Watching std:sort in sort sound- i can only wonder, why this second fine granular sorting pass, long after you pushed out the first sorting part out of cache?
Writing a Faster Sorting Algorithm
21–30 of 35 posts
Re: Writing a Faster Sorting Algorithm
#22Neat algorithm, but it is largely comparing apples to oranges. std::sort is universal, ska_sort is not: > Another problem is that I’m not sure what to do for data that I can’t sort. For example this algorithm can not sort a vector of std::sets > Another problem is that right now there can only be one sorting behavior per type. You have to provide me with a sort key, and if you provide me with an integer, I will sort…
Do you know of any resource which details these faster algorithms when given extra requirements?
Re: Writing a Faster Sorting Algorithm
#23Every sort algorithm should take locality of data-access into account. Without this, a comparison to other algorithms may not be fair.
Why pick this one detail to account for, and not one of the infinity of other details that may be relevant in certain circumstances?
Re: Writing a Faster Sorting Algorithm
#24Neat algorithm, but it is largely comparing apples to oranges. std::sort is universal, ska_sort is not: > Another problem is that I’m not sure what to do for data that I can’t sort. For example this algorithm can not sort a vector of std::sets > Another problem is that right now there can only be one sorting behavior per type. You have to provide me with a sort key, and if you provide me with an integer, I will sort…
With some generic programming, even complex data can be radix sorted in linear time. Google the Discriminator papers of Fritz Henglein.
Re: Writing a Faster Sorting Algorithm
#25Neat algorithm, but it is largely comparing apples to oranges. std::sort is universal, ska_sort is not: > Another problem is that I’m not sure what to do for data that I can’t sort. For example this algorithm can not sort a vector of std::sets > Another problem is that right now there can only be one sorting behavior per type. You have to provide me with a sort key, and if you provide me with an integer, I will sort…
Do you know of any resource which details these faster algorithms when given extra requirements?
Radix sort (and its variant trie sort) are the biggest ones:
https://en.wikipedia.org/wiki/Radix_sort
The O(n log n) lower bound is only for comparison based sorts. If you don't take a comparison function, but instead look at the internal structure, you can do better, but of course that depends on that internal structure and what order you want them in.
Re: Writing a Faster Sorting Algorithm
#26Neat algorithm, but it is largely comparing apples to oranges. std::sort is universal, ska_sort is not: > Another problem is that I’m not sure what to do for data that I can’t sort. For example this algorithm can not sort a vector of std::sets > Another problem is that right now there can only be one sorting behavior per type. You have to provide me with a sort key, and if you provide me with an integer, I will sort…
With some generic programming, even complex data can be radix sorted in linear time. Google the Discriminator papers of Fritz Henglein.
Re: Writing a Faster Sorting Algorithm
#27Can't test the code as there are numerous compilation errors with my version of gcc and libc++. Besides what others have written, the author seems to have taken only one measurement for each case. And more importantly there is no mention of randomization at all. If he sorted all input sizes using algorithm A and then all input sizes using algorithm B (instead of doing it in random order) it is possible that there is bias in the measurements (due to cache, branch prediction, etc), specially if he used the same input, just varying the size. I'm also worried about what he used to measure time, as he mentions that for a small number of elements it takes a few microseconds (gettimeofday(3) has a resolution of 1 microsecond, so the measured time is dangerously close to the clock resolution - clock_gettime(3) should be use instead if available, having a resolution of 1ns - or the C++ equivalent for nanosecond precision). Making the algorithm code available is not enough, it is just as important (perhaps even more) to make the benchmark code available. The "tests and benchmarks" code available on Github is a .hpp unfit for reproduction. I want to see the raw data used in the plots and the code to generate it, and I want to be able to run it myself. Otherwise there's no reason to trust it.
Re: Writing a Faster Sorting Algorithm
#28> This is a trivial function which is less complicated than the comparison operator that you would have to write for std::sort.
Unless it uses the same interface (aka call a comparator for each eval) it's apples to oranges. I also didn't see comparisons vs swaps enumerated, which matter for complex data structures or complex comparators.
Re: Writing a Faster Sorting Algorithm
#29Just to share since it is related, you can sort faster than a naive O(nlogn) algorithm whenever data is in the form of partitions where the partitions are sorted, but the contents of each partition are not. i.e. The largest element in each partition is smaller than the smallest in the next partition.
There the total number of elements are n and the average size of a partition is m. Then you need only apply any O(log(m)) algorithm to each partition to sort everything. That multiplies the time complexity by n/m. Then substitute n = m^c and the sort becomes O(n/c*log(n)) time. That gives you a factor of c speed up on what would otherwise be an O(nlogn) operation.
This trick is usable when you want to sort a B-Tree like structure where the data in each node is unsorted, but the nodes themselves are sorted. The file system hierarchy on a machine is like that. The default output of zfs list is also like that.
Re: Writing a Faster Sorting Algorithm
#30Earlier quoted context omitted.
AFAIK the STL is free to use any algorithm they want as long as it's O(n log n).
This is correct. The C++ standard specifies that sort must have a big-O complexity of n log(n). This can be seen on pdf page 925 of this working draft from 2014 [1]. [1] http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n429...
So providing a faster sort would be a violation?
The standard doesn't seem to say "or better" here, but I know that in other places it does (or least used to say something similar).