[Pharo-project] One order of magnitude faster sets/dicts?!
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. 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? 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. 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).
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. 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
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
Hi,
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 ... 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. I agree this benchmark is severely flawed. But it's generally also the ideal case for Pharo/Squeak dictionaries. Whatever might happen in other cases to degenerate the dictionary doesn't happen here since keys are added in the right order (increasing hashes). So whenever place should be made for a key with an already existing hash in Dictionary, this is avoided. In PDictionary however this scenario wouldn't even happen. Now the problem is that I don't have experience in writing proper benchmarks and this thread could provide us with some decent benchmarks that better show standard use of dictionaries (and as I hope, show that PDictionaries are more resistant to degradation than the current Dictionaries.
What we should take away from the existing benchmarks is that the differences aren't even that high for the ideal case of Pharo/Squeak dictionaries. The main point in favor of the PSet / PDictionary implementations is that you always know the exact ranges of keys/values that belong to a certain hash. So for includesKey: and at:ifAbsent: this is ideal since you never look further than the current hash-value. In Squeak/Pharo dictionary/set you don't know this, thanks to hash-collision/propagation, so you have to scan until you find an association that is nil. If you are unlucky and your whole dictionary has all the associations grouped together, even though they all represent keys with different hashes, you'll have to look through a whole bunch of totally unrelated keys. For example, I quickly changed one of the benchmarks to be a bit less optimal for the Squeak/Pharo collections, namely the includesKey: benchmark (which is also representative for at:) Rather than doing 1 to: dict size * 2 do: [ :i| dict at: (self key: i) ifAbsentPut: (self value: i)]. I changed it to: 1 to: dict size * 10 by: 10 do: [ :i| dict at: (self key: i) ifAbsentPut: (self value: i)]. This forces the dictionary to miss a lot more often. Since all the keys are grouped together this is the worst case for Squeak/Pharo dictionaries if you want to look something up that's not there. This are the results on Squeak 4.1: PBDictionary 0.0261 +/-0.0017 PBSTDictionary 5.235 +/-0.01 So around 200x faster in case of PDictionary. While PDictionary stays stable (even slightly faster, go figure), Squeak dictionaries degrade immensely! On Pharo the results are a bit better, but still very much degraded: PBDictionary 0.02387 +/-0.00014 PBSTDictionary 1.6577 +/-0.0036 So around 60x faster in case of PDictionary. If on Set I ensure that all elements will be found, I get the following result for #includes: PBSetIncludes 0.00440 +/-0.0001100000000000001 PBSTSetIncludes 0.004000 +/-9.6e-5 making PSet marginally slower. However, once you start not finding elements, with the same change as before, (to: size*10 by: 10): PBSetIncludes 0.00443 +/-0.0001 PBSTSetIncludes 0.9844000000000005 +/-0.0017 So around 220x faster in case of PSet. While PSet and PDictionary never really suffer from collisions, normal Squeak/Pharo dictionaries suffer immensely and grow up to 200x slower. Btw, this last benchmark was run on Pharo. Anyway, you are definitely right that the benchmarks suck and that we need better ones. As for PDictionaries and PSets, nowing that you know exactly what the range of existing hashes is I think it's only logical that it will be faster in many of the operations that need this knowledge, such as #includesKey:, #remove:, #at:put:/#add: and #at:/#includes:. Since PDictionary switches between SmallDictionary style and normal dictionary style, it also seems logical that it behaves faster in case that you are in SmallDictionary style without having to care about it. This is also not shown in any benchmark, but it is the case and makes it faster for that case. However, this added test makes it slower for the general dictionary style so removing it might improve things in that case to regain the marginal difference between both dictionaries in the ideal case for Pharo/Squeak. There is anyway no reason why PDictionary would be slower than Pharo/Squeak dictionaries except for design decisions like the previously mentioned one, and those can be changed if wanted. There is however a good reason why Squeak/Pharo dictionaries are slower and degrade faster, namely colliding hashes that even propagate collisions (also not benchmarked at the moment!). Ok, on the whole I wasn't really planning on pushing these dictionaries too much on Pharo / Squeak. We use them in Pinocchio and are happy with them (especially since we have native support). But they are there and can be used if you want to. I can understand that people are skeptical, but since I don't really have knowledge on building benchmarks (I didn't even set up the current ones, it was a student of mine) I don't think I can help much there. We use them throughout our project and some parts (such as our parser) have gotten a lot faster since we started using them. Btw, the code would indeed just be released under MIT license which is maybe already automatically the case since I signed that agreement as a contributor to Pharo? Some answers / comments to the other random questions and remarks:
I wonder what the #pPrimitive:plugin: pragmas stand for in PDictionary's #at:ifAbsent: and #at:put:. Those methods are implemented as primitives in the Pinocchio VM that we are developing. In primitive form most of the methods run ~ twice as fast. This is however not compatible with the way the stack is handled in Squeak/Pharo so can't really be ported. The current HashedCollections don't shrink. Remove is a rarely used operation. Ok, makes sense. One could implement the current dictionaries without associations, that's just less object-oriented. That would generate even less "garbage". True enough. 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. So you change the hash implementation to ensure that they take up the whole array of the Dictionary rather than just the first 4096 + collisions positions. Makes perfect sense for the squeak/pharo implementation but rather strange to enforce this on the whole system imho.
