On Dec 1, 2011, at 8:21 PM, Erik Trimble wrote:
> On 12/1/2011 6:44 PM, Ragnar Sundblad wrote:
>> Thanks for your answers!
>> On 2 dec 2011, at 02:54, Erik Trimble wrote:
>>> On 12/1/2011 4:59 PM, Ragnar Sundblad wrote:
>>>> I am sorry if these are dumb questions. If there are explanations
>>>> available somewhere for those questions that I just haven't found, please
>>>> let me know! :-)
>>>> 1. It has been said that when the DDT entries, some 376 bytes or so, are
>>>> rolled out on L2ARC, there still is some 170 bytes in the ARC to reference
>>>> them (or rather the ZAP objects I believe). In some places it sounds like
>>>> those 170 bytes refers to ZAP objects that contain several DDT entries.
>>>> In other cases it sounds like for each DDT entry in the L2ARC there must
>>>> be one 170 byte reference in the ARC. What is the story here really?
>>> Yup. Each entry (not just a DDT entry, but any cached reference) in the
>>> L2ARC requires a pointer record in the ARC, so the DDT entries held in
>>> L2ARC also consume ARC space. It's a bad situation.
>> Yes, it is a bad situation. But how many DDT entries can there be in each ZAP
>> object? Some have suggested an 1:1 relationship, others have suggested that
> I'm pretty sure it's NOT 1:1, but I'd have to go look at the code. In any
> case, it's not a very big number, so you're still looking at the same O(n) as
> the number of DDT entries (n).
It is not a "bad thing" it is what it is. Almost all non-trivial caches have a
called tags in the case of CPU caches). Trivial caches do trivial manipulation
of the address
to find the data in cache, a technique that would not work well for more
management systems, like databases or file systems. So, to implement the cache,
to put the cache directory somewhere. Again, in the case of CPU caches, the
size of the tags
is not counted as the size of the cache, but can be quite substantially large.
DDT is stored in an AVL tree. It is unlikely that each ZAP object will contain
only one DDT entry.
>>>> 2. Deletion with dedup enabled is a lot heavier for some reason that I
>>>> understand. It is said that the DDT entries have to be updated for each
>>>> deleted reference to that block. Since zfs already have a mechanism for
>>>> blocks (for example with snapshots), I don't understand why the DDT has to
>>>> contain any more block references at all, or why deletion should be much
>>>> just because there are checksums (DDT entries) tied to those blocks, and
>>>> if they have to, why it would be much harder than the other block reference
>>>> mechanism. If anyone could explain this (or give me a pointer to an
>>>> explanation), I'd be very happy!
>>> Remember that, when using Dedup, each block can potentially be part of a
>>> very large number of files. So, when you delete a file, you have to go look
>>> at the DDT entry FOR EACH BLOCK IN THAT FILE, and make the appropriate DDT
>>> updates. It's essentially the same problem that erasing snapshots has -
>>> for each block you delete, you have to find and update the metadata for all
>>> the other files that share that block usage. Dedup and snapshot deletion
>>> share the same problem, it's just usually worse for dedup, since there's a
>>> much larger number of blocks that have to be updated.
>> What is it that must be updated in the DDT entries - a ref count?
>> And how does that differ from the snapshot case, which seems like
>> a very similar mechanism?
> It is similar to the snapshot case, in that the block itself has a reference
> count in it's structure (for use in both dedup and snapshots) that would get
> updated upon "delete", but you also have to consider that the DDT entry
> itself, which is a separate structure from the block structure, also has to
> be updated. This is a whole new IOPS to get that additional structure. So,
> more or less, a dedup delete has to do two operations for every one that a
> snapshot delete does. Plus,
A snapshot does not modify blocks. Each block pointer has a birth txg entry.
The txg number is
guaranteed to be monotonically incremented, so we can tell the age of a block
by its birth txg.
When you delete a snapshot, the blocks that belong to that snapshot exclusively
to the free list.
>>> The problem is that you really need to have the entire DDT in some form of
>>> high-speed random-access memory in order for things to be efficient. If you
>>> have to search the entire hard drive to get the proper DDT entry every time
>>> you delete a block, then your IOPs limits are going to get hammered hard.
>>>> 3. I, as many others, would of course like to be able to have very large
>>>> datasets deduped without having to have enormous amounts of RAM.
>>>> Since the DDT is a AVL tree, couldn't just that entire tree be cached on
>>>> for example a SSD and be searched there without necessarily having to store
>>>> anything of it in RAM? That would probably require some changes to the DDT
>>>> lookup code, and some mechanism to gather the tree to be able to lift it
>>>> over to the SSD cache, and some other stuff, but still that sounds - with
>>>> my very basic (non-)understanding of zfs - like a not to overwhelming
>>> L2ARC typically sits on an SSD, and the DDT is usually held there, if the
>>> L2ARC device exists.
>> Well, it rather seems to be ZAP objects, referenced from the ARC, which
>> happens to contain DDT entries, that is in the L2ARC.
