macOS: queue for munmap operations
Subsystems (2)
PATCH 1/1 macOS: queue for munmap operations
From: Koji Nakamaru <[email protected]>
Executing many mmap/munmap calls alternately can cause a huge load on macOS. In order to reduce it, we should temporarily store munmap operations in a queue and process them all at once when the queue is filled. When the program terminates, we can discard any remaining munmap operations as corresponding mmaped regions are automatically reclaimed.
Add a queue for munmap operations to perform them all at once.
Here are some example timings. On the Linux kernel repository that requires about 1700 mmap/munmap calls:
time git ls-tree -r -l --full-tree 211ddde > /dev/null
Before:
real 0m2.083s
user 0m0.201s
sys 0m1.873sAfter:
real 0m0.243s
user 0m0.179s
sys 0m0.052sOn a private repository that requires about 943000 mmap/munmap calls:
time git ls-tree -r -l --full-tree xxxxxxx > /dev/null
Before:
real 27m15.138s
user 0m5.084s
sys 27m9.636sAfter:
real 0m24.209s
user 0m3.055s
sys 0m21.123s---
macOS: queue for munmap operations
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1993%2FKojiNakamaru%2Ffeature%2Fosx-queued-munmap-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1993/KojiNakamaru/feature/osx-queued-munmap-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/1993
Makefile | 1 + compat/osxmmap.c | 49 +++++++++++++++++++++++++++++ compat/posix.h | 7 +++++ contrib/buildsystems/CMakeLists.txt | 4 +++ meson.build | 2 ++ 5 files changed, 63 insertions(+) create mode 100644 compat/osxmmap.c
Makefile +1 -0
@@ -1654,6 +1654,7 @@ ifeq ($(uname_S),Darwin) COMPAT_CFLAGS += -DAPPLE_COMMON_CRYPTO endif PTHREAD_LIBS = + COMPAT_OBJS += compat/osxmmap.o endif ifdef NO_LIBGEN_H
compat/osxmmap.c new file +49 -0
49 lines not on this page · raw message
compat/posix.h +7 -0
@@ -278,6 +278,13 @@ int git_munmap(void *start, size_t length);#include <sys/mman.h> +#if defined(__APPLE__) +#define mmap git_mmap +#define munmap git_munmap +void *git_mmap (void *start,size_t length,int prot,int flags,int fd,off_t offset); +int git_munmap (void *start,size_t length); +#endif +#endif /* NO_MMAP || USE_WIN32_MMAP */ #ifndef MAP_FAILED
contrib/buildsystems/CMakeLists.txt +4 -0
@@ -271,6 +271,10 @@ if(CMAKE_SYSTEM_NAME STREQUAL "Windows") compat/strdup.c) set(NO_UNIX_SOCKETS 1) +elseif(CMAKE_SYSTEM_NAME STREQUAL "Darwin") + list(APPEND compat_SOURCES + compat/osxmmap.c) + elseif(CMAKE_SYSTEM_NAME STREQUAL "Linux") add_compile_definitions(PROCFS_EXECUTABLE_PATH="/proc/self/exe" HAVE_DEV_TTY ) list(APPEND compat_SOURCES unix-socket.c unix-stream-server.c compat/linux/procinfo.c)
meson.build +2 -0
@@ -1275,6 +1275,8 @@ elif host_machine.system() == 'windows' else libgit_sources += 'compat/mingw.c' endif +elif host_machine.system() == 'darwin' + libgit_sources += 'compat/osxmmap.c' endif if host_machine.system() == 'linux'
base-commit: 4253630c6f07a4bdcc9aa62a50e26a4d466219d1
Signature · 1 line
gitgitgadget
Some comments inline, all up to improvements
Quoted 10 lines · Koji Nakamaru via GitGitGadget
From: Koji Nakamaru <[email protected]>
Executing many mmap/munmap calls alternately can cause a huge load on macOS. In order to reduce it, we should temporarily store munmap operations in a queue and process them all at once when the queue is filled. When the program terminates, we can discard any remaining munmap operations as corresponding mmaped regions are automatically reclaimed.
