Skip to content
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.

Sign up for GitHub

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

Jump to bottom

Numa #1556

Merged
merged 20 commits into from
Jun 26, 2023
Merged

Numa #1556

Changes from 1 commit
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
6fc5f17
detect NUMA systems and pin work threads to nodes (linux)
zrm May 21, 2023
0d23f8c
disable mmap prefetch/readahead for NUMA systems
zrm May 21, 2023
9d058c2
avoid sending finalize op to thread pool if it does nothing
zrm May 21, 2023
2c1b5ae
silence robot
zrm May 23, 2023
8502d51
fix args
zrm May 23, 2023
bf83dcb
make --numa a param
zrm Jun 17, 2023
b71dfe6
recommendation that n_nodes evenly divide n_threads did not warrant s…
zrm Jun 17, 2023
adaad10
lower synchronization overhead
zrm Jun 18, 2023
c31d51d
statically allocate
zrm Jun 18, 2023
2f5bb46
move numa state to g_state
zrm Jun 18, 2023
4b94582
add description for --numa
zrm Jun 18, 2023
90a0e65
Merge branch 'master' into HEAD
ggerganov Jun 19, 2023
d0e3596
ggml : minor style changes
ggerganov Jun 19, 2023
67ba34e
ggml : minor style + try fix sanitizer build
ggerganov Jun 19, 2023
8f98035
Merge branch 'master' into HEAD
ggerganov Jun 26, 2023
0fe4b00
llama : allow to initialize backend with NUMA support
ggerganov Jun 26, 2023
875a1e1
llama : avoid ggml include in llama-util.h
ggerganov Jun 26, 2023
4a555b4
ggml : style / formatting
ggerganov Jun 26, 2023
81a40e9
ggml : fix handling of ops with n_threads > n_tasks > 1
ggerganov Jun 26, 2023
9aec2b7
server : utilize numa parameter
ggerganov Jun 26, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
disable mmap prefetch/readahead for NUMA systems
  • Loading branch information
@zrm
zrm committed May 21, 2023
commit 0d23f8ce8da6bb13d557233c7b76a5563055fcf5
15 changes: 14 additions & 1 deletion llama-util.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -163,6 +163,9 @@ static std::string llama_format_win_err(DWORD err) {
}
#endif

extern "C" {
bool ggml_is_numa();
}
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This has to be avoided. Probably utilize new GGML_USE_NUMA define

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think this function is still needed, because even if NUMA support is compiled in, there are some operations that only make sense if the system actually has more than one NUMA node. It also makes it easier to disable those operations, because that function always returns false if ggml_numa_init() is never called, so we can add a --numa option and not call ggml_numa_init() otherwise, causing those operations to be disabled. The latest commits do this.

struct llama_mmap {
void * addr;
size_t size;
Expand All @@ -176,8 +179,10 @@ struct llama_mmap {
size = file->size;
int fd = fileno(file->fp);
int flags = MAP_SHARED;
// prefetch/readahead impairs performance on NUMA systems
if (ggml_is_numa()) prefetch = 0;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

warning: statement should be inside braces [readability-braces-around-statements]

Suggested change
if (ggml_is_numa()) prefetch = 0;
if (ggml_is_numa()) { prefetch = 0;
}

#ifdef __linux__
flags |= MAP_POPULATE;
if (prefetch) flags |= MAP_POPULATE;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

warning: statement should be inside braces [readability-braces-around-statements]

Suggested change
if (prefetch) flags |= MAP_POPULATE;
if (prefetch) { flags |= MAP_POPULATE;
}

#endif
addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
if (addr == MAP_FAILED) {
Expand All @@ -191,6 +196,14 @@ struct llama_mmap {
strerror(errno));
}
}
if (ggml_is_numa()) {
// advise the kernel not to use readahead
// (because the next page might not belong on the same node)
if (madvise(addr, file->size, MADV_RANDOM)) {
fprintf(stderr, "warning: madvise(.., MADV_RANDOM) failed: %s\n",
strerror(errno));
}
}
}

~llama_mmap() {
Expand Down