[Linux-cluster] GFS performance

[Wendy Cheng]

Kamal Jain wrote:
> Feri,
>
> Thanks for the information.  A number of people have emailed me expressing some level of interest in the outcome of this, so hopefully I will soon be able to do some tuning and performance experiments and report back our results.
>
> On the demote_secs tuning parameter, I see you're suggesting 600 seconds, which appears to be longer than the default 300 seconds as stated by Wendy Cheng at http://people.redhat.com/wcheng/Patches/GFS/readme.gfs_glock_trimming.R4 -- we're running RHEL4.5.  Wouldn't a SHORTER demote period be better for lots of files, whereas perhaps a longer demote period might be more efficient for a smaller number of files being locked for long periods of time?
>   

This demote_secs tunable is a little bit tricky :) ... What happens here 
is that, GFS caches glocks that could get accumulated to a huge amount 
of count. Unless vm releases these inodes (files) associated with these 
glocks, current GFS internal daemons will do *fruitless* scan trying to 
remove these glock (but never succeed). If you set the demote_secs to a 
large number, it will *reduce* the wake-up frequencies of these daemons 
doing these fruitless works, that, in turns, leaving more CPU cycles for 
real works. Without glock trimming patch in place, that is a way to tune 
a system that is constantly touching large amount of files (such as 
rsync). Ditto for "scand" wake-up internal, making it larger will help 
the performance in this situation.

With the *new* glock trimming patch, we actually remove the memory 
reference count so glock can be "demoted" and subsequently removed from 
the system if in idle states. To demote the glock, we need gfs_scand 
daemon to wake up often - this implies we need smaller demote_secs for 
it to be effective.
> On a related note, I converted a couple of the clusters in our lab from GULM to DLM and while performance is not necessarily noticeably improved (though more detailed testing was done after the conversion), we did notice that both clusters became more stable in the DLM configuration.
>   
This is mostly because DLM is the current default lock manager (with 
on-going development efforts) while GULM is not actively maintained.

-- Wendy