Le 29/01/15 08:25, Howard Chu a écrit :
> Emmanuel Lécharny wrote:
>> Le 29/01/15 04:20, Howard Chu a écrit :
>>> ITS#8038 (syncrepl hanging onto its presentlist) only came to my
>>> attention due to the amount of memory involved. On a refresh of a DB
>>> with 2.8M entries I saw the consumer using about 320MB just for the
>>> presentlist. This list consists solely of 16 byte entryUUIDs; 2.8M
>>> items should have used no more than 48MB. An AVL node itself is 28
>>> bytes on 64-bit platform, plus 16 bytes for the struct berval wrapped
>>> around the UUID.
>>>
>>> I'm looking into adding an in-memory B+tree library to liblutil. For
>>> the type of fixed-size records we're usually storing in AVL trees, a
>>> Btree will be much more compact and higher performance since it will
>>> need rebalancing far less frequently.
>>>
>> Why using a B+tree ? A hash map wouldn't be a more appropriate data
>> structure ? EntryUUID ordering seems overkilling...
>
> I'm not fond of hashes, they're always cache-unfriendly and most of
> them have very poor dynamic growth behavior. Since we don't know in
> advance how many IDs are being stored, growth/resizing is a major
> concern. Tree structures are generally preferred because they have
> very good incremental growth performance, and B+trees have the best
> CPU cache behavior.
Hashes have three problems :
- first, as you say, growing a hash is a matter of copying the hash
completely (most of the time)
- second, they can degenerate
- third, they have an average emptiness of roughly 30%
Now, on average, with data that are well distributed, they have some
major advantages :
- they are faster than any other data structure, with a O(1) average
lookup cost
- the memory that it uses is minimal, as it's generally backed with an
array containing the data plus a flag that indicates a follow link if
the bucket is shared by more than one element
- adding and deleting elements in a hash map is generally not expensive
If you compare it with a B+tree, which is stable in O(logN), it's
faster, uses less memory, and it's easier to implement. The most
criticial point being that addition or removal from a hash is way less
expensive than for e BTree. You can also protect the hash against
concurrent access way more than a B+Tree, by splitting the buckets in
blocks of sub-buckets, with a lock being set on each separated block.
At this point, some real world experiment is needed to validate those
approaches.
