Re: [PATCH 10/10] pack-revindex: radix-sort the revindex
On Wed, Jul 10, 2013 at 06:47:49PM +0530, Ramkumar Ramachandra wrote: For a 64-bit off_t, using 16-bit digits gives us k=4. Wait, isn't off_t a signed data type? Did you account for that in your algorithm? It is signed, but the values we are storing in the revindex are all positive file offsets. Right-shifting a positive signed type is explicitly allowed in C. -static int cmp_offset(const void *a_, const void *b_) +/* + * This is a least-significant-digit radix sort. + */ Any particular reason for choosing LSD, and not MSD? Simplicity. An MSD implementation should have the same algorithmic complexity and in theory, one can do MSD in-place. I'm happy enough with the speedup here, but if you want to take a stab at beating my times with MSD, please feel free. The other usual downside of MSD is that it is typically not stable, but we don't care about that here. We know that our sort keys are unique. +#define DIGIT_SIZE (16) +#define BUCKETS (1 DIGIT_SIZE) Okay, NUMBER_OF_BUCKETS = 2^RADIX, and you choose a hex radix. Is off_t guaranteed to be fixed-length though? I thought only the ones in stdint.h were guaranteed to be fixed-length? I'm not sure what you mean by fixed-length. If you mean does it have the same size on every platform, then no. It will typically be 32-bit on platforms without largefile support, and 64-bit elsewhere. But it shouldn't matter. We'll first sort the entries by the lower 16 bits, and then if we have more bits, by the next 16 bits, and so on. We quit when the maximum value to sort (which we know ahead of time from the size of the packfile) is smaller than the 16-bits we are on. So we don't need to know the exact size of off_t, only the maximum value in our list (which must, by definition, be smaller than what can be represented by off_t). + /* +* We want to know the bucket that a[i] will go into when we are using +* the digit that is N bits from the (least significant) end. +*/ +#define BUCKET_FOR(a, i, bits) (((a)[(i)].offset (bits)) (BUCKETS-1)) Ouch! This is unreadable. Just write an inline function instead? A % would've been easier on the eyes, but you chose base-16. I specifically avoided an inline function because they are subject to compiler settings. This isn't just it would be a bit faster if this got inlined, and OK otherwise but this would be horribly slow if not inlined. I'm also not sure that static inline unsigned bucket_for(const struct revindex *a, unsigned i, unsigned bits) { return a[i].offset bits (BUCKETS-1); } is actually any more readable. I'm not sure what you mean by base-16. No matter the radix digit size, as long as it is an integral number of bits, we can mask it off, which is more efficient than modulo. A good compiler should see that it is a constant and convert it to a bit-mask, but I'm not sure I agree that modular arithmetic is more readable. This is fundamentally a bit-twiddling operation, as we are shifting and masking. I tried to explain it in the comment; suggestions to improve that are welcome. + /* +* We need O(n) temporary storage, so we sort back and forth between +* the real array and our tmp storage. To keep them straight, we always +* sort from a into buckets in b. +*/ + struct revindex_entry *tmp = xcalloc(n, sizeof(*tmp)); Shouldn't this be sizeof (struct revindex_entry), since tmp hasn't been declared yet? No, the variable is declared (but uninitialized) in its initializer. Despite its syntax, sizeof() is not a function and does not care about the state of the variable, only its type. Also, s/n/revindex_nr/, and something more appropriate for tmp? What name would you suggest be be more appropriate for tmp? + int bits = 0; + unsigned *pos = xmalloc(BUCKETS * sizeof(*pos)); sizeof(unsigned int), for clarity, if not anything else. I disagree; in general, I prefer using sizeof(*var) rather than sizeof(type), because it avoids repeating ourselves, and there is no compile-time check that you have gotten it right. In the initializer it is less important, because the type is right there. But when you are later doing: memset(pos, 0, BUCKETS * sizeof(*pos)); this is much more robust. If somebody changes the type of pos, the memset line does not need changed. If you used sizeof(unsigned), then the code is now buggy (and the compiler cannot notice). You picked malloc over calloc here, because you didn't want to incur the extra cost of zero-initializing the memory? Yes. We have to zero-initialize in each loop, so there is no point spending the extra effort on calloc. We could also xcalloc inside each loop iteration, but since we need the same-size allocation each time, I hoisted the malloc out of the loop. Also, pos is the actual buckets array, I presume (hence unsigned,
