Re: [PATCH 0/3] staging: zcache+ramster: move to new code base and re-merge

[Seth Jennings <sjenning@linux.vnet.ibm.com>]

On 08/16/2012 06:08 PM, Greg KH wrote:
> On a larger note, I _really_ don't want a set of 'delete and then add it
> back' set of patches.  That destroys all of the work that people had
> done up until now on the code base.
> 
> I understand your need, and want, to start fresh, but you still need to
> abide with the "evolve over time" model here.  Surely there is some path
> from the old to the new codebase that you can find?

I very much agree that this is the wrong way to do this.

I can't possibly inspect the code changes in this format, so
I'll just comment on some high level changes and mention
some performance results.

I like frontswap reclaiming memory from cleancache.  I think
that would work better than having the pages go back to the
kernel-wide page pool using the shrinker interface.

That being said, I can't test the impact of this alone
because all these changes are being submitted together.

I also like the sysfs->debugfs cleanup and zbud being moved
into its own file.

I do _not_ support replacing zsmalloc with zbud:
https://lkml.org/lkml/2012/8/14/347

I do not support the integration of ramster mixed in with
all the rest of these changes.  I have no way to see or
measure the impact of the ramster code.

I ran my kernel building benchmark twice on an unmodified
v3.5 kernel with zcache and then with these changes.  On
none-low memory pressure, <16 threads, they worked roughly
the same with low swap volume.  However, in mid-high
pressure, >20 threads, these changes degraded zcache runtime
and I/O savings by 30-80%.

I would suspect the low-density storage of zbud as the
culprit; however I can't confirm this because, again, it all
one huge change.

Some smaller issues:

1. This patchset breaks the build when CONFIG_SWAP in not
set.  FRONTSWAP depends on SWAP, but ZCACHE _selects_
FRONTSWAP.  If ZCACHE is selected and FRONTSWAP can't be
selected because SWAP isn't selected, then there is a break.

2. I get about 8 unsued/uninit'ed variable warnings at
compile time.

So I can't support this patchset, citing the performance
degradation and the fact that this submission is
unreviewable due to it being one huge monolithic patchset on
top of an existing codebase.

Seth