linux_dsm_epyc7002/Documentation/lzo.txt

197 lines
9.0 KiB
Plaintext
Raw Normal View History

===========================================================
LZO stream format as understood by Linux's LZO decompressor
===========================================================
Introduction
============
This is not a specification. No specification seems to be publicly available
for the LZO stream format. This document describes what input format the LZO
decompressor as implemented in the Linux kernel understands. The file subject
of this analysis is lib/lzo/lzo1x_decompress_safe.c. No analysis was made on
the compressor nor on any other implementations though it seems likely that
the format matches the standard one. The purpose of this document is to
better understand what the code does in order to propose more efficient fixes
for future bug reports.
Description
===========
The stream is composed of a series of instructions, operands, and data. The
instructions consist in a few bits representing an opcode, and bits forming
the operands for the instruction, whose size and position depend on the
opcode and on the number of literals copied by previous instruction. The
operands are used to indicate:
- a distance when copying data from the dictionary (past output buffer)
- a length (number of bytes to copy from dictionary)
- the number of literals to copy, which is retained in variable "state"
as a piece of information for next instructions.
Optionally depending on the opcode and operands, extra data may follow. These
extra data can be a complement for the operand (eg: a length or a distance
encoded on larger values), or a literal to be copied to the output buffer.
The first byte of the block follows a different encoding from other bytes, it
seems to be optimized for literal use only, since there is no dictionary yet
prior to that byte.
Lengths are always encoded on a variable size starting with a small number
of bits in the operand. If the number of bits isn't enough to represent the
length, up to 255 may be added in increments by consuming more bytes with a
rate of at most 255 per extra byte (thus the compression ratio cannot exceed
around 255:1). The variable length encoding using #bits is always the same::
length = byte & ((1 << #bits) - 1)
if (!length) {
length = ((1 << #bits) - 1)
length += 255*(number of zero bytes)
length += first-non-zero-byte
}
length += constant (generally 2 or 3)
For references to the dictionary, distances are relative to the output
pointer. Distances are encoded using very few bits belonging to certain
ranges, resulting in multiple copy instructions using different encodings.
Certain encodings involve one extra byte, others involve two extra bytes
forming a little-endian 16-bit quantity (marked LE16 below).
After any instruction except the large literal copy, 0, 1, 2 or 3 literals
are copied before starting the next instruction. The number of literals that
were copied may change the meaning and behaviour of the next instruction. In
practice, only one instruction needs to know whether 0, less than 4, or more
literals were copied. This is the information stored in the <state> variable
in this implementation. This number of immediate literals to be copied is
generally encoded in the last two bits of the instruction but may also be
taken from the last two bits of an extra operand (eg: distance).
End of stream is declared when a block copy of distance 0 is seen. Only one
instruction may encode this distance (0001HLLL), it takes one LE16 operand
for the distance, thus requiring 3 bytes.
.. important::
In the code some length checks are missing because certain instructions
are called under the assumption that a certain number of bytes follow
because it has already been guaranteed before parsing the instructions.
They just have to "refill" this credit if they consume extra bytes. This
is an implementation design choice independent on the algorithm or
encoding.
