2017-05-15 07:02:59 +07:00
|
|
|
===========================================================
|
2014-09-27 17:31:35 +07:00
|
|
|
LZO stream format as understood by Linux's LZO decompressor
|
|
|
|
===========================================================
|
|
|
|
|
|
|
|
Introduction
|
2017-05-15 07:02:59 +07:00
|
|
|
============
|
2014-09-27 17:31:35 +07:00
|
|
|
|
|
|
|
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
|
2017-05-15 07:02:59 +07:00
|
|
|
===========
|
2014-09-27 17:31:35 +07:00
|
|
|
|
|
|
|
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
|
2017-05-15 07:02:59 +07:00
|
|
|
operands are used to indicate:
|
2014-09-27 17:31:35 +07:00
|
|
|
|
|
|
|
- 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
|
2017-05-15 07:02:59 +07:00
|
|
|
around 255:1). The variable length encoding using #bits is always the same::
|
2014-09-27 17:31:35 +07:00
|
|
|
|
|
|
|
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.
|
|
|
|
|
2017-05-15 07:02:59 +07:00
|
|
|
.. 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.
|
2014-09-27 17:31:35 +07:00
|
|
|
|
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).
|
|
|
|
|
2019-03-08 07:30:44 +07:00
|
|
|
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.
|
|
|
|
|
2014-09-27 17:31:35 +07:00
|
|
|
Byte sequences
|
2017-05-15 07:02:59 +07:00
|
|
|
==============
|
2014-09-27 17:31:35 +07:00
|
|
|
|
2017-05-15 07:02:59 +07:00
|
|
|
First byte encoding::
|
2014-09-27 17:31:35 +07:00
|
|
|
|
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
|
2014-09-27 17:31:35 +07:00
|
|
|
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
|
2019-03-08 07:30:44 +07:00
|
|
|
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
|
|
|
|
2014-09-27 17:31:35 +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
|
|
|
|
|
2017-05-15 07:02:59 +07:00
|
|
|
Instruction encoding::
|
2014-09-27 17:31:35 +07:00
|
|
|
|
|
|
|
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
|
|
|
|
|
2019-03-08 07:30:44 +07:00
|
|
|
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.
|
|
|
|
|
2014-09-27 17:31:35 +07:00
|
|
|
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
|
2017-05-15 07:02:59 +07:00
|
|
|
=======
|
2014-09-27 17:31:35 +07:00
|
|
|
|
|
|
|
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.
|