Add a queue for munmap operations to perform them all at once.
Suggestions for rewording:
In order to reduce the peak load store all munmap operations in a queue. Process them all at once (and more efficient) when the queue is filled. The queue may be ignored when the git process terminates. The operating system will do all munmap() when the process exits.
Quoted 65 lines
Here are some example timings. On the Linux kernel repository that requires about 1700 mmap/munmap calls:
time git ls-tree -r -l --full-tree 211ddde > /dev/nullBefore:
real 0m2.083s user 0m0.201s sys 0m1.873sAfter:
real 0m0.243s user 0m0.179s sys 0m0.052sOn a private repository that requires about 943000 mmap/munmap calls:
time git ls-tree -r -l --full-tree xxxxxxx > /dev/nullBefore:
real 27m15.138s user 0m5.084s sys 27m9.636sAfter:
real 0m24.209s user 0m3.055s sys 0m21.123sSigned-off-by Koji Nakamaru <[email protected]>---
macOS: queue for munmap operationsPublished-As: https://github.com/gitgitgadget/git/releases/tag/pr-1993%2FKojiNakamaru%2Ffeature%2Fosx-queued-munmap-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1993/KojiNakamaru/feature/osx-queued-munmap-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/1993Makefile | 1 + compat/osxmmap.c | 49 +++++++++++++++++++++++++++++ compat/posix.h | 7 +++++ contrib/buildsystems/CMakeLists.txt | 4 +++ meson.build | 2 ++ 5 files changed, 63 insertions(+) create mode 100644 compat/osxmmap.cMakefile +1
-0@@ -1654,6 +1654,7 @@ ifeq ($(uname_S),Darwin) COMPAT_CFLAGS += -DAPPLE_COMMON_CRYPTO endif PTHREAD_LIBS = + COMPAT_OBJS += compat/osxmmap.o endif ifdef NO_LIBGEN_Hcompat/osxmmap.c new file +2
-0@@ -0,0 +1,49 @@ +#include <pthread.h> +#include "../git-compat-util.h"
compat/osxmmap.c line 12
Quoted 10 lines
compat/osxmmap.c new file +10
-0+/* We need original mmap/munmap here. */ +#undef mmap +#undef munmap + +/* +* OSX doesn't have any specific setting like Linux's vm.max_map_count, +* so COUNT_MAX can be any large number. We here set it to the default +* value of Linux's vm.max_map_count. +*/ +#define COUNT_MAX (65530)
Why the parantheses ?
And would a less generic name be better, likeMAX_UNMAP_COUNT
compat/osxmmap.c line 24
Quoted 12 lines
compat/osxmmap.c new file +12
-0+ +struct munmap_queue { +void *start; +size_t length; +}; + +void *git_mmap (void *start,size_t length,int prot,int flags,int fd,off_t offset) +{ +/* +* We can simply discard munmap operations in the queue by +* restricting mmap arguments. +*/
Should I read this as
The munmap queue is only ment to defere read-only mappings. And that is what Git does at the moment.
