Increase the size when either the distance between actual and optimal
slot grows too large, or when too many subsequent entries would have
to be moved.
This addresses reports that the simplehash performed, sometimes
considerably, worse than dynahash.
The reason turned out to be that insertions into the hashtable where,
due to the use of parallel query, in effect done from another
hashtable, in hash-value order. If the target hashtable, due to
mis-estimation, was sized a lot smaller than the source table(s) that
lead to very imbalanced tables; a lot of entries in many close-by
buckets from the source tables were inserted into a single, wider,
bucket on the target table. As the growth factor was solely computed
based on the fillfactor, the performance of the table decreased
further and further.
b81b5a96f4 was an attempt to address this problem for hash
aggregates (but not for bitmap scans), but it turns out that the
current method of mixing hash values often actually leaves neighboring
hash-values close to each other, just in different value range. It
might be worth revisiting that independently of the performance issues
addressed in this patch..
To address that problem resize tables in two additional cases: Firstly
when the optimal position for an entry would be far from the actual
position, secondly when many entries would have to be moved to make
space for the new entry (while satisfying the robin hood property).
Due to the additional resizing threshold it seems possible, and
testing confirms that so far, that a higher fillfactor doesn't hurt
performance and saves a bit of memory. It seems better to increase it
now, before a release containing any of this code, rather than wonder
in some later release.
The various boundaries aren't determined in a particularly scientific
manner, they might need some fine-tuning.
In all my tests the new code now, even with parallelism, performs at
least as good as the old code, in several scenarios significantly
better.
Reported-By: Dilip Kumar, Robert Haas, Kuntal Ghosh
Discussion:
https://postgr.es/m/CAFiTN-vagvuAydKG9VnWcoK=ADAhxmOa4ZTrmNsViBBooTnriQ@mail.gmail.comhttps://postgr.es/m/CAGz5QC+=fNTYgzMLTBUNeKt6uaWZFXJbkB5+7oWm-n9DwVxcLA@mail.gmail.com
Even if we don't emit definitions for SH_ALLOCATE and SH_FREE, we
still need prototypes. The user can't define them before including
simplehash.h because SH_TYPE isn't available yet.
For the allocator to be able to access private_data, it needs to
become an argument to SH_CREATE. Previously we relied on callers
to set that after returning from SH_CREATE, but SH_CREATE calls
SH_ALLOCATE before returning.
Dilip Kumar, reviewed by me.
This is infrastructure for a pending patch to allow parallel bitmap
heap scans.
Dilip Kumar, reviewed (in earlier versions) by Andres Freund and
(more recently) by me. Some further renaming by me, also.
dynahash.c hash tables aren't quite fast enough for some
use-cases. There are several reasons for lacking performance:
- the use of chaining for collision handling makes them cache
inefficient, that's especially an issue when the tables get bigger.
- as the element sizes for dynahash are only determined at runtime,
offset computations are somewhat expensive
- hash and element comparisons are indirect function calls, causing
unnecessary pipeline stalls
- it's two level structure has some benefits (somewhat natural
partitioning), but increases the number of indirections
to fix several of these the hash tables have to be adjusted to the
individual use-case at compile-time. C unfortunately doesn't provide a
good way to do compile code generation (like e.g. c++'s templates for
all their weaknesses do). Thus the somewhat ugly approach taken here is
to allow for code generation using a macro-templatized header file,
which generates functions and types based on a prefix and other
parameters.
Later patches use this infrastructure to use such hash tables for
tidbitmap.c (bitmap scans) and execGrouping.c (hash aggregation,
...). In queries where these use up a large fraction of the time, this
has been measured to lead to performance improvements of over 100%.
There are other cases where this could be useful (e.g. catcache.c).
The hash table design chosen is a variant of linear open-addressing. The
biggest disadvantage of simple linear addressing schemes are highly
variable lookup times due to clustering, and deletions leaving a lot of
tombstones around. To address these issues a variant of "robin hood"
hashing is employed. Robin hood hashing optimizes chaining lengths by
moving elements close to their optimal bucket ("rich" elements), out of
the way if a to-be-inserted element is further away from its optimal
position (i.e. it's "poor"). While that can make insertions slower, the
average lookup performance is a lot better, and higher fill factors can
be used in a still performant manner. To avoid tombstones - which
normally solve the issue that a deleted node's presence is relevant to
determine whether a lookup needs to continue looking or is done -
buckets following a deleted element are shifted backwards, unless
they're empty or already at their optimal position.
There's further possible improvements that can be made to this
implementation. Amongst others:
- Use distance as a termination criteria during searches. This is
generally a good idea, but I've been able to see the overhead of
distance calculations in some cases.
- Consider combining the 'empty' status into the hashvalue, and enforce
storing the hashvalue. That could, in some cases, increase memory
density and remove a few instructions.
- Experiment further with the, very conservatively choosen, fillfactor.
- Make maximum size of hashtable configurable, to allow storing very
very large tables. That'd require 64bit hash values to be more common
than now, though.
- some smaller memcpy calls could be optimized to copy larger chunks
But since the new implementation is already considerably faster than
dynahash it seem sensible to start using it.
Reviewed-By: Tomas Vondra
Discussion: <20160727004333.r3e2k2y6fvk2ntup@alap3.anarazel.de>
Bruce Momjian
Andres Freund
Robert Haas
Heikki Linnakangas
Peter Eisentraut