Ok, I hope this clears up some things. cheers, Toon
To be fair, the text below just shows that Dictionary / Set in Pharo degrade from O(1) to O(n) as it has to search through all the elements in the Set/Dictionary. It's not that those high factors are constants because of better implementation. Pinocchio dictionaries just stay O(1). The only way to degrade Pinocchio dictionaries is by giving it bad hashes; but that's something you can't solve in all cases anyway. cheers, Toon
This forces the dictionary to miss a lot more often. Since all the keys are grouped together this is the worst case for Squeak/Pharo dictionaries if you want to look something up that's not there. This are the results on Squeak 4.1:
PBDictionary 0.0261 +/-0.0017 PBSTDictionary 5.235 +/-0.01
So around 200x faster in case of PDictionary.
While PDictionary stays stable (even slightly faster, go figure), Squeak dictionaries degrade immensely! On Pharo the results are a bit better, but still very much degraded:
PBDictionary 0.02387 +/-0.00014 PBSTDictionary 1.6577 +/-0.0036
So around 60x faster in case of PDictionary.
If on Set I ensure that all elements will be found, I get the following result for #includes:
PBSetIncludes 0.00440 +/-0.0001100000000000001 PBSTSetIncludes 0.004000 +/-9.6e-5
making PSet marginally slower. However, once you start not finding elements, with the same change as before, (to: size*10 by: 10):
PBSetIncludes 0.00443 +/-0.0001 PBSTSetIncludes 0.9844000000000005 +/-0.0017
So around 220x faster in case of PSet. While PSet and PDictionary never really suffer from collisions, normal Squeak/Pharo dictionaries suffer immensely and grow up to 200x slower. Btw, this last benchmark was run on Pharo.
On Tue, 14 Sep 2010, Toon Verwaest wrote:
Hi,
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 ... 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. I agree this benchmark is severely flawed. But it's generally also the ideal case for Pharo/Squeak dictionaries. Whatever might happen in other cases to
In one sense, it's ideal. #at:put is fast, because the first check slot will be an empty slot. But #removeKey:ifAbsent: and #includes: will be very slow because those will have to check all subsequent slots. Normally a hash table which uses open addressing with linear probing has short chains. The average chain length is below a constant which only depends on the load factor, this is how it provides O(1) time for the operations. In Squeak the default hash function is very simple and therefore fast. But it has a caveat. If you provide low quality input, like integers from a small range (1..10000), it will have bad performance. With PluggableSet/PluggableDictionary you can change the hash function if you need it. There is a class method which creates a new dictionary that's more tolerant to the small integer range and not uniform distribution issue: PluggableSet integerSet.
degenerate the dictionary doesn't happen here since keys are added in the right order (increasing hashes). So whenever place should be made for a key
So in this case the dictionary is far from normal, since it has only a few chains which are very long.
with an already existing hash in Dictionary, this is avoided. In PDictionary however this scenario wouldn't even happen. Now the problem is that I don't have experience in writing proper benchmarks and this thread could provide us with some decent benchmarks that better show standard use of dictionaries (and as I hope, show that PDictionaries are more resistant to degradation than the current Dictionaries.
Here is the benchmark I used while I was changing Squeak's Dictionary implementation: http://leves.web.elte.hu/collections/DictionaryBenchmark.st I'm sure it could be improved and simplified a bit with the new collection methods.
What we should take away from the existing benchmarks is that the differences aren't even that high for the ideal case of Pharo/Squeak dictionaries.
The main point in favor of the PSet / PDictionary implementations is that you always know the exact ranges of keys/values that belong to a certain hash. So for includesKey: and at:ifAbsent: this is ideal since you never look further than the current hash-value. In Squeak/Pharo dictionary/set you don't know this, thanks to hash-collision/propagation, so you have to scan until you find an association that is nil. If you are unlucky and your whole dictionary has all the associations grouped together, even though they all represent keys with different hashes, you'll have to look through a whole bunch of totally unrelated keys.