Quoted 30 lines
compat/osxmmap.c new file +25
-0+if (start !=NULL || flags !=MAP_PRIVATE || prot !=PROT_READ ) +die ("invalid usage of mmap" ); +return mmap (start, length, prot, flags, fd, offset); +} + +int git_munmap (void *start,size_t length) +{ +static pthread_mutex_t mutex; +static struct munmap_queue *queue; +static int count; +int i; + +pthread_mutex_lock (&mutex); +if (!queue) + queue =xmalloc (COUNT_MAX *sizeof (struct munmap_queue )); + queue[count].start = start; + queue[count].length = length; +if (++count ==COUNT_MAX ) { +for (i =0 ; i <COUNT_MAX ; i++) +munmap (queue[i].start, queue[i].length); + count =0 ; + } +pthread_mutex_unlock (&mutex); +return 0 ; +}compat/posix.h +0
-0
compat/posix.h int git_munmap(void *start, size_t length); · line 281
Quoted 4 lines
compat/posix.h +1
-0@@ -278,6 +278,13 @@ int git_munmap(void *start, size_t length);#include <sys/mman.h> +#if defined(__APPLE__)
I think it would be better to have a global Makefile knob here. Which
a) allows to take out this patch once the MacOs kernel is improved
b) allows to hook in this code for other OS
Something like DEFER_MUNMAPS - better suggestions welcome
Quoted 42 lines
compat/posix.h +6
-0+#define mmap git_mmap +#define munmap git_munmap +void *git_mmap (void *start,size_t length,int prot,int flags,int fd,off_t offset); +int git_munmap (void *start,size_t length); +#endif +#endif /* NO_MMAP || USE_WIN32_MMAP */ #ifndef MAP_FAILED contrib/buildsystems/CMakeLists.txt +4
-0@@ -271,6 +271,10 @@ if(CMAKE_SYSTEM_NAME STREQUAL "Windows") compat/strdup.c) set(NO_UNIX_SOCKETS 1) +elseif(CMAKE_SYSTEM_NAME STREQUAL "Darwin") + list(APPEND compat_SOURCES + compat/osxmmap.c) + elseif(CMAKE_SYSTEM_NAME STREQUAL "Linux") add_compile_definitions(PROCFS_EXECUTABLE_PATH="/proc/self/exe" HAVE_DEV_TTY ) list(APPEND compat_SOURCES unix-socket.c unix-stream-server.c compat/linux/procinfo.c)meson.build +2
-0@@ -1275,6 +1275,8 @@ elif host_machine.system() == 'windows' else libgit_sources += 'compat/mingw.c' endif +elif host_machine.system() == 'darwin' + libgit_sources += 'compat/osxmmap.c' endif if host_machine.system() == 'linux'base-commit:
4253630c6f07a4bdcc9aa62a50e26a4d466219d1Signature · 1 line
gitgitgadget
Thank you for detailed suggestions. As I discussed in another thread, the root cause of many mmap/munmap calls was simply my ~/.gitconfig, so I'm withdrawing this patch. I'll answer some of your comments below.
Quoted 12 lines · Torsten Bögershausen
Some comments inline, all up to improvements
Quoted 8 lines · Koji Nakamaru via GitGitGadget
From: Koji Nakamaru <[email protected]>
Executing many mmap/munmap calls alternately can cause a huge load on macOS. In order to reduce it, we should temporarily store munmap operations in a queue and process them all at once when the queue is filled. When the program terminates, we can discard any remaining munmap operations as corresponding mmaped regions are automatically reclaimed.
Quoted 8 lines
Quoted 2 lines
Add a queue for munmap operations to perform them all at once.
Suggestions for rewording:
In order to reduce the peak load store all munmap operations in a queue. Process them all at once (and more efficient) when the queue is filled. The queue may be ignored when the git process terminates. The operating system will do allmunmap()when the process exits.
Thank you, it is much clear.
