Re: [RFC PATCH] f2fs: add extent cache base on rb-tree

[Jaegeuk Kim]

Hi Chao,

On Wed, Dec 24, 2014 at 04:01:16PM +0800, Chao Yu wrote:
> Hi Jaegeuk,
> 
> > -----Original Message-----
> > From: Jaegeuk Kim [mailto:jaegeuk@xxxxxxxxxx]
> > Sent: Tuesday, December 23, 2014 3:36 PM
> > To: Chao Yu
> > Cc: 'Changman Lee'; linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> > Subject: Re: [RFC PATCH] f2fs: add extent cache base on rb-tree
> > 
> > Hi Chao,
> > 
> > On Tue, Dec 23, 2014 at 11:01:39AM +0800, Chao Yu wrote:
> > > Hi Jaegeuk,
> > >
> > > > -----Original Message-----
> > > > From: Jaegeuk Kim [mailto:jaegeuk@xxxxxxxxxx]
> > > > Sent: Tuesday, December 23, 2014 7:16 AM
> > > > To: Chao Yu
> > > > Cc: 'Changman Lee'; linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> > > > Subject: Re: [RFC PATCH] f2fs: add extent cache base on rb-tree
> > > >
> > > > Hi Chao,
> > > >
> > > > On Mon, Dec 22, 2014 at 03:10:30PM +0800, Chao Yu wrote:
> > > > > Hi Changman,
> > > > >
> > > > > > -----Original Message-----
> > > > > > From: Changman Lee [mailto:cm224.lee@xxxxxxxxxxx]
> > > > > > Sent: Monday, December 22, 2014 10:03 AM
> > > > > > To: Chao Yu
> > > > > > Cc: Jaegeuk Kim; linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> > > > > > Subject: Re: [RFC PATCH] f2fs: add extent cache base on rb-tree
> > > > > >
> > > > > > Hi Yu,
> > > > > >
> > > > > > Good approach.
> > > > >
> > > > > Thank you. :)
> > > > >
> > > > > > As you know, however, f2fs breaks extent itself due to COW.
> > > > >
> > > > > Yes, and sometimes f2fs use IPU when override writing, in this condition,
> > > > > by using this approach we can cache more contiguous mapping extent for better
> > > > > performance.
> > > >
> > > > Hmm. When f2fs faces with this case, there is no chance to make an extent itself
> > > > at all.
> > >
> > > With new implementation of this patch f2fs will build extent cache when readpage/readpages.
> > 
> > I don't understand your points exactly. :(
> > If there are no on-disk extents, it doesn't matter when the caches are built.
> > Could you define what scenarios you're looking at?
> 
> What I mean is that IPU will not split the exist extent in extent cache; and
> this exist extent cache was been built when we init cache with last accessed extent
> (the only on-disk extent) which was stored in inode, or be built when we invoke
> get_data_block in readpage/readpages in IPU mode. So there is a chance to make
> extent in this scenario.

As long as I can understand your point, IPU can mitigate data fragmentation,
resulting in more extents, but current IPU policies except FORCE are activated
only when partition was already fragmented a lot such as high utilization or
SSR modes. That's why I said IPU is actually not helping extent caches.

[snip]

> > > IMO, node page cache belongs to system level cache, filesystem sub system can
> > > not control it completely, cached uptodate node page will be invalidated by
> > > using drop_caches from sysfs, or reclaimer of mm, result in more IO when we need
> > > these node page next time.
> > 
> > Yes, that's exactly what I wanted.
> 
> IMO, cost is expensive when we read node page again if these pages are invalidated
> by MM. In the worst case, we will read 3 ind-node blocks + 1 dnode block + 3 NAT blocks
> from disk for one blkaddr.

My point is who should control this sort of caches.
At least, we should focus on the hit ratio, not the worst case comparison.
In the worst case, extent cache has more overhead than node page cache alone.

> 
> > 
> > > New extent cache belongs to filesystem level cache, it is completely controlled
> > > by filesystem itself. What we can profit is: on the one hand, it is used as
> > > first level cache above the node page cache, which can also increase the cache
> > > hit ratio.
> > 
> > I don't think so. The hit ratio depends on the cache policy. The node page
> > cache is managed globally by kernel in LRU manner, so I think this can show
> > affordable hit ratio.
> 
> As I test, in this scenario we can have higher ratio hit in new extent cache
> than original extent cache:
> 1.write a large file
> 2.write randomly in this file
> 3.drop cache through drop_caches entry (or reclaim by MM)
> 4.read this large file
> 
> We cache all segregated extents in inode, so our ratio hit is 100% in
> above scenario. If we add cache policy in extent cache, our hit ratio
> will drop. But added policy can provide us more choose for different hit
> ratio (may affect IO count) requirement of users.

Of course! Even though system wants to get more memory, you're trying to grab
in-house memory for extents.
Or, new shrinker reclaims some of them, but it will drop hit raio obviously.

Please see below as a summary.

