the opposite approach, I actually have another patch locally.) One discrepancy was that V8's version of zlib had all symbols to be prefixed with `Cr_z_` per deps/v8/third_party/zlib/chromeconf.h, which seemed undesirable, so I added define `CHROMIUM_ZLIB_NO_CHROMECONF` to the build to remove it. (deps/zlib handled this by just commenting out the relevant include, but floating a patch seemed less desirable.) I tested this on Linux with the default build and a --shared-zlib build. I checked that the shared-zlib build correctly linked zlib according to ldd. I would appreciate if the reviewers could suggest some other build configurations to try. Refs: nodejs#47145 Refs: nodejs#47151
}, | ||
}], | ||
], | ||
}], |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It's pretty unclear why it would be safe to assume SSSE3 support here.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
this is copied from deps/zlib/zlib.gyp
; runtime CPU detection is used to decide which SIMD version to call into. (Or are you suggesting that we should add a comment to that effect?)
We currently have two copies of Chromium's zlib: one in deps/zlib and another in deps/v8/third_party/zlib. This has a couple of disadvantages: 1. There is an additional cost to keeping both dependencies up-to-date, and in fact they were already out-of-sync (see the refs). 2. People who compile with --shared-zlib (i.e. distro maintainers) will probably not be thrilled to learn that there is still a copy of zlib inside. 3. It's aesthetically unpleasing. This diff (per discussion in the refs) centralizes on deps/zlib and deletes deps/v8/third_party/zlib. That requires changing the include headers in two files in deps/v8/src, which was pretty straightforward. When the user requests compiling with a shared zlib build, we still need to compile the contents of deps/zlib/google. This is not ideal but it's also not a change in behavior: prior to this diff those files were being compiled in the deps/v8/third_party/zlib version. I tested this on Linux with the default build and a --shared-zlib build. I checked that the shared-zlib build dynamically linked zlib according to ldd, and that the regular build did not. I would appreciate if the reviewers could suggest some other build configurations to try. Refs: nodejs#47145 Refs: nodejs#47157
We currently have two copies of Chromium's zlib: one in deps/zlib and another in deps/v8/third_party/zlib. This has a couple of disadvantages: 1. There is an additional cost to keeping both dependencies up-to-date, and in fact they were already out-of-sync (see the refs). 2. People who compile with --shared-zlib (i.e. distro maintainers) will probably not be thrilled to learn that there is still a copy of zlib inside. 3. It's aesthetically unpleasing. This diff (per discussion in the refs) centralizes on deps/zlib and deletes deps/v8/third_party/zlib. That requires changing the include headers in two files in deps/v8/src, which was pretty straightforward. When the user requests compiling with a shared zlib build, we still need to compile the contents of deps/zlib/google. This is not ideal but it's also not a change in behavior: prior to this diff those files were being compiled in the deps/v8/third_party/zlib version. I tested this on Linux with the default build and a --shared-zlib build. I checked that the shared-zlib build dynamically linked zlib according to ldd, and that the regular build did not. I would appreciate if the reviewers could suggest some other build configurations to try. Refs: nodejs#47145 Refs: nodejs#47157
We currently have two copies of Chromium's zlib: one in deps/zlib and another in deps/v8/third_party/zlib. This has a couple of disadvantages: 1. There is an additional cost to keeping both dependencies up-to-date, and in fact they were already out-of-sync (see the refs). 2. People who compile with --shared-zlib (i.e. distro maintainers) will probably not be thrilled to learn that there is still a copy of zlib inside. 3. It's aesthetically unpleasing. This diff (per discussion in the refs) centralizes on deps/zlib and deletes deps/v8/third_party/zlib. That requires changing the include headers in two files in deps/v8/src, which was pretty straightforward. When the user requests compiling with a shared zlib build, we still need to compile the contents of deps/zlib/google. This is not ideal but it's also not a change in behavior: prior to this diff those files were being compiled in the deps/v8/third_party/zlib version. I tested this on Linux with the default build and a --shared-zlib build. I checked that the shared-zlib build dynamically linked zlib according to ldd, and that the regular build did not. I would appreciate if the reviewers could suggest some other build configurations to try. Refs: nodejs#47145 Refs: nodejs#47157
We currently have two copies of Chromium's zlib: one in deps/zlib and another in deps/v8/third_party/zlib. This has a couple of disadvantages: 1. There is an additional cost to keeping both dependencies up-to-date, and in fact they were already out-of-sync (see the refs). 2. People who compile with --shared-zlib (i.e. distro maintainers) will probably not be thrilled to learn that there is still a copy of zlib inside. 3. It's aesthetically unpleasing. This diff (per discussion in the refs) centralizes on deps/zlib and deletes deps/v8/third_party/zlib. That requires changing the include headers in two files in deps/v8/src, which was pretty straightforward. When the user requests compiling with a shared zlib build, we still need to compile the contents of deps/zlib/google. This is not ideal but it's also not a change in behavior: prior to this diff those files were being compiled in the deps/v8/third_party/zlib version. I tested this on Linux with the default build and a --shared-zlib build. I checked that the shared-zlib build dynamically linked zlib according to ldd, and that the regular build did not. I would appreciate if the reviewers could suggest some other build configurations to try. Refs: nodejs#47145 Refs: nodejs#47157
This is the first of a series of patches. This patch is contains changes to the existing zlib.gyp file to allow it to be used by our v8.gyp. --- We currently have two copies of Chromium's zlib: one in deps/zlib and another in deps/v8/third_party/zlib. This has a couple of disadvantages: 1. There is an additional cost to keeping both dependencies up-to-date, and in fact they were already out-of-sync (see the refs). 2. People who compile with --shared-zlib (i.e. distro maintainers) will probably not be thrilled to learn that there is still a copy of zlib inside. 3. It's aesthetically unpleasing. This diff (per discussion in the refs) centralizes on deps/zlib and deletes deps/v8/third_party/zlib. When the user requests compiling with a shared zlib build, we still need to compile the contents of deps/zlib/google. This is not ideal but it's also not a change in behavior: prior to this diff those files were being compiled in the deps/v8/third_party/zlib version. I tested this on Linux with the default build and a --shared-zlib build. I checked that the shared-zlib build dynamically linked zlib according to ldd, and that the regular build did not. I would appreciate if the reviewers could suggest some other build configurations to try. Refs: nodejs#47145 Refs: nodejs#47157
Closing for #47493. In addition this issue is sort of already fixed on the V8 side by https://chromium-review.googlesource.com/c/v8/v8/+/4571989. |
Snapshot checksum verification uses
v8::Internal::Checksum
, which uses zlib's adler32 checksum. However it was using the slow version of adler32 for machines with no SIMD. Let's change it to use the SIMD version when it is available. Fortunately the necessary incantations were already included in our other copy of Chromium's zlib indeps/zlib/zlib.gyp
, so this is mostly just copying from there.On a very recent Intel machine, this speeds up the snapshot checksumming from 870us to 300us (a 570us startup saving), according to:
P.S.: maybe we should just disable this entirely via
--no-verify-snapshot-checksum
? It doesn't feel like a super necessary protection.