git-commit-vandalism/alloc.c

117 lines
2.6 KiB
C
Raw Normal View History

Add specialized object allocator This creates a simple specialized object allocator for basic objects. This avoids wasting space with malloc overhead (metadata and extra alignment), since the specialized allocator knows the alignment, and that objects, once allocated, are never freed. It also allows us to track some basic statistics about object allocations. For example, for the mozilla import, it shows object usage as follows: blobs: 627629 (14710 kB) trees: 1119035 (34969 kB) commits: 196423 (8440 kB) tags: 1336 (46 kB) and the simpler allocator shaves off about 2.5% off the memory footprint off a "git-rev-list --all --objects", and is a bit faster too. [ Side note: this concludes the series of "save memory in object storage". The thing is, there simply isn't much more to be saved on the objects. Doing "git-rev-list --all --objects" on the mozilla archive has a final total RSS of 131498 pages for me: that's about 513MB. Of that, the object overhead is now just 56MB, the rest is going somewhere else (put another way: the fact that this patch shaves off 2.5% of the total memory overhead, considering that objects are now not much more than 10% of the total shows how big the wasted space really was: this makes object allocations much more memory- and time-efficient). I haven't looked at where the rest is, but I suspect the bulk of it is just the pack-file loading. It may be that we should pack the tree objects separately from the blob objects: for git-rev-list --objects, we don't actually ever need to even look at the blobs, but since trees and blobs are interspersed in the pack-file, we end up not being dense in the tree accesses, so we end up looking at more pages than we strictly need to. So with a 535MB pack-file, it's entirely possible - even likely - that most of the remaining RSS is just the mmap of the pack-file itself. We don't need to map in _all_ of it, but we do end up mapping a fair amount. ] Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2006-06-19 19:44:15 +02:00
/*
* alloc.c - specialized allocator for internal objects
*
* Copyright (C) 2006 Linus Torvalds
*
* The standard malloc/free wastes too much space for objects, partly because
* it maintains all the allocation infrastructure (which isn't needed, since
* we never free an object descriptor anyway), but even more because it ends
* up with maximal alignment because it doesn't know what the object alignment
* for the new allocation is.
*/
#include "cache.h"
#include "object.h"
#include "blob.h"
#include "tree.h"
#include "commit.h"
#include "tag.h"
#define BLOCKING 1024
union any_object {
struct object object;
struct blob blob;
struct tree tree;
struct commit commit;
struct tag tag;
};
struct alloc_state {
int count; /* total number of nodes allocated */
int nr; /* number of nodes left in current allocation */
void *p; /* first free node in current allocation */
};
static inline void *alloc_node(struct alloc_state *s, size_t node_size)
{
void *ret;
if (!s->nr) {
s->nr = BLOCKING;
s->p = xmalloc(BLOCKING * node_size);
}
s->nr--;
s->count++;
ret = s->p;
s->p = (char *)s->p + node_size;
memset(ret, 0, node_size);
return ret;
}
static struct alloc_state blob_state;
void *alloc_blob_node(void)
{
struct blob *b = alloc_node(&blob_state, sizeof(struct blob));
b->object.type = OBJ_BLOB;
return b;
}
static struct alloc_state tree_state;
void *alloc_tree_node(void)
{
struct tree *t = alloc_node(&tree_state, sizeof(struct tree));
t->object.type = OBJ_TREE;
return t;
}
static struct alloc_state tag_state;
void *alloc_tag_node(void)
{
struct tag *t = alloc_node(&tag_state, sizeof(struct tag));
t->object.type = OBJ_TAG;
return t;
}
static struct alloc_state object_state;
void *alloc_object_node(void)
{
struct object *obj = alloc_node(&object_state, sizeof(union any_object));
obj->type = OBJ_NONE;
return obj;
}
Add specialized object allocator This creates a simple specialized object allocator for basic objects. This avoids wasting space with malloc overhead (metadata and extra alignment), since the specialized allocator knows the alignment, and that objects, once allocated, are never freed. It also allows us to track some basic statistics about object allocations. For example, for the mozilla import, it shows object usage as follows: blobs: 627629 (14710 kB) trees: 1119035 (34969 kB) commits: 196423 (8440 kB) tags: 1336 (46 kB) and the simpler allocator shaves off about 2.5% off the memory footprint off a "git-rev-list --all --objects", and is a bit faster too. [ Side note: this concludes the series of "save memory in object storage". The thing is, there simply isn't much more to be saved on the objects. Doing "git-rev-list --all --objects" on the mozilla archive has a final total RSS of 131498 pages for me: that's about 513MB. Of that, the object overhead is now just 56MB, the rest is going somewhere else (put another way: the fact that this patch shaves off 2.5% of the total memory overhead, considering that objects are now not much more than 10% of the total shows how big the wasted space really was: this makes object allocations much more memory- and time-efficient). I haven't looked at where the rest is, but I suspect the bulk of it is just the pack-file loading. It may be that we should pack the tree objects separately from the blob objects: for git-rev-list --objects, we don't actually ever need to even look at the blobs, but since trees and blobs are interspersed in the pack-file, we end up not being dense in the tree accesses, so we end up looking at more pages than we strictly need to. So with a 535MB pack-file, it's entirely possible - even likely - that most of the remaining RSS is just the mmap of the pack-file itself. We don't need to map in _all_ of it, but we do end up mapping a fair amount. ] Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2006-06-19 19:44:15 +02:00