Quoted 71 lines
Quoted 71 lines
Here are some example timings. On the Linux kernel repository that requires about 1700 mmap/munmap calls:
time git ls-tree -r -l --full-tree 211ddde > /dev/nullBefore:
real 0m2.083s user 0m0.201s sys 0m1.873sAfter:
real 0m0.243s user 0m0.179s sys 0m0.052sOn a private repository that requires about 943000 mmap/munmap calls:
time git ls-tree -r -l --full-tree xxxxxxx > /dev/nullBefore:
real 27m15.138s user 0m5.084s sys 27m9.636sAfter:
real 0m24.209s user 0m3.055s sys 0m21.123sSigned-off-by Koji Nakamaru <[email protected]>---
macOS: queue for munmap operationsPublished-As: https://github.com/gitgitgadget/git/releases/tag/pr-1993%2FKojiNakamaru%2Ffeature%2Fosx-queued-munmap-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1993/KojiNakamaru/feature/osx-queued-munmap-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/1993Makefile | 1 + compat/osxmmap.c | 49 +++++++++++++++++++++++++++++ compat/posix.h | 7 +++++ contrib/buildsystems/CMakeLists.txt | 4 +++ meson.build | 2 ++ 5 files changed, 63 insertions(+) create mode 100644 compat/osxmmap.cMakefile +1
-0@@ -1654,6 +1654,7 @@ ifeq ($(uname_S),Darwin) COMPAT_CFLAGS += -DAPPLE_COMMON_CRYPTO endif PTHREAD_LIBS = + COMPAT_OBJS += compat/osxmmap.o endif ifdef NO_LIBGEN_Hcompat/osxmmap.c new file +8
-0@@ -0,0 +1,49 @@ +#include <pthread.h> +#include "../git-compat-util.h" +/* We need original mmap/munmap here. */ +#undef mmap +#undef munmap + +/* +* OSX doesn't have any specific setting like Linux's vm.max_map_count,
Quoted 8 lines
compat/osxmmap.c line 12
Quoted 4 lines
compat/osxmmap.c new file +4
-0+* so COUNT_MAX can be any large number. We here set it to the default +* value of Linux's vm.max_map_count. +*/ +#define COUNT_MAX (65530)Why the parantheses ?
And would a less generic name be better, likeMAX_UNMAP_COUNT
The parentheses are not required but I prefer them as discussed in [1]. I agree MAX_UNMAP_COUNT is more clear.
Quoted 7 lines
Quoted 7 lines
compat/osxmmap.c new file +7
-0+ +struct munmap_queue { +void *start; +size_t length; +}; + +void *git_mmap (void *start,size_t length,int prot,int flags,int fd,off_t offset)
Quoted 8 lines
compat/osxmmap.c line 24
Quoted 5 lines
compat/osxmmap.c new file +5
-0+{ +/* +* We can simply discard munmap operations in the queue by +* restricting mmap arguments. +*/ Should I read this as
The munmap queue is only ment to defere read-only mappings. And that is what Git does at the moment.
Yes. This part is actually borrowed from compat/mmap.c and I've also verified that the predicate is valid by searching all mmap calls.
Quoted 31 lines
Quoted 31 lines
compat/osxmmap.c new file +25
-0+if (start !=NULL || flags !=MAP_PRIVATE || prot !=PROT_READ ) +die ("invalid usage of mmap" ); +return mmap (start, length, prot, flags, fd, offset); +} + +int git_munmap (void *start,size_t length) +{ +static pthread_mutex_t mutex; +static struct munmap_queue *queue; +static int count; +int i; + +pthread_mutex_lock (&mutex); +if (!queue) + queue =xmalloc (COUNT_MAX *sizeof (struct munmap_queue )); + queue[count].start = start; + queue[count].length = length; +if (++count ==COUNT_MAX ) { +for (i =0 ; i <COUNT_MAX ; i++) +munmap (queue[i].start, queue[i].length); + count =0 ; + } +pthread_mutex_unlock (&mutex); +return 0 ; +}compat/posix.h +0
-0@@ -278,6 +278,13 @@ int git_munmap(void *start, size_t length);
Quoted 8 lines
compat/posix.h int git_munmap(void *start, size_t length); · line 281
Quoted 3 lines
compat/posix.h +1
-0#include <sys/mman.h> +#if defined(__APPLE__)I think it would be better to have a global
Makefileknob here. Which
a) allows to take out this patch once the MacOs kernel is improved
b) allows to hook in this code for other OS
Something likeDEFER_MUNMAPS- better suggestions welcome
I followed your suggestion and adjusted code and Makefile, etc. (locally)
Quoted 1 line
Quoted 1 line
[snip]
Quoted 8 lines · Koji Nakamaru via GitGitGadget
From: Koji Nakamaru <[email protected]>
Executing many mmap/munmap calls alternately can cause a huge load on macOS. In order to reduce it, we should temporarily store munmap operations in a queue and process them all at once when the queue is filled. When the program terminates, we can discard any remaining munmap operations as corresponding mmaped regions are automatically reclaimed.
Quoted 6 lines
Add a queue for munmap operations to perform them all at once.