> 
> > 
> > > On the other hand, it is more instable and controllable than node page
> > > cache.
> > 
> > It depends on how you can control the extent cache. But, I'm not sure that
> > would be better than page cache managed by MM.
> > 
> > So, my concerns are:
> > 
> > 1. Redundant memory overhead
> >  : The extent cache is likely on top of the node page cache, which will consume
> >  memory redundantly.
> > 
> > 2. CPU overhead
> >  : In every block address updates, it needs to traverse extent cache entries.
> 
> As I mentioned above, if extent cache has the ability of limitation and shrink,
> overhead of memory and CPU can be controlled.
> 
> > 
> > 3. Effectiveness
> >  : We have a node page cache that is managed by MM in LRU order. I think this
> >  provides good hit ratio, system-wide memory relciaming algorithms, and well-
> >  defined locking mechanism.
> 
> The effectiveness of new extent cache is the key point we should discuss.
> 
> IMO, hit ratio of extent cache and memory/CPU cost can be tradeoff.
> For example:
> 1.if limitation is on, max value is 1; shrink is off:
> 	our extent cache will be the same as original extent cache.
> 2.if limitation is off, shrink is off:
> 	our extent cache can provide high hit ratio when node page cache is
> 	invalid but there is more memory/CPU overhead.
> 3.there are more status between label 1 and label 2 as limitation and shrink is
> set to different value.
> So we just develop our extent cache for more functionality and selective.
> 
> Another point is that, now size of struct extent_info is 24 bytes which is much
> smaller than 4096 of node page size (ind-node page size is not calced), it's
> cost-efficient to use small memory to store contiguous mapping. (This is pointed
> out by Changman also)
> 
> > 
> > 4. Cache reclaiming policy
> >  a. global approach: it needs to consider lock contention, CPU overhead, and
> >                      shrinker. I don't think it is better than page cache.
> >  b. local approach: there still exists cold misses at the initial read
> >                     operations. After then, how does the extent cache increase
> > 		    hit ratio more than giving node page cache?
> > 
> > 		    For example, in the case of pretty normal scenario like
> > 		    open -> read -> close -> open -> read ..., we can't get
> > 		    benefits form locally-managed extent cache, while node
> > 		    page cache serves many block addresses.
> 
> If this case should be covered, how about remembering these recently accessed extent
> in a global list when evict, recovering extent cache from list when re-open?

Oh, surely, this example is from the local approach, not existing in
the global one. :)

I guess you've understood my concerns enoughly.
So, could you share a detailed design for all these stuffs?

For the design, I think these must be addressed seriously.
 1. decide one of global or local management
 2. should add a shrinker
 3. reduce lock contention and critical region

Please, limit this feature to consume memory as less as possible, since it will
increase memory footprint definitely.

Thanks,

> 
> Thanks,
> Yu
> 
> > 
> > This is my initial thought on the extent cache.
> > Definitely, it is worth to discuss further in more details.
> > 
> > Thanks,
> > 
> > >
> > > Thanks,
> > > Yu
> > >
> > > >
> > > > Thanks,
> > > >
> > > > >
> > > > > >
> > > > > > Anyway, mount option could be alternative for this patch.
> > > > >
> > > > > Yes, will do.
> > > > >
> > > > > Thanks,
> > > > > Yu
> > > > >
> > > > > >
> > > > > > On Fri, Dec 19, 2014 at 06:49:29PM +0800, Chao Yu wrote:
> > > > > > > Now f2fs have page-block mapping cache which can cache only one extent mapping
> > > > > > > between contiguous logical address and physical address.
> > > > > > > Normally, this design will work well because f2fs will expand coverage area of
> > > > > > > the mapping extent when we write forward sequentially. But when we write data
> > > > > > > randomly in Out-Place-Update mode, the extent will be shorten and hardly be
> > > > > > > expanded for most time as following reasons:
> > > > > > > 1.The short part of extent will be discarded if we break contiguous mapping in
> > > > > > > the middle of extent.
> > > > > > > 2.The new mapping will be added into mapping cache only at head or tail of the
> > > > > > > extent.
> > > > > > > 3.We will drop the extent cache when the extent became very fragmented.
> > > > > > > 4.We will not update the extent with mapping which we get from readpages or
> > > > > > > readpage.
> > > > > > >
> > > > > > > To solve above problems, this patch adds extent cache base on rb-tree like other
> > > > > > > filesystems (e.g.: ext4/btrfs) in f2fs. By this way, f2fs can support another
> > > > > > > more effective cache between dnode page cache and disk. It will supply high hit
> > > > > > > ratio in the cache with fewer memory when dnode page cache are reclaimed in
> > > > > > > environment of low memory.
> > > > > > >
> > > > > > > Todo:
> > > > > > > *introduce mount option for extent cache.
> > > > > > > *add shrink ability for extent cache.
> > > > > > >
> > > > > > > Signed-off-by: Chao Yu <chao2.yu@xxxxxxxxxxx>
> > > > > > > ---
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/