static struct alloc_state commit_state;
unsigned int alloc_commit_index(void)
{
static unsigned int count;
return count++;
}
void *alloc_commit_node(void)
{
struct commit *c = alloc_node(&commit_state, sizeof(struct commit));
c->object.type = OBJ_COMMIT;
c->index = alloc_commit_index();
c->graph_pos = COMMIT_NOT_FROM_GRAPH;
return c;
}
static void report(const char *name, unsigned int count, size_t size)
{
fprintf(stderr, "%10s: %8u (%"PRIuMAX" kB)\n",
name, count, (uintmax_t) size);
}
#define REPORT(name, type) \
report(#name, name##_state.count, name##_state.count * sizeof(type) >> 10)
Add specialized object allocator This creates a simple specialized object allocator for basic objects. This avoids wasting space with malloc overhead (metadata and extra alignment), since the specialized allocator knows the alignment, and that objects, once allocated, are never freed. It also allows us to track some basic statistics about object allocations. For example, for the mozilla import, it shows object usage as follows: blobs: 627629 (14710 kB) trees: 1119035 (34969 kB) commits: 196423 (8440 kB) tags: 1336 (46 kB) and the simpler allocator shaves off about 2.5% off the memory footprint off a "git-rev-list --all --objects", and is a bit faster too. [ Side note: this concludes the series of "save memory in object storage". The thing is, there simply isn't much more to be saved on the objects. Doing "git-rev-list --all --objects" on the mozilla archive has a final total RSS of 131498 pages for me: that's about 513MB. Of that, the object overhead is now just 56MB, the rest is going somewhere else (put another way: the fact that this patch shaves off 2.5% of the total memory overhead, considering that objects are now not much more than 10% of the total shows how big the wasted space really was: this makes object allocations much more memory- and time-efficient). I haven't looked at where the rest is, but I suspect the bulk of it is just the pack-file loading. It may be that we should pack the tree objects separately from the blob objects: for git-rev-list --objects, we don't actually ever need to even look at the blobs, but since trees and blobs are interspersed in the pack-file, we end up not being dense in the tree accesses, so we end up looking at more pages than we strictly need to. So with a 535MB pack-file, it's entirely possible - even likely - that most of the remaining RSS is just the mmap of the pack-file itself. We don't need to map in _all_ of it, but we do end up mapping a fair amount. ] Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2006-06-19 19:44:15 +02:00
void alloc_report(void)
{
REPORT(blob, struct blob);
REPORT(tree, struct tree);
REPORT(commit, struct commit);
REPORT(tag, struct tag);
REPORT(object, union any_object);
Add specialized object allocator This creates a simple specialized object allocator for basic objects. This avoids wasting space with malloc overhead (metadata and extra alignment), since the specialized allocator knows the alignment, and that objects, once allocated, are never freed. It also allows us to track some basic statistics about object allocations. For example, for the mozilla import, it shows object usage as follows: blobs: 627629 (14710 kB) trees: 1119035 (34969 kB) commits: 196423 (8440 kB) tags: 1336 (46 kB) and the simpler allocator shaves off about 2.5% off the memory footprint off a "git-rev-list --all --objects", and is a bit faster too. [ Side note: this concludes the series of "save memory in object storage". The thing is, there simply isn't much more to be saved on the objects. Doing "git-rev-list --all --objects" on the mozilla archive has a final total RSS of 131498 pages for me: that's about 513MB. Of that, the object overhead is now just 56MB, the rest is going somewhere else (put another way: the fact that this patch shaves off 2.5% of the total memory overhead, considering that objects are now not much more than 10% of the total shows how big the wasted space really was: this makes object allocations much more memory- and time-efficient). I haven't looked at where the rest is, but I suspect the bulk of it is just the pack-file loading. It may be that we should pack the tree objects separately from the blob objects: for git-rev-list --objects, we don't actually ever need to even look at the blobs, but since trees and blobs are interspersed in the pack-file, we end up not being dense in the tree accesses, so we end up looking at more pages than we strictly need to. So with a 535MB pack-file, it's entirely possible - even likely - that most of the remaining RSS is just the mmap of the pack-file itself. We don't need to map in _all_ of it, but we do end up mapping a fair amount. ] Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2006-06-19 19:44:15 +02:00
}