Re: [PATCH 10/10] pack-revindex: radix-sort the revindex
On Wed, Jul 10, 2013 at 10:10:16AM -0700, Brandon Casey wrote: On the linux.git repo, with about 3M objects to sort, this yields a 400% speedup. Here are the best-of-five numbers for running echo HEAD | git cat-file --batch-disk-size, which is dominated by time spent building the pack revindex: before after real0m0.834s 0m0.204s user0m0.788s 0m0.164s sys 0m0.040s 0m0.036s On a smaller repo, the radix sort would not be as impressive (and could even be worse), as we are trading the log(n) factor for the k=4 of the radix sort. However, even on git.git, with 173K objects, it shows some improvement: before after real0m0.046s 0m0.017s user0m0.036s 0m0.012s sys 0m0.008s 0m0.000s k should only be 2 for git.git. I haven't packed in a while, but I think it should all fit within 4G. :) The pathological case would be a pack file with very few very very large objects, large enough to push the pack size over the 2^48 threshold so we'd have to do all four radixes. Yeah, even linux.git fits into k=2. And that does more or less explain the numbers in both cases. For git.git, With 173K objects, log(n) is ~18, so regular sort is 18n. With a radix sort of k=2, which has a constant factor of 2 (you can see by looking at the code that we go through the list twice per radix), we have 4n. So there should be a 4.5x speedup. We don't quite get that, which is probably due to the extra bookkeeping on the buckets. For linux.git, with 3M objects, log(n) is ~22, so the speedup we hope for is 5.5x. We end up with 4x. It's probably worth mentioning here and/or in the code that k is dependent on the pack file size and that we can jump out early for small pack files. That's my favorite part of this code by the way. :) Yeah, I agree it is probably worth mentioning along with the numbers; it is where half of our speedup is coming from. I think the max bits loop condition deserves to be commented, too. I'll add that. Also note that my commit message still refers to --batch-disk-size which does not exist anymore. :) I didn't update the timings in the commit message for my re-roll, but I did confirm that they are the same. + /* +* We need O(n) temporary storage, so we sort back and forth between +* the real array and our tmp storage. To keep them straight, we always +* sort from a into buckets in b. +*/ + struct revindex_entry *tmp = xcalloc(n, sizeof(*tmp)); Didn't notice it the first time I read this, but do we really need calloc here? Or will malloc do? No, a malloc should be fine. I doubt it matters much, but there's no reason not to go the cheap route. + struct revindex_entry *a = entries, *b = tmp; + int bits = 0; + unsigned *pos = xmalloc(BUCKETS * sizeof(*pos)); + + while (max bits) { + struct revindex_entry *swap; + int i; You forgot to make i unsigned. See below too... Oops. Thanks for catching. + /* +* Now we can drop the elements into their correct buckets (in +* our temporary array). We iterate the pos counter backwards +* to avoid using an extra index to count up. And since we are +* going backwards there, we must also go backwards through the +* array itself, to keep the sort stable. +*/ + for (i = n - 1; i = 0; i--) + b[--pos[BUCKET_FOR(a, i, bits)]] = a[i]; ...which is why the above loop still works. Since we are iterating by ones, I guess I can just compare to UINT_MAX. -Peff -- To unsubscribe from this list: send the line unsubscribe git in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
[PATCH 10/10] pack-revindex: radix-sort the revindex