>> I mean that you could just move the entire AVL tree onto the SSD, completely
>> outside of zfs if you will, and have it being searched there, not dependent
>> of what is in RAM at all.
>> Every DDT lookup would take up to [tree depth] number of reads, but that
>> be OK if you have a SSD which is fast on reading (which many are).
> ZFS currently treats all metadata (of which DDT entries are) and data slabs
> the same when it comes to choosing to migrate them from ARC to L2ARC, so the
> most-frequently-accessed info is in the ARC (regardless of what that info
> is), and everything else sits in the L2ARC.
The ARC has both most-frequently used and most-recently used data (hence the
Adaptive Replacement Cache) Therefore L2ARC contains data that is soon to be
from with the most-recent or most-frequent list.
> But, ALL entries in the L2ARC require an ARC reference pointer.
> Under normal operation, you really should have an L2ARC device capable of
> holding the entire DDT, to get the random IOPS benefit from that. However,
> using the current design, that still consumes a rather large amount of ARC
> space to hold the L2ARC reference pointers. A redesign effort should
> definitely reconsider how this is done - probably the most efficient way
> would be to delete L2ARC ref pointers completely in ARC, and just force a
> search of L2ARC if the data isn't found in the ARC. But, that's just a guess
> at a new implementation; I'm sure there's gotchas around that, and, like I
> said, I suspect that the only way to save dedup is to kill dedup (then redo
> it from scratch).
All deduplication implementations have a DDT. In some cases, they use a
The problem of a fixed, dedicated device is that when space runs out, they stop
It should be noted that the DDT necessarily contains critical data. On-disk
there is at least 2
copies of the DDT (and other metadata) that are spread around the pool for
implies that the worst case is a pool constructed of a single, large, slow HDD.
>>> There does need to be serious work on changing how the DDT in the L2ARC is
>>> referenced, however; the ARC memory requirements for DDT-in-L2ARC
>>> definitely need to be removed (which requires a non-trivial rearchitecting
>>> of dedup). There are some other changes that have to happen for Dedup to
>>> be really usable. Unfortunately, I can't see anyone around willing to do
>>> those changes, and my understanding of the code says that it is much more
>>> likely that we will simply remove and replace the entire dedup feature
>>> rather than trying to fix the existing design.
>> Yes, replacing it is certainly one possibility.
>> Is there any work going on for a replacement mechanism?
> Not that I know of, and there hasn't been any talk on any of these lists
> about it.
Greenbytes has an interesting implementation. Very different than stock ZFS.
Due to the critical nature of the DDT, it needs to be protected. For those who
cheap to buy one, fast L2ARC device, buying 2 fast devices to be used only for
is a tough sell.
>>>> 4. Now and then people mention that the problem with bp_rewrite has been
>>>> explained, on this very mailing list I believe, but I haven't found that
>>>> explanation. Could someone please give me a pointer to that description
>>>> (or perhaps explain it again :-) )?
>>>> Thanks for any enlightenment!
>>> bp_rewrite is a feature which stands for the (as yet unimplemented) system
>>> call of the same name, which does Block Pointer re-writing. That is, it
>>> would allow ZFS to change the physical location on media of an existing ZFS
>>> data slab. That is, bp_rewrite is necessary to allow ZFS to change the
>>> Physical layout of data on media, without changing the Conceptual
>>> arrangement of such data.
>>> It's been the #1 most-wanted feature of ZFS since I can remember, probably
>>> for 10 years now.
>> Yes, I got that much. :-)
>> But what is the problem really?
>> Being naive/ignorant (and completely ignoring any possible dependencies
>> the different layers in the zfs stack), it doesn't seem that magic or
>> when compared to the rest of the stuff in there.
> Conceptually, it's not *that* bad. From an implementation point of view,
> it's a major feature add, which touches a big chunk of the code. As always,
> the Devil is in the details. One area of problem is how to guaranty the move
> has taken place - that is, when I say I'm going to move Slab A from disk
> location X to location Y, how can I atomically guaranty this? While I'm
> doing other I/O. When there might be a power loss (or other pool loss). Plus
> lots of other non-best-case events happening....
> The major problem with "active" (vs off-line) deduplication is that no matter
> what strategy you use, you MUST keep a *complete* copy of all blocks
> currently in the pool, with their checksums. So, for something like ZFS, you
> need a structure that holds the physical block location, a 256-bit checksum,
> and a reference count, at the minimum, for each and every block in the entire
> pool. If you want good performance, this lookup table has to be on something
> that has very good random I/O performance.
As you relocate the blocks, you also have to COW the metadata for all
IMHO, this workload can be a far worse workload than DDT lookups or reference
For those who are too cheap to purchase fast disks, life will be unpleasant… it
is likely to be more
efficient to just build a new pool and migrate the data.
ZFS and performance consulting
LISA '11, Boston, MA, December 4-9
zfs-discuss mailing list