In the previous post, we’ve established that we want to use SIMD for speed.
We’d also like our CSV parser stream the data to avoid excessive memory usage so we’re going to have to read the CSV input in chunks.
Given that SIMD registers are currently up to 512-bits in size, the chunk size will need to be multiples of 64-bytes to work with arbitrary SIMD instructions.
This post will look at the chunk size Haskell’s bytestring library actually gives us and explore some ways we can get the required chunk size we need.
Lazy IO
Due to laziness, streaming in Haskell is straightforward. The following function lazily reads the entire contents of the input file and writes them into the output file.
import qualified Data.ByteString.Lazy as LBS
cat :: FilePath -> FilePath -> IO ()
cat inputFile outputFile = do
bs <- LBS.readFile inputFile
LBS.writeFile outputFile bsFor efficiency, the lazy bytestring is actually very similar in structure to a list of strict bytestrings. Each bytestring represents a chunk of the input file contents with a carefully chosen chunk size to maximise performance.
defaultChunkSize :: Int
defaultChunkSize = 32 * k - chunkOverhead
where k = 1024
chunkOverhead :: Int
chunkOverhead = 2 * sizeOf (undefined :: Int)Evidence of this behaviour is observable by using the toChunks function to convert the lazy bytestring and inspecting their size.
The following command reads a file with lazy IO and counts the frequency of each chunk size:
$ git clone git@github.com:haskell-works/hw-simd-cli.git
$ cd hw-simd-cli
$ git checkout 2a9dadf9291cad68f2b05917619861eead3c31dd
$ ./project.sh install
$ hw-simd chunks -i ~/7g.csv -m chunked -r classic
Total chunks: 232230
Chunk histogram:
2350,1
32752,232229Unfortunately for us, this chunk size is no good for using SIMD instructions that can use registers up to 512-bits or 64-bytes.
More interestingly, even had the defaultChunkSize been set to a more convenient size of being a multiple of 64-bytes, a 64-byte multiple chunk size is not guaranteed, as shown in this example where we read from standard input instead:
$ cat ~/7g.csv | time hw-simd chunks -i - -m chunked -r classic
Total chunks: 279332
Chunk histogram:
16,5780
32,5882
48,1793
64,739
80,400
96,239
112,130
128,77
144,55
160,50
176,32
192,20
208,16
224,12
240,5
256,7
272,9
288,8
304,8
320,2
336,5
368,3
384,2
400,1
416,1
432,3
448,1
464,1
480,1
496,4
512,1
544,3
560,3
592,1
640,1
704,2
1088,1
1184,1
1280,1
16384,8555
16400,10125
16416,23140
16432,9269
16448,4729
16464,2883
16480,1683
16496,1005
16512,669
16528,433
16544,303
16560,202
16576,160
16592,95
16608,86
16624,60
16640,53
16656,36
16672,43
16688,29
16704,38
16720,23
16736,13
16752,13
16768,20
16784,15
16800,8
16816,14
16832,9
16848,8
16864,7
16880,7
16896,2
16912,4
16928,9
16944,11
16960,10
16976,9
16992,11Rechunking & resegmenting
One strategy we could use to ensure our bytestrings are always chunked to 64-byte multiples is to rechunk the bytestrings into equal chunk sizes like the following:
|---------------a---------------|---------b----------|-c-|-------------d---------------|
|-e--|-f--|-g--|-h--|-i--|-j--|=k==|-l--|-m--|-n--|=o==|=p==|-q--|-r--|-s--|-t--|-u--|v|In the above, the chunks e-j, l-n, and q-v don’t require any byte copying because they are strict substrings of the chunks a, b and d respectively.
k, o, and p however do require copying because their bytes come from multipe source chunks.
The need for copying is denoted by using the = characters.
The above scheme may minimise the amount of byte copying, but it is still fairly expensive because many bytestring objects are created.
To reduce the number of bytestring objects, another approach is to resegment the data instead.
This process is shown below:
|---------------a---------------|---------b----------|-c-|-------------d---------------|
|--------------w--------------|=k==|------x-------|=o==|=p==|-----------y------------|v|Here, segments v, w and y are substrings of a, b and d respectively with a size that is the largest multiple of the chunk size allowed by the source segments. The segments are equivalent to the concatenation of the e-j, l-n, and q-v chunks in the rechunk example.
This gets us to the point where all except the last segment is a multiple of the chunk size.
