lkml.org 
[lkml]   [2015]   [Jul]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRE: [f2fs-dev] [PATCH 1/2] f2fs: refactor shrink flow for extent cache
Date
Hi Jaegeuk,

> -----Original Message-----
> From: Jaegeuk Kim [mailto:jaegeuk@kernel.org]
> Sent: Wednesday, July 01, 2015 9:26 AM
> To: Chao Yu; Chao Yu
> Cc: linux-kernel@vger.kernel.org; linux-f2fs-devel@lists.sourceforge.net;
> linux-kernel@vger.kernel.org; linux-f2fs-devel@lists.sourceforge.net
> Subject: Re: [f2fs-dev] [PATCH 1/2] f2fs: refactor shrink flow for extent cache
>
> Hi Chao,
>
> On Tue, Jun 30, 2015 at 06:42:09PM +0800, Chao Yu wrote:
> > For now, in extent cache, we have a global lru list which links all extent
> > node in the cache, and the list is protected by a global spinlock.
> >
> > If we want to shrink extent cache, we will:
> > 1. delete all target extent node from global lru list under spinlock;
> > 2. traverse all per-inode extent tree in global radix tree;
> > 2.a. traverse all extent node in per-inode extent tree, try to free extent
> > node if it is not in global lru list already.
> >
> > This method is inefficient when there is huge number of inode extent tree in
> > global extent tree.
> >
> > In this patch we introduce a new method for extent cache shrinking:
> > When we attach a new extent node, we record extent tree pointer in extent node.
> > In shrink flow, we can try to find and lock extent tree of inode directly by
> > this backward pointer, and then detach the extent node from extent tree.
> >
> > This can help to shrink extent cache more efficiently.
>
> Yes, but as we discussed before, this way will consume 4 bytes per each
> extent_node. Can it be acceptable?

Yes, this method will increase memory overhead obviously.

Maybe there is a better way to reduce lock contention and block time
of shrinker maked. I will rethink about it.

What I think now is I should firstly tests on our new shrinker with
extreme case to see how bad it shows.

>
> Instead, IMO, we need to focus on how to increase its hit ratio first.
> Actually, I wrote a patch for that.
> Could you check that first?

OK

Thanks,

>
> Thanks,



\
 
 \ /
  Last update: 2015-07-02 15:01    [W:0.045 / U:0.140 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site