lib/lzo: implement run-length encoding Patch series "lib/lzo: run-length encoding support", v5. Following on from the previous lzo-rle patchset: https://lkml.org/lkml/2018/11/30/972 This patchset contains only the RLE patches, and should be applied on top of the non-RLE patches ( https://lkml.org/lkml/2019/2/5/366 ). Previously, some questions were raised around the RLE patches. I've done some additional benchmarking to answer these questions. In short: - RLE offers significant additional performance (data-dependent) - I didn't measure any regressions that were clearly outside the noise One concern with this patchset was around performance - specifically, measuring RLE impact separately from Matt Sealey's patches (CTZ & fast copy). I have done some additional benchmarking which I hope clarifies the benefits of each part of the patchset. Firstly, I've captured some memory via /dev/fmem from a Chromebook with many tabs open which is starting to swap, and then split this into 4178 4k pages. I've excluded the all-zero pages (as zram does), and also the no-zero pages (which won't tell us anything about RLE performance). This should give a realistic test dataset for zram. What I found was that the data is VERY bimodal: 44% of pages in this dataset contain 5% or fewer zeros, and 44% contain over 90% zeros (30% if you include the no-zero pages). This supports the idea of special-casing zeros in zram. Next, I've benchmarked four variants of lzo on these pages (on 64-bit Arm at max frequency): baseline LZO; baseline + Matt Sealey's patches (aka MS); baseline + RLE only; baseline + MS + RLE. Numbers are for weighted roundtrip throughput (the weighting reflects that zram does more compression than decompression). https://drive.google.com/file/d/1VLtLjRVxgUNuWFOxaGPwJYhl_hMQXpHe/view?usp=sharing Matt's patches help in all cases for Arm (and no effect on Intel), as expected. RLE also behaves as expected: with few zeros present, it makes no difference; above ~75%, it gives a good improvement (50 - 300 MB/s on top of the benefit from Matt's patches). Best performance is seen with both MS and RLE patches. Finally, I have benchmarked the same dataset on an x86-64 device. Here, the MS patches make no difference (as expected); RLE helps, similarly as on Arm. There were no definite regressions; allowing for observational error, 0.1% (3/4178) of cases had a regression > 1 standard deviation, of which the largest was 4.6% (1.2 standard deviations). I think this is probably within the noise. https://drive.google.com/file/d/1xCUVwmiGD0heEMx5gcVEmLBI4eLaageV/view?usp=sharing One point to note is that the graphs show RLE appears to help very slightly with no zeros present! This is because the extra code causes the clang optimiser to change code layout in a way that happens to have a significant benefit. Taking baseline LZO and adding a do-nothing line like "__builtin_prefetch(out_len);" immediately before the "goto next" has the same effect. So this is a real, but basically spurious effect - it's small enough not to upset the overall findings. This patch (of 3): When using zram, we frequently encounter long runs of zero bytes. This adds a special case which identifies runs of zeros and encodes them using run-length encoding. This is faster for both compression and decompresion. For high-entropy data which doesn't hit this case, impact is minimal. Compression ratio is within a few percent in all cases. This modifies the bitstream in a way which is backwards compatible (i.e., we can decompress old bitstreams, but old versions of lzo cannot decompress new bitstreams). Link: http://lkml.kernel.org/r/20190205155944.16007-2-dave.rodgman@arm.com Signed-off-by: Dave Rodgman <dave.rodgman@arm.com> Cc: David S. Miller <davem@davemloft.net> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Markus F.X.J. Oberhumer <markus@oberhumer.com> Cc: Matt Sealey <matt.sealey@arm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Nitin Gupta <nitingupta910@gmail.com> Cc: Richard Purdie <rpurdie@openedhand.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: Sonny Rao <sonnyrao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-08 07:30:40 +07:00
Versions
0: Original version
1: LZO-RLE
Version 1 of LZO implements an extension to encode runs of zeros using run
length encoding. This improves speed for data with many zeros, which is a
common case for zram. This modifies the bitstream in a backwards compatible way
(v1 can correctly decompress v0 compressed data, but v0 cannot read v1 data).
For maximum compatibility, both versions are available under different names
(lzo and lzo-rle). Differences in the encoding are noted in this document with
e.g.: version 1 only.
