[linux-mm-cc] compcache performance data

[Nitin Gupta]

Hi,

On Fri, Jun 12, 2009 at 4:04 AM, John Mitchell <johnlmitchell at gmail.com>wrote:

> - what's the "sweet spot" for which compcache is most appropriate?
>
> I've read about using this for netbooks, where CPU is relatively free, it
> has limited memory, and disk is at a premium.
>

Wherever RAM is a limiting factor and CPU is powerful enough (say, 1GHz)
should be good use case for compcache.



>
> - for this sweet spot, what's the best test load, to measure performance?
> How can this be made repeatable?
>
> The wiki mentions "load Firefox, OpenOffice, and 3 evince windows", but
> that doesn't sound very repeatable.


I think such tests can be made repeatable. There must be some tools out
there that can simulate user interaction with such applications - quick
googling brought up this result:
http://en.wikipedia.org/wiki/Linux_Desktop_Testing_Project
However, setting up this test, collecting and presenting numbers is a
project in itself :)



> The "scan" benchmark is very easy to use and is repeatable, but how
> comparable is it to real-world information?
>


'scan' benchmark is just an approximation of workload where memory access
happens to be linear and working set is greater than RAM. I cannot say if
any real application exhibits such behaviour. This benchmark is primarily
useful to stress compcache code (correctness, performance regressions etc.).
I also find it useful when profiling the code as its an easy way to trigger
swapping.


> For my tests, I ran three configurations of swap: traditional swap to disk,
> compcache only, and compcache with disk as backing store.  The plot is
> linked below, and I'm happy to publish data & scripts if that is useful.  I
> found that using compcache completed the workload quite a bit faster than
> just with disk-based swap, and compcache+backing store was faster.
>


I observed similar results in my testing: compcache+backing_swap is
generally faster than compcache alone. I could not explain this behaviour.


>
>
> ** How can I tune my system and load to best get results from compcache?
> For a VM, how much 1) RAM, 2) swap disk, and 3) how much compcache
> allocation?  What's the test load?
>


There are few things I do:
 - If compcache is almost full (reached its memlimit) and no. of hits to the
cache is low (non-increasing value of NumReads in /proc/ramzswap) then
unload and then reload compcache modules (unuse_compcache.sh;
use_compcache.sh)
 - In above, note that simple swapoff /dev/ramzswap0 will not release memory
currently allocated for compressed pages. You must do swapoff and then
unload the ramzswap module (which is what unuse script does).
 - memlimit > 15% of RAM triggers oom-killer under heavy memory pressure.

Ideally, above should be done automatically by compcache: dynamically adjust
memlimit based on hit ratio, run-time defragmentation, moving unused
compressed pages to backing swap etc. Though memory assigned to compressed
pages grows/shrinks on demand, compcache is still fairly "static" in nature.


http://www.flickr.com/photos/shavenwarthog/3618101530/
>
>
I could not understand what this graph shows.

Thanks,
Nitin
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.laptop.org/pipermail/linux-mm-cc/attachments/20090614/f35277c8/attachment.htm