Here are some example timings. On the Linux kernel repository that requires about 1700 mmap/munmap calls:
time git ls-tree -r -l --full-tree 211ddde > /dev/null
Why is it doing so many mmap calls? Do you have a ton of loose objects? We have to mmap loose objects individually (because they're all in separate files), but each pack only gets a single map (well, there's a window parameter, but it's 1GB on 64-bit systems, so you should get a handful of maps at most).
If you run "git gc", how does the resulting ls-tree perform? I have only 27 mmap() calls on my system.
I know that running "git gc" is relatively expensive, but it is also bringing other optimizations (like the fact that we don't have to open() and map each of those files in the first place!).
Quoted 3 lines
On a private repository that requires about 943000 mmap/munmap calls:
time git ls-tree -r -l --full-tree xxxxxxx > /dev/null
Ditto here. I'd be curious how well packed the repo is, and how it does after a repack. If it has a very large packfile, you might also try:
git config core.packedGitWindowSize 4G
or similar (though for just an ls-tree, we should only be looking at tree objects, which in general I'd expect to be in a confined area of the packfile; so the 1GB window is probably plenty).
Quoted 10 lines
compat/osxmmap.c new file +10
-0+int git_munmap (void *start,size_t length) +{ +static pthread_mutex_t mutex; +static struct munmap_queue *queue; +static int count; +int i; + +pthread_mutex_lock (&mutex); +if (!queue) + queue =xmalloc (COUNT_MAX *sizeof (struct munmap_queue ));
compat/osxmmap.c line 49
Quoted 10 lines
compat/osxmmap.c new file +10
-0+ queue[count].start = start; + queue[count].length = length; +if (++count ==COUNT_MAX ) { +for (i =0 ; i <COUNT_MAX ; i++) +munmap (queue[i].start, queue[i].length); + count =0 ; + } +pthread_mutex_unlock (&mutex); +return 0 ; +}
Does batching those unmaps actually make them faster? Or is it just that the commands you showed did not fill the queue, so we essentially just leaked all of those maps until the program exited?
If the latter, then I'd wonder:
1. Does this increase memory pressure, since the OS has no idea we're not actually interested in those maps anymore? Some of them can be quite large, if the command is looking at blobs.
2. How does it perform on a command that actually fills the queue? I guess something like "git log --raw" might do it (though if my guesses above are right, you'd need on the order of 64,000 loose trees).
-Peff
Thank you for pointing out many unusual mmap calls and other details. As discussed below, the root cause was simply my ~/.gitconfig. This patch may be useful in some rare/edge cases but a somewhat unusual hack, so I'm withdrawing it.
Quoted 35 lines · Jeff King
Quoted 8 lines · Koji Nakamaru via GitGitGadget
From: Koji Nakamaru <[email protected]>
Executing many mmap/munmap calls alternately can cause a huge load on macOS. In order to reduce it, we should temporarily store munmap operations in a queue and process them all at once when the queue is filled. When the program terminates, we can discard any remaining munmap operations as corresponding mmaped regions are automatically reclaimed.
Quoted 6 lines
Add a queue for munmap operations to perform them all at once.
Here are some example timings. On the Linux kernel repository that requires about 1700 mmap/munmap calls:
time git ls-tree -r -l --full-tree 211ddde > /dev/nullWhy is it doing so many mmap calls? Do you have a ton of loose objects? We have to mmap loose objects individually (because they're all in separate files), but each pack only gets a single map (well, there's a window parameter, but it's 1GB on 64-bit systems, so you should get a handful of maps at most).
If you run "git gc", how does the resulting ls-tree perform? I have only 27
mmap()calls on my system.I know that running "git gc" is relatively expensive, but it is also bringing other optimizations (like the fact that we don't have to
open()and map each of those files in the first place!).Quoted 3 lines
On a private repository that requires about 943000 mmap/munmap calls:
time git ls-tree -r -l --full-tree xxxxxxx > /dev/null
Quoted 8 lines
Ditto here. I'd be curious how well packed the repo is, and how it does after a repack. If it has a very large packfile, you might also try:
git config core.packedGitWindowSize 4G
or similar (though for just an ls-tree, we should only be looking at tree objects, which in general I'd expect to be in a confined area of the packfile; so the 1GB window is probably plenty).
