From: Atsushi Kumagai <kumagai-atsushi@xxxxxxxxxxxxxxxxx> Subject: Re: [PATCH 0/8] makedumpfile: Add Snappy Compression Support Date: Thu, 5 Jul 2012 14:05:59 +0900 > On Tue, 03 Jul 2012 12:07:27 +0900 > HATAYAMA Daisuke <d.hatayama at jp.fujitsu.com> wrote: > > Thank you for your work. > > I will merge your patches into the next version with small addition: > > diff --git a/README b/README > index ae986d1..638b111 100644 > --- a/README > +++ b/README > @@ -45,6 +45,9 @@ > 6.Build with lzo support: > # make USELZO=on ; make install > The user has to prepare lzo library. > + 7.Build with snappy support: > + # make USESNAPPY=on ; make install > + The user has to prepare snappy library. > > * SUPPORTED KERNELS > This makedumpfile supports the following kernels. > Thanks. I overlooked README file... > > In addition, I did brief performance test with your patches: > - The source data is a vmcore saved on the disk, it might be sparse data. > - makedumpfile writes dumpfile to the same disk as the source data. > - execution time is average of 5 times. > > | source | zlib | LZO | snappy > ------------------------+---------------+-------------+-------------+------------- > size (byte) | 5,107,498,116 | 242,398,239 | 309,549,499 | 487,542,710 > compression ratio(%) | --- | 4.75 | 6.06 | 9.55 > execution time (sec) | --- | 143 | 49.6 | 51.2 > > > It seems that LZO still has worth in the case that good compression ratio is expected, > as you said in your benchmark report. > Yes, this is just as I'm assuming. I think it's better to use snappy for the machine with hundreads of gigabytes or more. For the machine with less than that, it's still reasonable to use LZO because risk at the worst case is small. Also, snappy is well optimized for little endian, 64-bit machines. Though I don't see this actually to be honest, this might show considerably bad performance on other archs except for x86_64. Thanks. HATAYAMA, Daisuke