The pack revindex stores the offsets of the objects in the pack in sorted order, allowing us to easily find the on-disk size of each object. To compute it, we populate an array with the offsets from the sha1-sorted idx file, and then use qsort to order it by offsets. That does O(n log n) offset comparisons, and profiling shows that we spend most of our time in cmp_offset. However, since we are sorting on a simple off_t, we can use numeric sorts that perform better. A radix sort can run in O(k*n), where k is the number of digits in our number. For a 64-bit off_t, using 16-bit digits gives us k=4. On the linux.git repo, with about 3M objects to sort, this yields a 400% speedup. Here are the best-of-five numbers for running echo HEAD | git cat-file --batch-disk-size, which is dominated by time spent building the pack revindex: before after real0m0.834s 0m0.204s user0m0.788s 0m0.164s sys 0m0.040s 0m0.036s On a smaller repo, the radix sort would not be as impressive (and could even be worse), as we are trading the log(n) factor for the k=4 of the radix sort. However, even on git.git, with 173K objects, it shows some improvement: before after real0m0.046s 0m0.017s user0m0.036s 0m0.012s sys 0m0.008s 0m0.000s Signed-off-by: Jeff King p...@peff.net --- I changed a few things from the original, including: 1. We take an unsigned number of objects to match the fix in the last patch. 2. The 16-bit digit size is factored out to a single place, which avoids magic numbers and repeating ourselves. 3. The digits variable is renamed to bits, which is more accurate. 4. The outer loop condition uses the simpler while (max bits). 5. We use memcpy instead of an open-coded loop to copy the whole array at the end. The individual bucket-assignment is still done by struct assignment. I haven't timed if memcpy would make a difference there. 6. The 64K*sizeof(int) pos array is now heap-allocated, in case there are platforms with a small stack. I re-ran my timings to make sure none of the above impacted them; it turned out the same. pack-revindex.c | 84 + 1 file changed, 79 insertions(+), 5 deletions(-) diff --git a/pack-revindex.c b/pack-revindex.c index 1aa9754..9365bc2 100644 --- a/pack-revindex.c +++ b/pack-revindex.c @@ -59,11 +59,85 @@ static int cmp_offset(const void *a_, const void *b_) /* revindex elements are lazily initialized */ } -static int cmp_offset(const void *a_, const void *b_) +/* + * This is a least-significant-digit radix sort. + */ +static void sort_revindex(struct revindex_entry *entries, unsigned n, off_t max) { - const struct revindex_entry *a = a_; - const struct revindex_entry *b = b_; - return (a-offset b-offset) ? -1 : (a-offset b-offset) ? 1 : 0; + /* +* We use a digit size of 16 bits. That keeps our memory +* usage reasonable, and we can generally (for a 4G or smaller +* packfile) quit after two rounds of radix-sorting. +*/ +#define DIGIT_SIZE (16) +#define BUCKETS (1 DIGIT_SIZE) + /* +* We want to know the bucket that a[i] will go into when we are using +* the digit that is N bits from the (least significant) end. +*/ +#define BUCKET_FOR(a, i, bits) (((a)[(i)].offset (bits)) (BUCKETS-1)) + + /* +* We need O(n) temporary storage, so we sort back and forth between +* the real array and our tmp storage. To keep them straight, we always +* sort from a into buckets in b. +*/ + struct revindex_entry *tmp = xcalloc(n, sizeof(*tmp)); + struct revindex_entry *a = entries, *b = tmp; + int bits = 0; + unsigned *pos = xmalloc(BUCKETS * sizeof(*pos)); + + while (max bits) { + struct revindex_entry *swap; + int i; + + memset(pos, 0, BUCKETS * sizeof(*pos)); + + /* +* We want pos[i] to store the index of the last element that +* will go in bucket i (actually one past the last element). +* To do this, we first count the items that will go in each +* bucket, which gives us a relative offset from the last +* bucket. We can then cumulatively add the index from the +* previous bucket to get the true index. +*/ + for (i = 0; i n; i++) + pos[BUCKET_FOR(a, i, bits)]++; + for (i = 1; i BUCKETS; i++) + pos[i] += pos[i-1]; + + /* +* Now we can drop the elements into their correct buckets (in +* our temporary array). We iterate the pos counter backwards +* to avoid using an extra index to count up. And since we are +* going backwards there, we must also go
Re: [PATCH 10/10] pack-revindex: radix-sort the revindex
On Wed, Jul 10, 2013 at 07:55:57AM -0400, Jeff King wrote: 5. We use memcpy instead of an open-coded loop to copy the whole array at the end. The individual bucket-assignment is still done by struct assignment. I haven't timed if memcpy would make a difference there. I just timed this, and I can't measure any difference. I think the struct assignment is the more readable option, and I do not think any compilers should have trouble with it. But if they do, we can switch it for a memcpy. -Peff -- To unsubscribe from this list: send the line unsubscribe git in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCH 10/10] pack-revindex: radix-sort the revindex
