It was only speeding up compression a small amount, yet adversely affected compression and would segfault due to the size not being consistent on successive passes.
As we can work out what that compression overhead is, we can factor that into testing how much ram we can allocate.
There is no advantage to running multiple threads when there is no compression back end so drop to 1 only.
Limit ram for compression back end to 1/3 ram regardless for when OSs lie due to heavy overcommit.
Add one more thread on compression and decompression to account for the staggered nature of thread recruitment.
Make the initial buffer slightly smaller and make it progressively larger, thus recruiting threads sooner and more evenly.
This also speeds up decompression for the same reason.
Check the amount of memory being used by each thread on decompression to ensure we don't try to recruit too much ram.
Update main help screen to include environment settings.
Update to respect $TMP environment variable for TMP files.
Updated control structure to include tmpdir pointer.
Update lrzip.conf parser to respect -U -M options.
Update lrzip.conf example to include new parameters.
Reorder main Switch loop in main.c for readability.
Have MAXRAM and control.window be exclusive. MAXRAM wins.
Have UNLIMITED and control.window be exclusive. UNLIMITED wins.
Have UNLIMITED and MAXRAM be exclusive. UNLIMITED wins.
Corrects heuristic computation in rzip.c which would override
MAXRAM or UNLIMITED if control.window set
Show heuristically computed control.window when computed.
Remove display compression level from control.window verbose output.
Update print_verbose format for Testing for incompressible data in stream.c
to omit extra \n.
Changes by Peter Hyman <pete@peterhyman.com>
This is done by making the semaphore wrappers null functions on osx and closing the thread in the creation wrapper.
Move the wrappers to rzip.h to make this change clean.
Instead of failing completely, detect when a failure has occurred, and serialise work for that thread to decrease the memory required to complete compression / decompression.
Do that by waiting for the thread before it to complete before trying to work on that block again.
Check internally when lzma compress has failed and try a lower compression level before falling back to bzip2 compression.
This overlapping of compressing streams means that when files are large enough to be split into multiple blocks, all CPUs will be used more effectively throughout the compression, affording a nice speedup.
Move the writing of the chunk byte size and initial headers into the compthread to prevent any races occurring.
Fix a few dodgy callocs that may have been overflowing!
The previous commit reverts were done because the changes designed to speed it up actually slowed it down instead.
Scale the 9 levels down to the 7 that lzma has.
This makes the default lzma compression level 5 which is what lzma normally has, and uses a lot less ram and is significantly faster than previously, but at the cost of giving slightly less compression.
Change suggested maximum compression in README to disable threading with -p 1.
Use bzip2 as a fallback compression when lzma fails due to internal memory errors as may happen on 32 bits.
Limit lzma windows to 300MB in the right place on 32 bit only.
Make the main process less nice than the backend threads since it tends to be the rate limiting step.
Choose sane defaults for memory usage since linux ludicriously overcommits.
Use sliding mmap for any compression windows greater than 2/3 ram.
Consolidate and simplify testing of allocatable ram.
Minor tweaks to output.
Round up the size of the high buffer in sliding mmap to one page.
Squeeze a little more out of 32 bit compression windows.
This is done by taking each stream of data on read in into separate buffers for up to as many threads as CPUs.
As each thread's data becomes available, feed it into runzip once it is requests more of the stream.
Provided there are enough chunks in the originally compressed data, this provides a massive speedup potentially proportional to the number of CPUs. The slower the backend compression, the better the speed up (i.e. zpaq is the best sped up).
Fix the output of zpaq compress and decompress from trampling on itself and racing and consuming a lot of CPU time printing to the console.
When limiting cwindow to 6 on 32 bits, ensure that control.window is also set.
When testing for the maximum size of testmalloc, the multiple used was out by one, so increase it.
Minor output tweaks.
Place the data from each stream into a buffer that then is handed over to one thread which is allowed to begin doing the backend compression while the main rzip stream continues operating.
Fork up to as many threads as CPUs and feed data to them in a ring fashion, parallelising the workload as much as possible.
This causes a big speed up on the compression side on SMP machines.
Thread compression is limited to a minimum of 10MB compressed per thread to minimise the compromise to compression of smaller windows.
Alter the progress output to match some of the changes in verbose modes.
This facilitates a massive speed up on SMP machines proportional to the number of CPUs during the back end compression phase, but does so at some cost to the final size.
Limit the number of threads to ensure that each thread at least works on a window of STREAM_BUFSIZE.
Disable the lzma threading library as it does not contribute any more to the scalability of this new approach, yet compromises compression.
Increase the size of the windows passed to all the compression back end types now as they need more to split them up into multiple threads, and the number of blocks increases the compressed size slightly.
Prevent failure when offset is not a multiple of page size.
Add chunk percentage complete to output.
Tweak output at various verbosities.
Update documentation to reflect improved performance of unlimited mode.
Update benchmark results.
More tidying.
Modify the sliding mmap window to have a 64k smaller buffer which matches the size of the search size, and change the larger lower buffer to make it slide with the main hash search progress. This makes for a MUCH faster unlimited mode, making it actually usable.
Limit windows to 2GB again on 32 bit, but do it when determining the largest size possible in rzip.c.
Implement a linux-kernel like unlikely() wrapper for inbuilt expect, and modify most fatal warnings to be unlikely, and a few places where it's also suitable.
Minor cleanups.
accessing the buffer directly, thus allowing us to have window sizes larger than
available ram. This is implemented through the use of a "sliding mmap"
implementation. Sliding mmap uses two mmapped buffers, one large one as
previously, and one page sized smaller one. When an attempt is made to read
beyond the end of the large buffer, the small buffer is remapped to the file
area that's being accessed. While this implementation is 100x slower than direct
mmapping, it allows us to implement unlimited sized compression windows.
Implement the -U option with unlimited sized windows.
Rework the selection of compression windows. Instead of trying to guess how
much ram the machine might be able to access, we try to safely buffer as much
ram as we can, and then use that to determine the file buffer size. Do not
choose an arbitrary upper window limit unless -w is specified.
Rework the -M option to try to buffer the entire file, reducing the buffer
size until we succeed.
Align buffer sizes to page size.
Clean up lots of unneeded variables.
Fix lots of minor logic issues to do with window sizes accepted/passed to rzip
and the compression backends.
More error handling.
Change -L to affect rzip compression level directly as well as backend
compression level and use 9 by default now.
More cleanups of information output.
Use 3 point release numbering in case one minor version has many subversions.
Numerous minor cleanups and tidying.
Updated docs and manpages.
fsync after flushing buffer.
Remove unnecessary memset after anonymous mmap.
Do test malloc before compression backend to see how big a chunk can be passed.
Make stdout write directly to stdout on decompression without the need for temporary files since there is no need to seek backwards.
Make file testing not actually write the file during test.
More tidying up.
This will increase speed of compression and generate a smaller file, but not be backward compatible.
Tweak the way memory is allocated to optimise chances of success and minimise slowdown for the machine.
fsync to empty dirty data before allocating large ram to increase chance of mem allocation and decrease disk thrash of write vs read.
Add lots more information to verbose mode.
Lots of code tidying and minor tweaks.
Premalloc ram to improve early detection of being unable to allocate that much ram.
Make sure to always make chunk size a multiple of page size for mmap to work.
Begin changes to make variable byte width offsets in rzip chunks.
Decrease header entries to only 2 byte wide as per original rzip.
Random other tidying.
