On Mon, 13 Sep 2010, Levente Uzonyi wrote:
On Mon, 13 Sep 2010, Adrian Lienhard wrote:
This mail is quite long as I copied it from a conversation with Toon Verwaest. Toon and Camillo re-implemented hashed collections for Pinocchio (http://scg.unibe.ch/research/pinocchio). Their benchmarks have shown that the new implementation is significantly faster than the one in Pharo 1.1. The source code is available from www.squeaksource.com/p.html (they haven't set a license, but I'm pretty sure they'll release their code as MIT).
I post this here to find somebody knowledgeable in the area to take a look at the code and figure out whether this implementation would be interesting for Pharo...
Cheers, Adrian
citing Toon:
We ran some benchmarks on our data structures (It's in comparison with and benchmarked in standard Pharo 1.1). We just add / remove integers within a certain range thus always perfectly distributed making it ideal for Pharo dictionaries.
The numbers show how long a certain benchmark took on average in 20 runs of the benchmark and its standard deviation over those runs. A single benchmark run generally consists of many individual operations on the collection in question. For example:
benchIncludes 1 to: set size * 2 do: [ :i| set includes: i ]
where the set has size 10000. Evaluating this once results in a single run. So the number at the top of the benchmark is the sum of all averages with the average stdev.
The results:
PDictionary: 15.8 +/-1.7 Do 0.472 +/-0.079 RemoveKey 0.010 +/-0.021 AtPut 0.947 +/-0.02 AtIfAbsentPut 1.64 +/-0.96 AtPutExisting 0.198 +/-0.011 KeysAndValuesDo 0.502 +/-0.09 IncludesKey 0.365 +/-0.024 AtPutNew 4.2 +/-1.3 Includes 7.40 +/-0.33
Dictionary: 138.2 +/-4.7 Do 0.830 +/-0.098 RemoveKey 10.292 +/-0.077 AtPut 0.957 +/-0.044 AtIfAbsentPut 1.52 +/-0.95 AtPutExisting 0.203 +/-0.011 KeysAndValuesDo 0.850 +/-0.096 IncludesKey 80.25 +/-0.25 AtPutNew 3.5 +/-1.3 Includes 39.8 +/-4.4
PSet: 1.767 +/-0.057 Remove 0.008 +/-0.018 Add 0.845 +/-0.039 AddExisting 0.310 +/-0.021 AddNew 0.240 +/-0.021 Includes 0.365 +/-0.024
Set: 70.9 +/-1.2 Remove 7.737 +/-0.051 Add 0.755 +/-0.022 AddExisting 0.305 +/-0.015 AddNew 2.473 +/-0.03 Includes 59.6 +/-1.2
Obviously the very slight differences have to be taken with a grain of salt, but whenever it's more than 10x faster it's quite unlikely that it's due to some random runtime variation that might totally change in another run.
I'm a bit skeptic, because 10x improvement is pretty hard to get if the benchmark is not flawed and the code doesn't use VM support.
Okay, I found the benchmark code, and it's flawed. The distribution of the keys is not uniform. Actually it's far from uniform, it's just 1..dictSize. Anyway I ran the benchmarks in Squeak 4.2 alpha and got the following results: PBDictionary: 0.1249 +/-0.0067 RemoveKey 6.7e-500 +/-4.6e-5 AtPutNew 0.0648 +/-0.0048 Do 0.00160 +/-0.0001 AtPutExisting 0.003300000000000002 +/-0.00018 AtIfAbsentPut 0.0192 +/-0.0035 AtPut 0.01557000000000001 +/-0.00028 IncludesKey 0.0180 +/-0.0029 KeysAndValuesDo 0.00223 +/-0.0001200000000000001 Includes 0.000133 +/-6.3e-5 PBSTDictionary: 0.2204 +/-0.0063 RemoveKey 0.07203 +/-0.00032 AtPutNew 0.0434 +/-0.0055 Do 0.00 +/-0.0 AtPutExisting 0.00260 +/-0.00013 AtIfAbsentPut 0.01193 +/-0.0002 AtPut 0.0147 +/-0.0031 IncludesKey 0.01157000000000001 +/-0.0002400000000000001 KeysAndValuesDo 0.002200 +/-7.4e-5 Includes 0.05990000000000003 +/-0.00019 The PDictionary is only better at 2 benchmarks: Includes (not IncludesKey!) and RemoveKey which are both rarely used. In all other tests Squeak's Dictionary implementation is faster in this flawed benchmark. Levente
I wonder what the #pPrimitive:plugin: pragmas stand for in PDictionary's #at:ifAbsent: and #at:put:.