Jeff King wrote: That does O(n log n) offset comparisons, and profiling shows that we spend most of our time in cmp_offset. However, since we are sorting on a simple off_t, we can use numeric sorts that perform better. A radix sort can run in O(k*n), where k is the number of digits in our number. For a 64-bit off_t, using 16-bit digits gives us k=4. Wait, isn't off_t a signed data type? Did you account for that in your algorithm? On the linux.git repo, with about 3M objects to sort, this yields a 400% speedup. Here are the best-of-five numbers for running echo HEAD | git cat-file --batch-disk-size, which is dominated by time spent building the pack revindex: Okay. diff --git a/pack-revindex.c b/pack-revindex.c index 1aa9754..9365bc2 100644 --- a/pack-revindex.c +++ b/pack-revindex.c @@ -59,11 +59,85 @@ static int cmp_offset(const void *a_, const void *b_) /* revindex elements are lazily initialized */ } -static int cmp_offset(const void *a_, const void *b_) +/* + * This is a least-significant-digit radix sort. + */ Any particular reason for choosing LSD, and not MSD? +#define DIGIT_SIZE (16) +#define BUCKETS (1 DIGIT_SIZE) Okay, NUMBER_OF_BUCKETS = 2^RADIX, and you choose a hex radix. Is off_t guaranteed to be fixed-length though? I thought only the ones in stdint.h were guaranteed to be fixed-length? + /* +* We want to know the bucket that a[i] will go into when we are using +* the digit that is N bits from the (least significant) end. +*/ +#define BUCKET_FOR(a, i, bits) (((a)[(i)].offset (bits)) (BUCKETS-1)) Ouch! This is unreadable. Just write an inline function instead? A % would've been easier on the eyes, but you chose base-16. + /* +* We need O(n) temporary storage, so we sort back and forth between +* the real array and our tmp storage. To keep them straight, we always +* sort from a into buckets in b. +*/ + struct revindex_entry *tmp = xcalloc(n, sizeof(*tmp)); Shouldn't this be sizeof (struct revindex_entry), since tmp hasn't been declared yet? Also, s/n/revindex_nr/, and something more appropriate for tmp? + struct revindex_entry *a = entries, *b = tmp; It's starting to look like you have something against descriptive names ;) + int bits = 0; + unsigned *pos = xmalloc(BUCKETS * sizeof(*pos)); sizeof(unsigned int), for clarity, if not anything else. You picked malloc over calloc here, because you didn't want to incur the extra cost of zero-initializing the memory? Also, pos is the actual buckets array, I presume (hence unsigned, because there can't be a negative number of keys in any bucket)? + while (max bits) { No clue what max is. Looked at the caller and figured out that it's the pack-size, although I'm still clueless about why it's appearing here. + struct revindex_entry *swap; + int i; + + memset(pos, 0, BUCKETS * sizeof(*pos)); Ah, so that's why you used malloc there. Wait, shouldn't this be memset(pos, 0, sizeof(*pos))? + for (i = 0; i n; i++) + pos[BUCKET_FOR(a, i, bits)]++; Okay, so you know how many numbers are in each bucket. + for (i = 1; i BUCKETS; i++) + pos[i] += pos[i-1]; Cumulative sums; right. + for (i = n - 1; i = 0; i--) + b[--pos[BUCKET_FOR(a, i, bits)]] = a[i]; Classical queue. You could've gone for something more complex, but I don't think it would have been worth the extra complexity. + swap = a; + a = b; + b = swap; Wait a minute: why don't you just throw away b? You're going to rebuild the queue in the next iteration anyway, no? a is what is being sorted. + /* And bump our bits for the next round. */ + bits += DIGIT_SIZE; I'd have gone for a nice for-loop. + /* +* If we ended with our data in the original array, great. If not, +* we have to move it back from the temporary storage. +*/ + if (a != entries) + memcpy(entries, tmp, n * sizeof(*entries)); How could a be different from entries? It has no memory allocated for itself, no? Why did you even create a, and not directly operate on entries? + free(tmp); + free(pos); Overall, I found it quite confusing :( +#undef BUCKET_FOR Why not DIGIT_SIZE and BUCKETS too, while at it? -- To unsubscribe from this list: send the line unsubscribe git in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCH 10/10] pack-revindex: radix-sort the revindex