Byte sequences
==============
First byte encoding::
lib/lzo: implement run-length encoding Patch series "lib/lzo: run-length encoding support", v5. Following on from the previous lzo-rle patchset: https://lkml.org/lkml/2018/11/30/972 This patchset contains only the RLE patches, and should be applied on top of the non-RLE patches ( https://lkml.org/lkml/2019/2/5/366 ). Previously, some questions were raised around the RLE patches. I've done some additional benchmarking to answer these questions. In short: - RLE offers significant additional performance (data-dependent) - I didn't measure any regressions that were clearly outside the noise One concern with this patchset was around performance - specifically, measuring RLE impact separately from Matt Sealey's patches (CTZ & fast copy). I have done some additional benchmarking which I hope clarifies the benefits of each part of the patchset. Firstly, I've captured some memory via /dev/fmem from a Chromebook with many tabs open which is starting to swap, and then split this into 4178 4k pages. I've excluded the all-zero pages (as zram does), and also the no-zero pages (which won't tell us anything about RLE performance). This should give a realistic test dataset for zram. What I found was that the data is VERY bimodal: 44% of pages in this dataset contain 5% or fewer zeros, and 44% contain over 90% zeros (30% if you include the no-zero pages). This supports the idea of special-casing zeros in zram. Next, I've benchmarked four variants of lzo on these pages (on 64-bit Arm at max frequency): baseline LZO; baseline + Matt Sealey's patches (aka MS); baseline + RLE only; baseline + MS + RLE. Numbers are for weighted roundtrip throughput (the weighting reflects that zram does more compression than decompression). https://drive.google.com/file/d/1VLtLjRVxgUNuWFOxaGPwJYhl_hMQXpHe/view?usp=sharing Matt's patches help in all cases for Arm (and no effect on Intel), as expected. RLE also behaves as expected: with few zeros present, it makes no difference; above ~75%, it gives a good improvement (50 - 300 MB/s on top of the benefit from Matt's patches). Best performance is seen with both MS and RLE patches. Finally, I have benchmarked the same dataset on an x86-64 device. Here, the MS patches make no difference (as expected); RLE helps, similarly as on Arm. There were no definite regressions; allowing for observational error, 0.1% (3/4178) of cases had a regression > 1 standard deviation, of which the largest was 4.6% (1.2 standard deviations). I think this is probably within the noise. https://drive.google.com/file/d/1xCUVwmiGD0heEMx5gcVEmLBI4eLaageV/view?usp=sharing One point to note is that the graphs show RLE appears to help very slightly with no zeros present! This is because the extra code causes the clang optimiser to change code layout in a way that happens to have a significant benefit. Taking baseline LZO and adding a do-nothing line like "__builtin_prefetch(out_len);" immediately before the "goto next" has the same effect. So this is a real, but basically spurious effect - it's small enough not to upset the overall findings. This patch (of 3): When using zram, we frequently encounter long runs of zero bytes. This adds a special case which identifies runs of zeros and encodes them using run-length encoding. This is faster for both compression and decompresion. For high-entropy data which doesn't hit this case, impact is minimal. Compression ratio is within a few percent in all cases. This modifies the bitstream in a way which is backwards compatible (i.e., we can decompress old bitstreams, but old versions of lzo cannot decompress new bitstreams). Link: http://lkml.kernel.org/r/20190205155944.16007-2-dave.rodgman@arm.com Signed-off-by: Dave Rodgman <dave.rodgman@arm.com> Cc: David S. Miller <davem@davemloft.net> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Markus F.X.J. Oberhumer <markus@oberhumer.com> Cc: Matt Sealey <matt.sealey@arm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Nitin Gupta <nitingupta910@gmail.com> Cc: Richard Purdie <rpurdie@openedhand.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: Sonny Rao <sonnyrao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-08 07:30:40 +07:00
0..16 : follow regular instruction encoding, see below. It is worth
noting that code 16 will represent a block copy from the
dictionary which is empty, and that it will always be
invalid at this place.