Where can I find the benchmark code?
An important thing for our datastructures is that they don't degrade since we don't have colliding hashes; but we didn't really test that yet (although it becomes obvious in some of the benchmark results such as removing of elements and testing presence of elements).
Is there anything special that one would neet to consider when replacing the old implementation in Pharo with yours?
What doesn't happen at the moment is shrinking dictionaries after they have grown significantly. I wouldn't know at what point to do so either; nor do I think Dictionary does that?
The current HashedCollections don't shrink. Remove is a rarely used operation.
There is a parameter indicating how many elements can be added before it switches from a SmallDictionary version to a dictionary with buckets. This is set to 20 by default, and I have no idea if it makes sense to change it; I didn't really profile what the optimal number is. Maybe it makes sense to make that a constant in the code rather than an instance variable to save a bit of space and time ...
We have our own PHashedCollection subclassing from Collection; so the biggest part of the hierarchy is compatible with what Pharo has. Mostly while porting Pinocchio-specific stuff will have to be removed; mostly annotations related to Pinocchio Primitives and exporting directives. This is all straightforward though.
The current "ratio" of hashes to elements is set to 500%. So basically when you have 32 buckets it will only grow when it contains 160 elements (or key-value pairs), so 5 elements per bucket on average. This seems to not make it slower which is understandable since SmallDictionary is pretty fast for small amounts of elements; but this ratio might have to be tweaked depending on the kind of objects and is currently a parameter. The advantage of using 500% is that you have A LOT less garbage.
Oh, and our dictionaries do use a lot less garbage, since Pharo's dictionaries use association objects for each key-value pair. We don't, so we never generate garbage when you remove an object. Our Sets on the other hand do use a bit more memory because of the bucket scheme. Sets don't use associations so we don't have an edge there :) Only performance-wise in that case.
One could implement the current dictionaries without associations, that's just less object-oriented. That would generate even less "garbage".
The hashes are masked with the amount of buckets - 1, since the buckets are always a power of 2. This implies that the lower bits of the hash are the most significant ones. This does not work together with the current identityHash / IdentityDictionary since the 12bit hash is shifted to the left by 18 (or so). This would make all the lower 18 bits 0 and all objects thus ending up in the first bucket.
That is a very important thing to note enough. However, I also think it's a bit strange to just bitshift the identityHash by 18. You don't win any significant numbers with it ... you have 12bit hashes; no way to improve that (without changing the image format). Since we overfill our dictionaries however this works quite ok with those bad hashes; and we'll only start wasting a bit of space after you have added 20480 elements to the dictionary (you have 4096 different hashes with 12 bits, * 5).
Let's imagine that we want to put 10000 objects into a IdentitySet. Each object has it's identityHash in the 0..4095 range. The hash function is very simple: hash \\ capacity + 1. Therefore the values of the hash function will all be in 1..4096, even if capacity is ~15000. This causes a lot of collisions (the table will be consist of a few very long chains degrading performance to unacceptable leves). These collisions are avoided by the shift.
Levente
_______________________________________________ Pharo-project mailing list Pharo-project@lists.gforge.inria.fr http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
_______________________________________________ Pharo-project mailing list Pharo-project@lists.gforge.inria.fr http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project