Show HN: Rapidgzip – Truly Parallel Gzip Decompression with 10 GB/s | Hacker News Reader

Show HN: Rapidgzip – Truly Parallel Gzip Decompression with 10 GB/s

github.com

35 pointsmxmlnkn12 comments

I have posted a much earlier version of this over a year ago [0].

Since then a lot has changed. Obviously, the name has changed. This happened for the paper publication [1].

I have also optimized the speed, integrated ISA-L for special cases, limited the compression-ratio-dependent maximum memory consumption, and finally added parallelized CRC32 computation, which adds ~5% overhead no matter the number of cores used. At this point, I am leaning towards calling it production-ready although there are still many ideas for improvements.

Redoing the benchmarks of the older Show HN, would look like this:

    time pigz      -d -c 4GiB-base64.gz | wc -c     # real ~13.4 s -> ~320 MB/s
    time rapidgzip -d -c 4GiB-base64.gz | wc -c     # real ~1.26 s -> ~3.4 GB/s

However, at this point, the piping itself becomes a problem. Rapidgzip is actually slightly faster than cat when comparing the piped bandwidth! E.g., compare these additional benchmarks:

    time cat 4GiB-base64.gz | wc -c                 # real ~1.06 s -> ~3.1 GB/s
    time fcat 4GiB-base64.gz | wc -c                # real ~0.41 s -> ~8.0 GB/s
    time rapidgzip -o /dev/null -d 4GiB-base64.gz   # real ~0.68 s -> ~6.5 GB/s

fcat is an alternative cat implementation that uses vmsplice to speed up piping. According to the ReadMe it currently is broken, but it works fine on my system and piping it to md5sum yields consistent results [2].

So, at this point, I/O and actually also allocations have become a limiting factor and if you want full speed, you would have to interface with the rapidgzip library interface directly (in C++ or via the Python bindings) and process the decompressed data in memory.

The project ReadMe contains further benchmarks with Silesia and FASTQ data and scaling up to 128 cores, for which rapidgzip achieves 12 GB/s for Silesia and 24 GB/s when an index has been created with --export-index and is used with --import-index.

It can also be tested with ratarmount 0.14.0, which now uses rapidgzip as a backend by default for .gz and .tar.gz files [3].

[0] https://news.ycombinator.com/item?id=32366959 [1] https://dl.acm.org/doi/10.1145/3588195.3592992 [2] https://github.com/mre/fcat [3] https://github.com/mxmlnkn/ratarmount

Open articleView on HN

Does this work with files compressed using the usual gzip command line program?

I thought you had to write "sync points" inside a gzip file to support reading from locations other than the start. If you somehow bypass that requirement, can you explain how?

Yes, it works with files produced by the usual gzip! That's why it is novel. If you make requirements on the compressor, then a multitude of competing gzip implementations exist already, the most popular of which, I think, is bgzip [0]. Rapidgzip can actually leverage the additional metadata stored in bgzip files to speed up decompression by roughly a factor of 2 compared to files compressed with the usual gzip.

Rapidgzip does not require sync points because it simply looks for them. The approach is similar to gzrecover [1], i.e., you "simply" try to start decompression at every possible >bit< position. Of course that would be too slow (~200 kB/s), so instead I use several skip and lookup tables to speed up this search to achieve roughly (~70 MB/s).

However, that search for deflate block boundaries can yield positions that look valid but which actually aren't, e.g., think of gzip-compressed gzip files. These cases also need to be detected and filtered. Rapidgzip does this filtering and the parallelization by using a cache which stores prefetched decompressed results with the start bit position as key. When the previous block has finished decoding, we know the start position of the next block, and if it exists in the cache, then that result is used. This implicitly filters out valid-looking duds, i.e., false positives.

The proof of concept of looking for block boundaries and to resolve and parallelize interdependencies between consecutive blocks, was introduced by pugz [1]. However, it had no detection for false positives and therefore did only work for compressed text files. I built upon this to make it work without assumptions on the compressed data, and I also added capabilities for random access, for which the prefetch-cache architecture also is very suited, so that I can use it in ratarmount.

[0] http://www.htslib.org/doc/bgzip.html

[1] https://github.com/arenn/gzrt

[2] https://arxiv.org/abs/1905.07224

Thank you for the thorough answer. That is all quite methodical and impressive, bringing together good ideas from disparate projects into a single solution with new benefits.

For people like me who have battled this demon before, it would help to explicitly say that you support any gzip file by way of this special magic. At least on my cursory reading of your page I did not understand this. And I never heard of gzrecover, so thanks for that too.

I like how the pugz paper starts with "Decompressing a file made by the gzip program at an arbitrary location is in principle impossible." Until today I believed the same!

Long time ago (probably around 2010 or something) I had a similar idea [1] and concluded that it is infeasible due to the LZ77 window but... wow, I never thought that starting a speculative parsing at every possible position can be good enough! Great job!

Just in case, my own idea was that, due to the fact that there is exactly one end-of-block symbol per block and the longest codes in dynamic Huffman trees should be lexicographically last due to the canonical encoding. So if you have a long row of 1 bits there is a good chance that this codes the EOB symbol (because it should be the rarest symbol and any reasonable compressor will assign the longest code) and a new block follows. I don't know if this is better than the current brute-force approach.

[1] https://news.ycombinator.com/item?id=29586865

Interesting, I never thought about looking for the EOB symbol like that but it makes sense. I would guess that some blocks you wouldn't be finding with this are, e.g., compressed zeros, because there would basically be only two Huffman codes, 0 for repeating the last symbol as much times as possible, and 1 for EOB it might even be the other way around.

