Page History

...

Compression is performed by a background operation (currently called as part of Put) which maintains a pointer into the Cached List and runs toward the head compressing entries. Decompression occurs on demand during a Get. In the case of objects tagged "copy" the compressed version is reinserted in the LRU since we need to make a copy anyway. Those not tagged "copy" are inserted uncompressed in the hope that they can be reused in uncompressed form. This is a compile time option and may be something we want to change.

There are 3 algorithms and levels of compression (speed on 1 thread i7 920) :

• fastlz: 173 MB/sec compression, 442 MB/sec decompression : basically free since disk or network will limit first, ~53% final size
• libz: 55 MB/sec compression, 234 MB/sec decompression : almost free, particularly decompression, ~37% final size
• liblzma: 3 MB/sec compression, 50 MB/sec decompression : expensive, ~27% final size

These are ballpark numbers, and your millage will vary enormously. JPEG for example will not compress with any of these. The RamCache does detect compression level and will declare something "incompressible" if it doesn't get below 90% of the original size. This value is cached so that the RamCache will not attempt to compress it again (at least as long as it is in the history).