On Wed, Jul 10, 2013 at 4:55 AM, Jeff King p...@peff.net wrote: The pack revindex stores the offsets of the objects in the pack in sorted order, allowing us to easily find the on-disk size of each object. To compute it, we populate an array with the offsets from the sha1-sorted idx file, and then use qsort to order it by offsets. That does O(n log n) offset comparisons, and profiling shows that we spend most of our time in cmp_offset. However, since we are sorting on a simple off_t, we can use numeric sorts that perform better. A radix sort can run in O(k*n), where k is the number of digits in our number. For a 64-bit off_t, using 16-bit digits gives us k=4. On the linux.git repo, with about 3M objects to sort, this yields a 400% speedup. Here are the best-of-five numbers for running echo HEAD | git cat-file --batch-disk-size, which is dominated by time spent building the pack revindex: before after real0m0.834s 0m0.204s user0m0.788s 0m0.164s sys 0m0.040s 0m0.036s On a smaller repo, the radix sort would not be as impressive (and could even be worse), as we are trading the log(n) factor for the k=4 of the radix sort. However, even on git.git, with 173K objects, it shows some improvement: before after real0m0.046s 0m0.017s user0m0.036s 0m0.012s sys 0m0.008s 0m0.000s k should only be 2 for git.git. I haven't packed in a while, but I think it should all fit within 4G. :) The pathological case would be a pack file with very few very very large objects, large enough to push the pack size over the 2^48 threshold so we'd have to do all four radixes. It's probably worth mentioning here and/or in the code that k is dependent on the pack file size and that we can jump out early for small pack files. That's my favorite part of this code by the way. :) Signed-off-by: Jeff King p...@peff.net --- I changed a few things from the original, including: 1. We take an unsigned number of objects to match the fix in the last patch. 2. The 16-bit digit size is factored out to a single place, which avoids magic numbers and repeating ourselves. 3. The digits variable is renamed to bits, which is more accurate. 4. The outer loop condition uses the simpler while (max bits). 5. We use memcpy instead of an open-coded loop to copy the whole array at the end. The individual bucket-assignment is still done by struct assignment. I haven't timed if memcpy would make a difference there. 6. The 64K*sizeof(int) pos array is now heap-allocated, in case there are platforms with a small stack. I re-ran my timings to make sure none of the above impacted them; it turned out the same. pack-revindex.c | 84 + 1 file changed, 79 insertions(+), 5 deletions(-) diff --git a/pack-revindex.c b/pack-revindex.c index 1aa9754..9365bc2 100644 --- a/pack-revindex.c +++ b/pack-revindex.c @@ -59,11 +59,85 @@ static int cmp_offset(const void *a_, const void *b_) /* revindex elements are lazily initialized */ } -static int cmp_offset(const void *a_, const void *b_) +/* + * This is a least-significant-digit radix sort. + */ +static void sort_revindex(struct revindex_entry *entries, unsigned n, off_t max) { - const struct revindex_entry *a = a_; - const struct revindex_entry *b = b_; - return (a-offset b-offset) ? -1 : (a-offset b-offset) ? 1 : 0; + /* +* We use a digit size of 16 bits. That keeps our memory +* usage reasonable, and we can generally (for a 4G or smaller +* packfile) quit after two rounds of radix-sorting. +*/ +#define DIGIT_SIZE (16) +#define BUCKETS (1 DIGIT_SIZE) + /* +* We want to know the bucket that a[i] will go into when we are using +* the digit that is N bits from the (least significant) end. +*/ +#define BUCKET_FOR(a, i, bits) (((a)[(i)].offset (bits)) (BUCKETS-1)) + + /* +* We need O(n) temporary storage, so we sort back and forth between +* the real array and our tmp storage. To keep them straight, we always +* sort from a into buckets in b. +*/ + struct revindex_entry *tmp = xcalloc(n, sizeof(*tmp)); Didn't notice it the first time I read this, but do we really need calloc here? Or will malloc do? + struct revindex_entry *a = entries, *b = tmp; + int bits = 0; + unsigned *pos = xmalloc(BUCKETS * sizeof(*pos)); + + while (max bits) { + struct revindex_entry *swap; + int i; You forgot to make i unsigned. See below too... + + memset(pos, 0, BUCKETS * sizeof(*pos)); + + /* +* We want pos[i] to store the index of the last element that +* will go in bucket i (actually one past the last