> From: Anthony Liguori [mailto:anth...@codemonkey.ws]
> Sent: Tuesday, August 02, 2011 6:01 PM
> To: Alexander Graf
> Cc: Shribman, Aidan; Kevin Wolf; Stefan Hajnoczi; qemu-devel
> Developers
> Subject: Re: [Qemu-devel] [PATCH v3] XBZRLE delta for live
> migration of large memory apps
>
> On 08/02/2011 09:01 AM, Alexander Graf wrote:
> >
> >>
> >> XBZRLE has a sustained bandwidth of 1.5-2.2 GB/s for
> typical workloads making it
> >> ideal for in-line, real-time encoding such as is needed
> for live-migration.
>
> How does this compare to just doing gzip compression for the
> same workload?
>
> Regards,
>
> Anthony Liguori
For the live-migration case using a synthetic benchmark (which represents the
discussed enterprise workload) you can see that zlib (59-92 MB/s), lzo
(418-1123 MB/s) and snappy (372-1362 MB/s) compressions are much slower than
xbzrle delta compression (1778-2286 MB/s) in addition to having a much bigger
variance in results for various scenarios.
Aidan
==========================================================
Scenario SPARSE with diff segment of step 1111 len 12
==========================================================
zlib: ENC{1.08s 59 MB/s 2.36%} DEC{0.24s 269 MB/s 100.00%} .. ok
xbzlib: ENC{1.14s 56 MB/s 2.62%} DEC{0.19s 335 MB/s 100.00%} .. ok
lzo: ENC{0.06s 1123 MB/s 2.82%} DEC{0.03s 2286 MB/s 100.00%} .. ok
xblzo: ENC{0.06s 1067 MB/s 2.82%} DEC{0.05s 1391 MB/s 100.00%} .. ok
snappy: ENC{0.05s 1362 MB/s 6.14%} DEC{0.04s 1641 MB/s 100.00%} .. ok
xbsnappy: ENC{0.07s 985 MB/s 6.14%} DEC{0.06s 1103 MB/s 100.00%} .. ok
xbrle: ENC{0.29s 218 MB/s 3.08%} DEC{0.10s 627 MB/s 100.00%} .. ok
xbzrle: ENC{0.03s 2286 MB/s 3.55%} DEC{0.03s 2560 MB/s 100.00%} .. ok
==========================================================
Scenario MEDIUM with diff segment of step 701 len 33
==========================================================
zlib: ENC{0.70s 92 MB/s 5.68%} DEC{0.26s 244 MB/s 100.00%} .. ok
xbzlib: ENC{0.81s 79 MB/s 6.08%} DEC{0.21s 298 MB/s 100.00%} .. ok
lzo: ENC{0.07s 955 MB/s 7.07%} DEC{0.03s 2207 MB/s 100.00%} .. ok
xblzo: ENC{0.07s 865 MB/s 6.34%} DEC{0.05s 1362 MB/s 100.00%} .. ok
snappy: ENC{0.06s 1049 MB/s 9.80%} DEC{0.04s 1641 MB/s 100.00%} .. ok
xbsnappy: ENC{0.07s 914 MB/s 9.27%} DEC{0.06s 1164 MB/s 100.00%} .. ok
xbrle: ENC{0.28s 228 MB/s 10.31%} DEC{0.11s 561 MB/s 100.00%} .. ok
xbzrle: ENC{0.03s 2065 MB/s 8.37%} DEC{0.02s 3368 MB/s 100.00%} .. ok
==========================================================
Scenario DENSE with diff segment of step 203 len 41
==========================================================
zlib: ENC{1.54s 42 MB/s 20.76%} DEC{0.42s 151 MB/s 100.00%} .. ok
xbzlib: ENC{1.39s 46 MB/s 18.68%} DEC{0.41s 156 MB/s 100.00%} .. ok
lzo: ENC{0.11s 561 MB/s 25.58%} DEC{0.03s 2133 MB/s 100.00%} .. ok
xblzo: ENC{0.11s 561 MB/s 21.37%} DEC{0.07s 928 MB/s 100.00%} .. ok
snappy: ENC{0.12s 516 MB/s 25.87%} DEC{0.04s 1561 MB/s 100.00%} .. ok
xbsnappy: ENC{0.15s 435 MB/s 22.80%} DEC{0.07s 928 MB/s 100.00%} .. ok
xbrle: ENC{0.30s 211 MB/s 41.44%} DEC{0.11s 587 MB/s 100.00%} .. ok
xbzrle: ENC{0.04s 1778 MB/s 31.92%} DEC{0.03s 2560 MB/s 100.00%} .. ok
==========================================================
Scenario VERY-DENSE with diff segment of step 121 len 43
==========================================================
zlib: ENC{1.90s 34 MB/s 36.43%} DEC{0.49s 130 MB/s 100.00%} .. ok
xbzlib: ENC{1.60s 40 MB/s 26.71%} DEC{0.48s 132 MB/s 100.00%} .. ok
lzo: ENC{0.15s 418 MB/s 43.34%} DEC{0.03s 2133 MB/s 100.00%} .. ok
xblzo: ENC{0.17s 386 MB/s 32.29%} DEC{0.08s 800 MB/s 100.00%} .. ok
snappy: ENC{0.17s 372 MB/s 41.86%} DEC{0.04s 1488 MB/s 100.00%} .. ok
xbsnappy: ENC{0.21s 305 MB/s 33.46%} DEC{0.08s 762 MB/s 100.00%} .. ok
xbrle: ENC{0.29s 217 MB/s 72.78%} DEC{0.14s 451 MB/s 100.00%} .. ok
xbzrle: ENC{0.03s 2207 MB/s 54.92%} DEC{0.03s 2133 MB/s 100.00%} .. ok