lib/lzo: implement run-length encoding Patch series "lib/lzo: run-length encoding support", v5. Following on from the previous lzo-rle patchset: https://lkml.org/lkml/2018/11/30/972 This patchset contains only the RLE patches, and should be applied on top of the non-RLE patches ( https://lkml.org/lkml/2019/2/5/366 ). Previously, some questions were raised around the RLE patches. I've done some additional benchmarking to answer these questions. In short: - RLE offers significant additional performance (data-dependent) - I didn't measure any regressions that were clearly outside the noise One concern with this patchset was around performance - specifically, measuring RLE impact separately from Matt Sealey's patches (CTZ & fast copy). I have done some additional benchmarking which I hope clarifies the benefits of each part of the patchset. Firstly, I've captured some memory via /dev/fmem from a Chromebook with many tabs open which is starting to swap, and then split this into 4178 4k pages. I've excluded the all-zero pages (as zram does), and also the no-zero pages (which won't tell us anything about RLE performance). This should give a realistic test dataset for zram. What I found was that the data is VERY bimodal: 44% of pages in this dataset contain 5% or fewer zeros, and 44% contain over 90% zeros (30% if you include the no-zero pages). This supports the idea of special-casing zeros in zram. Next, I've benchmarked four variants of lzo on these pages (on 64-bit Arm at max frequency): baseline LZO; baseline + Matt Sealey's patches (aka MS); baseline + RLE only; baseline + MS + RLE. Numbers are for weighted roundtrip throughput (the weighting reflects that zram does more compression than decompression). https://drive.google.com/file/d/1VLtLjRVxgUNuWFOxaGPwJYhl_hMQXpHe/view?usp=sharing Matt's patches help in all cases for Arm (and no effect on Intel), as expected. RLE also behaves as expected: with few zeros present, it makes no difference; above ~75%, it gives a good improvement (50 - 300 MB/s on top of the benefit from Matt's patches). Best performance is seen with both MS and RLE patches. Finally, I have benchmarked the same dataset on an x86-64 device. Here, the MS patches make no difference (as expected); RLE helps, similarly as on Arm. There were no definite regressions; allowing for observational error, 0.1% (3/4178) of cases had a regression > 1 standard deviation, of which the largest was 4.6% (1.2 standard deviations). I think this is probably within the noise. https://drive.google.com/file/d/1xCUVwmiGD0heEMx5gcVEmLBI4eLaageV/view?usp=sharing One point to note is that the graphs show RLE appears to help very slightly with no zeros present! This is because the extra code causes the clang optimiser to change code layout in a way that happens to have a significant benefit. Taking baseline LZO and adding a do-nothing line like "__builtin_prefetch(out_len);" immediately before the "goto next" has the same effect. So this is a real, but basically spurious effect - it's small enough not to upset the overall findings. This patch (of 3): When using zram, we frequently encounter long runs of zero bytes. This adds a special case which identifies runs of zeros and encodes them using run-length encoding. This is faster for both compression and decompresion. For high-entropy data which doesn't hit this case, impact is minimal. Compression ratio is within a few percent in all cases. This modifies the bitstream in a way which is backwards compatible (i.e., we can decompress old bitstreams, but old versions of lzo cannot decompress new bitstreams). Link: http://lkml.kernel.org/r/20190205155944.16007-2-dave.rodgman@arm.com Signed-off-by: Dave Rodgman <dave.rodgman@arm.com> Cc: David S. Miller <davem@davemloft.net> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Markus F.X.J. Oberhumer <markus@oberhumer.com> Cc: Matt Sealey <matt.sealey@arm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Nitin Gupta <nitingupta910@gmail.com> Cc: Richard Purdie <rpurdie@openedhand.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: Sonny Rao <sonnyrao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-08 07:30:40 +07:00
17 : bitstream version. If the first byte is 17, the next byte
gives the bitstream version (version 1 only). If the first byte
is not 17, the bitstream version is 0.
lib/lzo: implement run-length encoding Patch series "lib/lzo: run-length encoding support", v5. Following on from the previous lzo-rle patchset: https://lkml.org/lkml/2018/11/30/972 This patchset contains only the RLE patches, and should be applied on top of the non-RLE patches ( https://lkml.org/lkml/2019/2/5/366 ). Previously, some questions were raised around the RLE patches. I've done some additional benchmarking to answer these questions. In short: - RLE offers significant additional performance (data-dependent) - I didn't measure any regressions that were clearly outside the noise One concern with this patchset was around performance - specifically, measuring RLE impact separately from Matt Sealey's patches (CTZ & fast copy). I have done some additional benchmarking which I hope clarifies the benefits of each part of the patchset. Firstly, I've captured some memory via /dev/fmem from a Chromebook with many tabs open which is starting to swap, and then split this into 4178 4k pages. I've excluded the all-zero pages (as zram does), and also the no-zero pages (which won't tell us anything about RLE performance). This should give a realistic test dataset for zram. What I found was that the data is VERY bimodal: 44% of pages in this dataset contain 