The overhead of looking for deflate blocks can actually be made arbitrarily small by simply increasing the chunk size because you only have to check the first ~64 kiB of the chunk for a valid start position. 99% of compressors limit the produced deflate block size to something below 128 KiB, that's why.

Of course, lager chunk sizes have their own problems, like decreased performance for "small" files and higher memory usage because at least one chunk per core has to be kept in memory. The performance improvements on the block finder I made over the one in pugz made it possible to reduce the optimal chunk size from 16 MiB down to 4 MiB and thereby slash the memory usage by 4. 1 MiB chunks might also be fine if you can stomach the ~10% slowdown.

The LZ77 windows are something that I skipped over in my other answer. They are a problem. In order to decompress at a position without knowing the LZ77 window, the idea proposed by pugz was to use a dummy 16-bit LZ77 window initialized with unique IDs. That way, you can decompress into a 16-bit data stream and in a second step, when the window becomes known, you do a one-to-one replacement of the 16-bit unique IDs into 8-bit actual data from the real LZ77 window that has become known at this point. Decoding to 16-bit instead of 8-bit requires to modify the decompressor or write one from scratch and it also adds overhead because more memory needs to be traversed and of course because of the second replacement step that now has to be done.

> [...] there would basically be only two Huffman codes, 0 for repeating the last symbol as much times as possible, and 1 for EOB it might even be the other way around.

You are correct. More generally EOB won't be the lexicographically last symbol if there are length codes used only once in given block. But it's already possible that the Huffman tree was not optimal anyway, so such cases are safe to ignore as we can rely on the fallback.

> The overhead of looking for deflate blocks can actually be made arbitrarily small by simply increasing the chunk size because you only have to check the first ~64 kiB of the chunk for a valid start position.

But it still means that you need to probe all 2^19 positions. The fact that such probing can be made efficient surprised me the most. After reading your paper (especially the Table 1), it seems that rejecting valid trees with unused symbols were the defining factor, given that allowing them will increase false positives by at least 8 orders of magnitude.

> [...] the idea proposed by pugz was to use a dummy 16-bit LZ77 window initialized with unique IDs.

Unique IDs per block, right? I also assume there is some encoding for definitive literals (okay if only the standard DEFLATE is supported), and that second pass has to be serialized so it should make heavy use of gather instructions for performance (another reason that it didn't come to my mind at that time). Still in awe, to be honest.

> it seems that rejecting valid trees with unused symbols were the defining factor

You are correct. This assumption helps a lot to make the block finder faster. I tested with lots of compressors and I have never found one that creates "bloating Huffman codes", as I called the error code in rapidgzip. Note that, ironically, the Fixed Huffman blocks actually contain such a (predefined) Huffman code with two unused symbols. I'm not sure why it was defined in such a way, but it would allow searching Fixed Huffman blocks by looking for long bit sequences without these forbidden codes. However, Fixed Huffman blocks are rather rare, mostly for very small files (<256 B) or for the last blocks at the end of a compressed file. I currently do not look for those.

> Unique IDs per block, right?

I try to differentiate between deflate blocks (~8-128 KiB) and decompression chunks (~4 MiB). The unique IDs are per chunk, i.e., per unknown window preceding each chunk, which is at maximum 32 KiB. The encoding is something like: 0-255: fully resolved literals, 32 * 1024 to 64 * 1024 - 1: unique IDs corresponding to the index pointing in the unknown window. The second pass can also be mostly parallelized. The only serialized part is the propagation of the windows right before each chunk. The unique IDs in the last 32 KiB of each chunk can be resolved knowing only the last 32 KiB of the preceding chunk. This needs to be serial. However, the rest of the 4 MiB chunks can then be resolved in parallel and this is also done in parallel. A prior version of rapidgzip did not parallelize this and it didn't scale to a lot of cores.

Nice work! How much manual SIMD are you using?

Currently, none. ISA-L might use some, I'm not sure. I have also tried to accelerate some crucial algorithms with SIMD and my trial and errors are still inside the src/benchmarks subfolder. But, in all of those cases, a simple std::string_view::find, std::transform, or lookup tables turned out to be equally fast or faster and, of course, is more portable. Even the CRC32 algorithm uses "simple" lookup tables (slice-by-16) and avoids any SIMD while still achieving ~4.5 GB/s per core. However, there is some Intel Whitepaper [0] showing how to use PCLMULQDQ for CRC32, which might not be much faster but it would reduce cache pressure. AVX-512 even has a VPCLMULQDQ.

[0] https://www.intel.com/content/dam/develop/external/us/en/doc...

The Intel algorithm should be much faster. The slicing algorithms are limited by load unit bandwidth, which appears to be 2 loads/cycle on Zen 2, while PCLMULQDQ should be able to process 16 bytes every 2 cycles (8 bytes/cycle peak). 4.5GB/s is about 1-1.5 bytes/cycle. zlib-ng has an implementation.

That sounds more than I, for some reason, expected. I'm not sure why, but I was expecting "only" 2x speedup at maximum, while the CRC32 adds "only" 5% overhead. But, 5-8x speedup would be really nice. With that, the the CRC32 overhead would become immeasurable. I'll definitely have to look into it now.

For portable SIMD, have a look at ISPC. It allows you to write a function once, compile it for multiple instruction sets, and then automatically select the best one to use at runtime. You don't get the precision of hand-crafted SIMD, but it can grant some easy wins!

Reply on news.ycombinator.com