Following your suggestion, I investigated the number of mmap calls in other environments and found much smaller counts. I tracked how xmmap_gently() was called in packfile.c and found settings->packed_git_window_size was different between environments. My ~/.gitconfig defined "packedGitLimit = 128m" and this caused many calls.
Quoted 27 lines
Quoted 10 lines
compat/osxmmap.c new file +10
-0+int git_munmap (void *start,size_t length) +{ +static pthread_mutex_t mutex; +static struct munmap_queue *queue; +static int count; +int i; + +pthread_mutex_lock (&mutex); +if (!queue) + queue =xmalloc (COUNT_MAX *sizeof (struct munmap_queue ));compat/osxmmap.c line 49
Quoted 10 lines
compat/osxmmap.c new file +10
-0+ queue[count].start = start; + queue[count].length = length; +if (++count ==COUNT_MAX ) { +for (i =0 ; i <COUNT_MAX ; i++) +munmap (queue[i].start, queue[i].length); + count =0 ; + } +pthread_mutex_unlock (&mutex); +return 0 ; +}Does batching those unmaps actually make them faster? Or is it just that the commands you showed did not fill the queue, so we essentially just leaked all of those maps until the program exited?
If the latter, then I'd wonder:
Quoted 8 lines
1. Does this increase memory pressure, since the OS has no idea we're not actually interested in those maps anymore? Some of them can be quite large, if the command is looking at blobs.
2. How does it perform on a command that actually fills the queue? I guess something like "git log --raw" might do it (though if my guesses above are right, you'd need on the order of 64,000 loose trees).
In my extreme cases, this batching makes them faster. Queue flushing has occurred several times for the private repository case and not occurred for the Linux kernel case. Though I haven't investigated in detail, memory pressure doesn't seem to be critical (and it could also be possible to adopt smarter thresholds).
I tested git log --raw for the Linux kernel repository. For reference, the results are shown below:
without "packedGitLimit = 128m":
mmap 9
# without batching
real 1m3.970s
user 1m2.232s
sys 0m1.725s
# with batching
real 1m5.991s
user 0m58.637s
sys 0m4.315swith "packedGitLimit = 128m":
mmap 3072538
# without batching
(It took too long so I stopped the execution)
real 518m6.928s
user 0m41.126s
sys 517m24.072s
# with batching
real 2m26.276s
user 1m8.495s
sys 1m3.230sQuoted 9 lines · Koji Nakamaru
Quoted 8 lines
Ditto here. I'd be curious how well packed the repo is, and how it does after a repack. If it has a very large packfile, you might also try:
git config core.packedGitWindowSize 4G
or similar (though for just an ls-tree, we should only be looking at tree objects, which in general I'd expect to be in a confined area of the packfile; so the 1GB window is probably plenty).
Quoted 5 lines
Following your suggestion, I investigated the number of mmap calls in other environments and found much smaller counts. I tracked how
xmmap_gently()was called inpackfile.cand foundsettings->packed_git_window_sizewas different between environments. My ~/.gitconfig defined "packedGitLimit = 128m" and this caused many calls.
Ah, very interesting. Yes, I think that helps explain why there were so many mmap calls. I don't think there's a good reason to lower that number in general, assuming the OS is reasonably good at dropping mapped pages from RAM when there's memory pressure.
Quoted 5 lines
In my extreme cases, this batching makes them faster. Queue flushing has occurred several times for the private repository case and not occurred for the Linux kernel case. Though I haven't investigated in detail, memory pressure doesn't seem to be critical (and it could also be possible to adopt smarter thresholds).
OK, that's quite interesting that batching makes such a difference. I guess somebody with more knowledge of macOS kernel internals could probably explain it. Though it sounds like your problem was sufficiently solved by dropping the extra config, it's a good fact for us to know about in general.
-Peff