[OpenAFS] improving cache partition performance

Simon Wilkinson sxw@inf.ed.ac.uk
Mon, 29 Aug 2011 19:27:05 +0100

On 29 Aug 2011, at 17:33, Harald Barth wrote:
> PS: Some years ago I did some benchmarking of AFS cache performance. I
> was astonished how much slower ext2 on ramdisk was compared to
> memcache. So the question is, how fast could a AFS disk cache be if it
> did not needed to bother with a file system and could use swap?

We've made significant improvements here with 1.6.0 - the performance of =
the AFS cache is now similar to the performance of the file system which =
underlies the cache. It's still slower than memcache (which has =
significantly less overhead), but does have the advantage that you don't =
have to permanently devote a chunk of your operating system's memory to =
the disk cache.

Creating a block-device based cache is often discussed. The idea here is =
that you interface with the underlying storage at a block-based, rather =
than file-based level. You'd split your device into chunk sized records, =
and simply ship (portions of) those records to and from disk as =
required. Memory caching would be performed by the operating system =
caching AFS pages, so this would also avoid the double caching problem. =
It's a very interesting research project for anyone that has the time to =
actually implement it.=20