For doing our SIMD operations, we’d like all the segments to be a multiple of the chunk size so resegmentPadded will pad the last segment to the chunk size with 0 bytes:
|---------------a---------------|---------b----------|-c-|-------------d---------------|
|--------------w--------------|=k==|------x-------|=o==|=p==|-----------y------------|=z==|This padded segment denoted by z will require byte copying from d and zero-filling the remaining buffer up to the chunk size.
For clarity, I provide the diagrams for each strategy side-by-side:
rechunk: |-e--|-f--|-g--|-h--|-i--|-j--|=k==|-l--|-m--|-n--|=o==|=p==|-q--|-r--|-s--|-t--|-u--|v|
resegment: |--------------w--------------|=k==|------x-------|=o==|=p==|-----------y------------|v|
resegmentPadded: |--------------w--------------|=k==|------x-------|=o==|=p==|-----------y------------|=z==|Some benchmarking will give us some idea of how much rechunking and resegmenting costs us:
$ git clone git@github.com:haskell-works/hw-simd-cli.git
$ cd hw-simd-cli
$ ./project.sh install
$ cat ~/7g.csv | pv -t -e -b -a | hw-simd cat -i - -o - -m default > /dev/null
7.08GiB 0:00:05 [1.27GiB/s]
$ cat ~/7g.csv | pv -t -e -b -a | hw-simd cat -i - -o - -m rechunk -c 64 > /dev/null
7.08GiB 0:00:22 [ 317MiB/s]
$ cat ~/7g.csv | pv -t -e -b -a | hw-simd cat -i - -o - -m resegment -c 64 > /dev/null
7.08GiB 0:00:06 [1.15GiB/s]The results show the cost of using small chunks is drastic compared to the much more modest overhead of resegmenting.
Pre-chunked reading
An alternative to resegmenting the lazy bytestring is to read the bytes with the desired segment size in the first place.
The hGetContents function from the bytestring library is implemented in terms of hGetContentsN like this:
hGetContentsN :: Int -> Handle -> IO ByteString
hGetContentsN k h = lazyRead -- TODO close on exceptions
where
lazyRead = unsafeInterleaveIO loop
loop = do
c <- S.hGetSome h k -- only blocks if there is no data available
if S.null c
then hClose h >> return Empty
else do cs <- lazyRead
return (Chunk c cs)hGetContentsChunkedBy, a different version of the function which guarantees that every chunk is the same size (except the last) can be implemented by using hGetBuf and createAndTrim instead of hGetSome and keeping everything else the same:
hGetContentsChunkedBy :: Int -> IO.Handle -> IO LBS.ByteString
hGetContentsChunkedBy k h = lazyRead
where lazyRead = IO.unsafeInterleaveIO loop
loop = do
c <- BS.createAndTrim k $ \p -> IO.hGetBuf h p k
if BS.null c
then IO.hClose h >> return LBS.Empty
else LBS.Chunk c <$> lazyReadBenchmarking this shows the performance is comparable to resegmenting.
$ for x in {1..5}; do cat ~/7g.csv | pv -t -e -b -a | hw-simd cat -i - -o - -m default > /dev/null; done
7.08GiB 0:00:06 [1.10GiB/s]
7.08GiB 0:00:05 [1.26GiB/s]
7.08GiB 0:00:05 [1.29GiB/s]
7.08GiB 0:00:05 [1.27GiB/s]
7.08GiB 0:00:05 [1.27GiB/s]
$ for x in {1..5}; do cat ~/7g.csv | pv -t -e -b -a | hw-simd cat -i - -o - -m prechunk -c 32704 > /dev/null; done
7.08GiB 0:00:06 [1.12GiB/s]
7.08GiB 0:00:06 [1.16GiB/s]
7.08GiB 0:00:06 [1.10GiB/s]
7.08GiB 0:00:06 [1.13GiB/s]
7.08GiB 0:00:06 [1.14GiB/s]
$ for x in {1..5}; do cat ~/7g.csv | pv -t -e -b -a | hw-simd cat -i - -o - -m resegment -c 64 > /dev/null; done
7.08GiB 0:00:06 [1.14GiB/s]
7.08GiB 0:00:06 [1.16GiB/s]
7.08GiB 0:00:06 [1.06GiB/s]
7.08GiB 0:00:05 [1.18GiB/s]
7.08GiB 0:00:05 [1.19GiB/s]This is likely because the chunks returned by hGetContents were already large enough that the extra effort to resegment the lazy bytestring is negligible.
Closing Remarks
This post looked at how we can resegment our lazy bytestring to make the chunk sizes compatible with SIMD instructions at a reasonable cost.
The next post will look at using FFI to call into C functions that use SIMD to do the heavy lifting.