5% or fewer zeros, and 44% contain over 90% zeros (30% if you include the no-zero pages). This supports the idea of special-casing zeros in zram. Next, I've benchmarked four variants of lzo on these pages (on 64-bit Arm at max frequency): baseline LZO; baseline + Matt Sealey's patches (aka MS); baseline + RLE only; baseline + MS + RLE. Numbers are for weighted roundtrip throughput (the weighting reflects that zram does more compression than decompression). https://drive.google.com/file/d/1VLtLjRVxgUNuWFOxaGPwJYhl_hMQXpHe/view?usp=sharing Matt's patches help in all cases for Arm (and no effect on Intel), as expected. RLE also behaves as expected: with few zeros present, it makes no difference; above ~75%, it gives a good improvement (50 - 300 MB/s on top of the benefit from Matt's patches). Best performance is seen with both MS and RLE patches. Finally, I have benchmarked the same dataset on an x86-64 device. Here, the MS patches make no difference (as expected); RLE helps, similarly as on Arm. There were no definite regressions; allowing for observational error, 0.1% (3/4178) of cases had a regression > 1 standard deviation, of which the largest was 4.6% (1.2 standard deviations). I think this is probably within the noise. https://drive.google.com/file/d/1xCUVwmiGD0heEMx5gcVEmLBI4eLaageV/view?usp=sharing One point to note is that the graphs show RLE appears to help very slightly with no zeros present! This is because the extra code causes the clang optimiser to change code layout in a way that happens to have a significant benefit. Taking baseline LZO and adding a do-nothing line like "__builtin_prefetch(out_len);" immediately before the "goto next" has the same effect. So this is a real, but basically spurious effect - it's small enough not to upset the overall findings. This patch (of 3): When using zram, we frequently encounter long runs of zero bytes. This adds a special case which identifies runs of zeros and encodes them using run-length encoding. This is faster for both compression and decompresion. For high-entropy data which doesn't hit this case, impact is minimal. Compression ratio is within a few percent in all cases. This modifies the bitstream in a way which is backwards compatible (i.e., we can decompress old bitstreams, but old versions of lzo cannot decompress new bitstreams). Link: http://lkml.kernel.org/r/20190205155944.16007-2-dave.rodgman@arm.com Signed-off-by: Dave Rodgman <dave.rodgman@arm.com> Cc: David S. Miller <davem@davemloft.net> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Markus F.X.J. Oberhumer <markus@oberhumer.com> Cc: Matt Sealey <matt.sealey@arm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Nitin Gupta <nitingupta910@gmail.com> Cc: Richard Purdie <rpurdie@openedhand.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: Sonny Rao <sonnyrao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-08 07:30:40 +07:00
18..21 : copy 0..3 literals
state = (byte - 17) = 0..3 [ copy <state> literals ]
skip byte
22..255 : copy literal string
length = (byte - 17) = 4..238
state = 4 [ don't copy extra literals ]
skip byte
Instruction encoding::
0 0 0 0 X X X X (0..15)
Depends on the number of literals copied by the last instruction.
If last instruction did not copy any literal (state == 0), this
encoding will be a copy of 4 or more literal, and must be interpreted
like this :
0 0 0 0 L L L L (0..15) : copy long literal string
length = 3 + (L ?: 15 + (zero_bytes * 255) + non_zero_byte)
state = 4 (no extra literals are copied)
If last instruction used to copy between 1 to 3 literals (encoded in
the instruction's opcode or distance), the instruction is a copy of a
2-byte block from the dictionary within a 1kB distance. It is worth
noting that this instruction provides little savings since it uses 2
bytes to encode a copy of 2 other bytes but it encodes the number of
following literals for free. It must be interpreted like this :
0 0 0 0 D D S S (0..15) : copy 2 bytes from <= 1kB distance
length = 2
state = S (copy S literals after this block)
Always followed by exactly one byte : H H H H H H H H
distance = (H << 2) + D + 1
If last instruction used to copy 4 or more literals (as detected by
state == 4), the instruction becomes a copy of a 3-byte block from the
dictionary from a 2..3kB distance, and must be interpreted like this :
0 0 0 0 D D S S (0..15) : copy 3 bytes from 2..3 kB distance
length = 3
state = S (copy S literals after this block)
Always followed by exactly one byte : H H H H H H H H
distance = (H << 2) + D + 2049
0 0 0 1 H L L L (16..31)
Copy of a block within 16..48kB distance (preferably less than 10B)
length = 2 + (L ?: 7 + (zero_bytes * 255) + non_zero_byte)
Always followed by exactly one LE16 : D D D D D D D D : D D D D D D S S
distance = 16384 + (H << 14) + D
state = S (copy S literals after this block)
End of stream is reached if distance == 16384
In version 1 only, this instruction is also used to encode a run of
zeros if distance = 0xbfff, i.e. H = 1 and the D bits are all 1.
