On Fri, Sep 18, 2015 at 02:19:15PM +0900, Joonsoo Kim wrote: > This patchset makes zram to use crypto API in order to support > more compression algorithm. > > The reason we need to support vairous compression algorithms is that > zram's performance is highly depend on workload and compression algorithm > and architecture. Every compression algorithm has it's own strong point. > For example, zlib is the best for compression ratio, but, it's > (de)compression speed is rather slow. Against my expectation, in my kernel > build test with zram swap, in low-memory condition on x86, zlib shows best > performance than others. In this case, I guess that compression ratio is > the most important factor. Unlike this situation, on ARM, maybe fast > (de)compression speed is the most important because it's computation speed > is slower than x86.
Fair enough. lzo and lz4 cannot cover all of workloads so surely, there are workloads other compressor algorithm can win. As well, we could support H/W compressor with crypto support so I think crypto is a way to go. One thing I have a concern is I don't want to bind some of compressor algorithm to zram statically. User can see what kinds of crypto compressor support in his system via /proc/crypto and use it dynamically. I hope zram supports it. > > Anyway, there is a concern from Sergey to use crypto API in zram. Current > crypto API has a limitation that always require tfm object to (de)compress > something because some of (de)compression function requires scratch buffer > embedded on tfm even if some of (de)compression function doesn't > require it. Due to above reason, using crypto API rather than calling Nice catch from Sergey! > compression library directly causes more memory footprint and this is > why zram doesn't use crypto API before. > > In this patchset, crypto compress noctx API is introduced to reduce memory > footprint caused by maintaining multiple tfm and zram uses it. Before > noctx API, zram's performace is down-graded if tfm is insufficient. But, > after applying noctx API, performace is restored. > > This addresses Sergey's concern perfectly and provides possibility to use > various compression algorithm in zram. > > Following is zram's read performance number. > > * iozone -t 4 -R -r 16K -s 60M -I +Z -i 0 -i 1 > * max_stream is set to 1 > * Output is in Kbytes/sec > > zram-base vs zram-crypto vs zram-crypto-noctx > > Read 10411701.88 6426911.62 9423894.38 > Re-read 10017386.62 6428218.88 11000063.50 Another patch and number we need to merge this patchset is zlib support patchset and number you had experiement. > > Thanks. > > Joonsoo Kim (7): > crypto: introduce decompression API that can be called via sharable > tfm object > crypto/lzo: support decompress_noctx > crypyo/lz4: support decompress_noctx > crypto/lz4hc: support decompress_noctx > crypto/842: support decompress_noctx > zram: use crypto API for compression > zram: use crypto decompress_noctx API > > Sergey Senozhatsky (2): > zram: make stream find and release functions static > zram: pass zstrm down to decompression path > > crypto/842.c | 9 +++- > crypto/compress.c | 36 +++++++++++++++ > crypto/crypto_null.c | 3 +- > crypto/deflate.c | 3 +- > crypto/lz4.c | 9 +++- > crypto/lz4hc.c | 9 +++- > crypto/lzo.c | 9 +++- > drivers/block/zram/Kconfig | 8 ++-- > drivers/block/zram/Makefile | 4 +- > drivers/block/zram/zcomp.c | 102 > +++++++++++++++++++++++++++++++---------- > drivers/block/zram/zcomp.h | 42 +++++++---------- > drivers/block/zram/zcomp_lz4.c | 47 ------------------- > drivers/block/zram/zcomp_lz4.h | 17 ------- > drivers/block/zram/zcomp_lzo.c | 47 ------------------- > drivers/block/zram/zcomp_lzo.h | 17 ------- > drivers/block/zram/zram_drv.c | 32 +++++++++---- > include/linux/crypto.h | 20 ++++++++ > 17 files changed, 211 insertions(+), 203 deletions(-) > delete mode 100644 drivers/block/zram/zcomp_lz4.c > delete mode 100644 drivers/block/zram/zcomp_lz4.h > delete mode 100644 drivers/block/zram/zcomp_lzo.c > delete mode 100644 drivers/block/zram/zcomp_lzo.h > > -- > 1.9.1 > -- To unsubscribe from this list: send the line "unsubscribe linux-crypto" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
