SNAPPY，ZLIB，LZ4，gzip几个压缩算法的对比

以下是各种压缩方式下，应用服务器的性能数据

 

场景                                响应时间         QPS       CPU         网络带宽（out）    GC数目

 

====================================================================

 

压缩：无               |          21.6ms         11500      81            118m                      0            

 

压缩：SNAPPY     |          28ms            9900        77            111m                     19

 

压缩：ZLIB           |          25ms            10000      78            112m                     27

 

压缩：LZ4            |          23ms            10900      78            110m                     0

 

压缩：gzip            |          30ms            9900        78            113m                     44

 

下面是用本地机器做的对比：

 

deflater，gzip，lz4，snappy各项对比如下：

 

压缩比：

 

source byte size:      2974090

 

snappy byte size:     499280

 

gzip byte size:          383723

 

deflater byte size:    463823

 

lz4 byte size:           170805

 

性能比：

name	Mode	Cnt	Score	Error	Units
deflater	thrpt	20	≈ 10^-4		ops/us
gzip	thrpt	20	≈ 10^-5		ops/us
lz4	thrpt	20	≈ 10^-3		ops/us
snappy	thrpt	20	≈ 10^-4		ops/us
deflater	avgt	20	30595.443 ± 3117.520		us/op
gzip	avgt	20	79238.541 ± 7208.015		us/op
lz4	avgt	20	3149.854 ± 395.372		us/op
snappy	avgt	20	11257.427 ± 1219.347		us/op
deflater	sample	431	47889.938 ± 826.705		us/op
deflater:deflater·p0.00	sample		36700.16		us/op
deflater:deflater·p0.50	sample		46989.312		us/op
deflater:deflater·p0.90	sample		53202.125		us/op
deflater:deflater·p0.95	sample		55351.706		us/op
deflater:deflater·p0.99	sample		68073.554		us/op
deflater:deflater·p0.999	sample		81788.928		us/op
deflater:deflater·p0.9999	sample		81788.928		us/op
deflater:deflater·p1.00	sample		81788.928		us/op
gzip	sample	222	93336.844 ± 1869.349		us/op
gzip:gzip·p0.00	sample		72089.6		us/op
gzip:gzip·p0.50	sample		94044.16		us/op
gzip:gzip·p0.90	sample		100545.331		us/op
gzip:gzip·p0.95	sample		107026.842		us/op
gzip:gzip·p0.99	sample		118761.718		us/op
gzip:gzip·p0.999	sample		128450.56		us/op
gzip:gzip·p0.9999	sample		128450.56		us/op
gzip:gzip·p1.00	sample		128450.56		us/op
lz4	sample	5419	3866.797 ± 54.781		us/op
lz4:lz4·p0.00	sample		2082.816		us/op
lz4:lz4·p0.50	sample		4009.984		us/op
lz4:lz4·p0.90	sample		4530.176		us/op
lz4:lz4·p0.95	sample		4980.736		us/op
lz4:lz4·p0.99	sample		6859.981		us/op
lz4:lz4·p0.999	sample		22632.202		us/op
lz4:lz4·p0.9999	sample		32178.176		us/op
lz4:lz4·p1.00	sample		32178.176		us/op
snappy	sample	1775	11473.738 ± 189.907		us/op
snappy:snappy·p0.00	sample		6299.648		us/op
snappy:snappy·p0.50	sample		11665.408		us/op
snappy:snappy·p0.90	sample		12900.762		us/op
snappy:snappy·p0.95	sample		14139.392		us/op
snappy:snappy·p0.99	sample		17627.873		us/op
snappy:snappy·p0.999	sample		56448.516		us/op
snappy:snappy·p0.9999	sample		59703.296		us/op
snappy:snappy·p1.00	sample		59703.296		us/op
deflater	ss	20	49073.148 ± 4697.407		us/op
gzip	ss	20	108874.516 ± 9298.844		us/op
lz4	ss	20	7779.001 ± 4453.597		us/op
snappy	ss	20	11424.403 ± 2107.730		us/op

注：

 

Mode标注：

 

thrpt：Throughput: operations per unit of time.                

 

（Runs by continuously calling methods, counting the total throughput over all worker threads. This mode is time-based, and it will run until the iteration time expires.）

 

avgt：Average time: average time per per operation.      

 

（Runs by continuously calling methods, counting the average time to call over all worker threads. This is the inverse of , but with different aggregation policy. This mode is time-based, and it will run until the iteration time expires.）

 

sample：Sample time: samples the time for each operation.

 

（Runs by continuously calling methods, and randomly samples the time needed for the call. This mode automatically adjusts the sampling frequency, but may omit some pauses which missed the sampling measurement. This mode is time-based, and it will run until the iteration time expires.）

 

ss：Single shot time: measures the time for a single operation.

 

（Runs by calling once and measuring its time. This mode is useful to estimate the "cold" performance when you don't want to hide the warmup invocations, or if you want to see the progress from call to call, or you want to record every single sample. This mode is work-based, and will run only for a single invocation of method.）

 

注意：lz4使用的是LZ4Factory.fastestJavaInstance().fastCompressor()，如果用fastestInstance而不是fastestJavaInstance的时候也会使用Native通过JNI实现

 

          deflater使用的是Deflater(Deflater.BEST_SPEED)模式

 

lz4在压缩比上完胜

 

lz4在首次初始化速度上完胜，

 

lz4在速度上完胜

 

lz4在稳定性上完胜

 

但是lz4的解压比较麻烦，需要指定原byte数组大小，所以开发起来工作量要大些，但是它性能好压缩比高啊。