lib/lzo: implement run-length encoding Patch series "lib/lzo: run-length encoding support", v5. Following on from the previous lzo-rle patchset: https://lkml.org/lkml/2018/11/30/972 This patchset contains only the RLE patches, and should be applied on top of the non-RLE patches ( https://lkml.org/lkml/2019/2/5/366 ). Previously, some questions were raised around the RLE patches. I've done some additional benchmarking to answer these questions. In short: - RLE offers significant additional performance (data-dependent) - I didn't measure any regressions that were clearly outside the noise One concern with this patchset was around performance - specifically, measuring RLE impact separately from Matt Sealey's patches (CTZ & fast copy). I have done some additional benchmarking which I hope clarifies the benefits of each part of the patchset. Firstly, I've captured some memory via /dev/fmem from a Chromebook with many tabs open which is starting to swap, and then split this into 4178 4k pages. I've excluded the all-zero pages (as zram does), and also the no-zero pages (which won't tell us anything about RLE performance). This should give a realistic test dataset for zram. What I found was that the data is VERY bimodal: 44% of pages in this dataset contain 5% or fewer zeros, and 44% contain over 90% zeros (30% if you include the no-zero pages). This supports the idea of special-casing zeros in zram. Next, I've benchmarked four variants of lzo on these pages (on 64-bit Arm at max frequency): baseline LZO; baseline + Matt Sealey's patches (aka MS); baseline + RLE only; baseline + MS + RLE. Numbers are for weighted roundtrip throughput (the weighting reflects that zram does more compression than decompression). https://drive.google.com/file/d/1VLtLjRVxgUNuWFOxaGPwJYhl_hMQXpHe/view?usp=sharing Matt's patches help in all cases for Arm (and no effect on Intel), as expected. RLE also behaves as expected: with few zeros present, it makes no difference; above ~75%, it gives a good improvement (50 - 300 MB/s on top of the benefit from Matt's patches). Best performance is seen with both MS and RLE patches. Finally, I have benchmarked the same dataset on an x86-64 device. Here, the MS patches make no difference (as expected); RLE helps, similarly as on Arm. There were no definite regressions; allowing for observational error, 0.1% (3/4178) of cases had a regression > 1 standard deviation, of which the largest was 4.6% (1.2 standard deviations). I think this is probably within the noise. https://drive.google.com/file/d/1xCUVwmiGD0heEMx5gcVEmLBI4eLaageV/view?usp=sharing One point to note is that the graphs show RLE appears to help very slightly with no zeros present! This is because the extra code causes the clang optimiser to change code layout in a way that happens to have a significant benefit. Taking baseline LZO and adding a do-nothing line like "__builtin_prefetch(out_len);" immediately before the "goto next" has the same effect. So this is a real, but basically spurious effect - it's small enough not to upset the overall findings. This patch (of 3): When using zram, we frequently encounter long runs of zero bytes. This adds a special case which identifies runs of zeros and encodes them using run-length encoding. This is faster for both compression and decompresion. For high-entropy data which doesn't hit this case, impact is minimal. Compression ratio is within a few percent in all cases. This modifies the bitstream in a way which is backwards compatible (i.e., we can decompress old bitstreams, but old versions of lzo cannot decompress new bitstreams). Link: http://lkml.kernel.org/r/20190205155944.16007-2-dave.rodgman@arm.com Signed-off-by: Dave Rodgman <dave.rodgman@arm.com> Cc: David S. Miller <davem@davemloft.net> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Markus F.X.J. Oberhumer <markus@oberhumer.com> Cc: Matt Sealey <matt.sealey@arm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Nitin Gupta <nitingupta910@gmail.com> Cc: Richard Purdie <rpurdie@openedhand.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: Sonny Rao <sonnyrao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-08 07:30:40 +07:00
In this case, it is followed by a fourth byte, X.
run length = ((X << 3) | (0 0 0 0 0 L L L)) + 4.
0 0 1 L L L L L (32..63)
Copy of small block within 16kB distance (preferably less than 34B)
length = 2 + (L ?: 31 + (zero_bytes * 255) + non_zero_byte)
Always followed by exactly one LE16 : D D D D D D D D : D D D D D D S S
distance = D + 1
state = S (copy S literals after this block)
0 1 L D D D S S (64..127)
Copy 3-4 bytes from block within 2kB distance
state = S (copy S literals after this block)
length = 3 + L
Always followed by exactly one byte : H H H H H H H H
distance = (H << 3) + D + 1
1 L L D D D S S (128..255)
Copy 5-8 bytes from block within 2kB distance
state = S (copy S literals after this block)
length = 5 + L
Always followed by exactly one byte : H H H H H H H H
distance = (H << 3) + D + 1
Authors
=======
This document was written by Willy Tarreau <w@1wt.eu> on 2014/07/19 during an
lib/lzo: implement run-length encoding Patch series "lib/lzo: run-length encoding support", v5. Following on from the previous lzo-rle patchset: https://lkml.org/lkml/2018/11/30/972 This patchset contains only the RLE patches, and should be applied on top of the non-RLE patches ( https://lkml.org/lkml/2019/2/5/366 ). Previously, some questions were raised around the RLE patches. I've done some additional benchmarking to answer these questions. In short: - RLE offers significant additional performance (data-dependent) - I didn't measure any regressions that were clearly outside the noise One concern with this patchset was around performance - specifically, measuring RLE impact separately from Matt Sealey's patches (CTZ & fast copy). I have done some additional benchmarking which I hope clarifies the benefits of each part of the patchset. Firstly, I've captured some memory via /dev/fmem from a Chromebook with many tabs open which is starting to swap, and then split this into 4178 4k pages. I've excluded the all-zero pages (as zram does), and also the no-zero pages (which won't tell us anything about RLE performance). This should give a realistic test dataset for zram. What I found was that the data is VERY bimodal: 44% of pages in this dataset contain 5% or fewer zeros, and 44% contain over 90% zeros (30% if you include the no-zero pages). This supports the idea of special-casing zeros in zram. Next, I've benchmarked four variants of lzo on these pages (on 64-bit Arm at max frequency): baseline LZO; baseline + Matt Sealey's patches (aka MS); baseline + RLE only; baseline + MS + RLE. Numbers are for weighted roundtrip throughput (the weighting reflects that zram does more compression than decompression). https://drive.google.com/file/d/1VLtLjRVxgUNuWFOxaGPwJYhl_hMQXpHe/view?usp=sharing Matt's patches help in all cases for Arm (and no effect on Intel), as expected. RLE also behaves as expected: with few zeros present, it makes no difference; above ~75%, it gives a good improvement (50 - 300 MB/s on top of the benefit from Matt's patches). Best performance is seen with both MS and RLE patches. Finally, I have benchmarked the same dataset on an x86-64 device. Here, the MS patches make no difference (as expected); RLE helps, similarly as on Arm. There were no definite regressions; allowing for observational error, 0.1% (3/4178) of cases had a regression > 1 standard deviation, of which the largest was 4.6% (1.2 standard deviations). I think this is probably within the noise. https://drive.google.com/file/d/1xCUVwmiGD0heEMx5gcVEmLBI4eLaageV/view?usp=sharing One point to note is that the graphs show RLE appears to help very slightly with no zeros present! This is because the extra code causes the clang optimiser to change code layout in a way that happens to have a significant benefit. Taking baseline LZO and adding a do-nothing line like "__builtin_prefetch(out_len);" immediately before the "goto next" has the same effect. So this is a real, but basically spurious effect - it's small enough not to upset the overall findings. This patch (of 3): When using zram, we frequently encounter long runs of zero bytes. This adds a special case which identifies runs of zeros and encodes them using run-length encoding. This is faster for both compression and decompresion. For high-entropy data which doesn't hit this case, impact is minimal. Compression ratio is within a few percent in all cases. This modifies the bitstream in a way which is backwards compatible (i.e., we can decompress old bitstreams, but old versions of lzo cannot decompress new bitstreams). Link: http://lkml.kernel.org/r/20190205155944.16007-2-dave.rodgman@arm.com Signed-off-by: Dave Rodgman <dave.rodgman@arm.com> Cc: David S. Miller <davem@davemloft.net> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Markus F.X.J. Oberhumer <markus@oberhumer.com> Cc: Matt Sealey <matt.sealey@arm.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Nitin Gupta <nitingupta910@gmail.com> Cc: Richard Purdie <rpurdie@openedhand.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: Sonny Rao <sonnyrao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-08 07:30:40 +07:00
analysis of the decompression code available in Linux 3.16-rc5, and updated
by Dave Rodgman <dave.rodgman@arm.com> on 2018/10/30 to introduce run-length
encoding. The code is tricky, it is possible that this document contains
mistakes or that a few corner cases were overlooked. In any case, please
report any doubt, fix, or proposed updates to the author(s) so that the
document can be updated.