It's the classical separate chaining vs open addressing problem. IMHO open addressing is better in general: it uses less space, has less cache misses and has better performance with a proper hash function. Though there are special cases when it's inferior to separate chaining.
For example, I quickly changed one of the benchmarks to be a bit less optimal for the Squeak/Pharo collections, namely the includesKey: benchmark (which is also representative for at:) Rather than doing
1 to: dict size * 2 do: [ :i| dict at: (self key: i) ifAbsentPut: (self value: i)].
I changed it to:
1 to: dict size * 10 by: 10 do: [ :i| dict at: (self key: i) ifAbsentPut: (self value: i)].
This forces the dictionary to miss a lot more often. Since all the keys are grouped together this is the worst case for Squeak/Pharo dictionaries if you want to look something up that's not there. This are the results on Squeak 4.1:
PBDictionary 0.0261 +/-0.0017 PBSTDictionary 5.235 +/-0.01
So around 200x faster in case of PDictionary.
The values of the hash function are still in 1..100000 and the distribution of the values is very far from uniform in that small range.
While PDictionary stays stable (even slightly faster, go figure), Squeak dictionaries degrade immensely! On Pharo the results are a bit better, but still very much degraded:
PBDictionary 0.02387 +/-0.00014 PBSTDictionary 1.6577 +/-0.0036
So around 60x faster in case of PDictionary.
If on Set I ensure that all elements will be found, I get the following result for #includes:
PBSetIncludes 0.00440 +/-0.0001100000000000001 PBSTSetIncludes 0.004000 +/-9.6e-5
making PSet marginally slower. However, once you start not finding elements, with the same change as before, (to: size*10 by: 10):
PBSetIncludes 0.00443 +/-0.0001 PBSTSetIncludes 0.9844000000000005 +/-0.0017
So around 220x faster in case of PSet. While PSet and PDictionary never really suffer from collisions, normal Squeak/Pharo dictionaries suffer immensely and grow up to 200x slower. Btw, this last benchmark was run on Pharo.
See above.
Anyway, you are definitely right that the benchmarks suck and that we need better ones.
As for PDictionaries and PSets, nowing that you know exactly what the range of existing hashes is I think it's only logical that it will be faster in many of the operations that need this knowledge, such as #includesKey:, #remove:, #at:put:/#add: and #at:/#includes:.
See above.
Since PDictionary switches between SmallDictionary style and normal dictionary style, it also seems logical that it behaves faster in case that you are in SmallDictionary style without having to care about it. This is also not shown in any benchmark, but it is the case and makes it faster for that case. However, this added test makes it slower for the general dictionary style so removing it might improve things in that case to regain the marginal difference between both dictionaries in the ideal case for Pharo/Squeak.
There is anyway no reason why PDictionary would be slower than Pharo/Squeak dictionaries except for design decisions like the previously mentioned one, and those can be changed if wanted. There is however a good reason why Squeak/Pharo dictionaries are slower and degrade faster, namely colliding hashes that even propagate collisions (also not benchmarked at the moment!).
Set has better cache locality than PSet, I think Dictionary is similar to PDictionary in this regard.
Ok, on the whole I wasn't really planning on pushing these dictionaries too much on Pharo / Squeak. We use them in Pinocchio and are happy with them (especially since we have native support). But they are there and can be used if you want to. I can understand that people are skeptical, but since I don't really have knowledge on building benchmarks (I didn't even set up the current ones, it was a student of mine) I don't think I can help much there. We use them throughout our project and some parts (such as our parser) have gotten a lot faster since we started using them.
I'm not saying that a hashed collection can't be faster than the current implementation, but a few people (including me) implemented serveral variants over the years and the current idea was found the best so far for general purpose. Performance can easily be boosted by primitives, especially in special cases, like this: http://leves.web.elte.hu/LargeIdentityDictionary/ , where I used the existing #pointsTo: primitive to boost #includesKey: (http://leves.web.elte.hu/LargeIdentityDictionary/LargeIdentityDictionary2.pn... ). With a better primitive (for this purpose) all operations can be boosted drastically, but I prefer code written in Smalltalk. I think Cog will be a lot faster in the near future, so the performance gap will be smaller.
Btw, the code would indeed just be released under MIT license which is maybe already automatically the case since I signed that agreement as a contributor to Pharo?
If you didn't upload your code to a Pharo repository, then it's not automatically MIT just because you signed the agreement.
Some answers / comments to the other random questions and remarks:
I wonder what the #pPrimitive:plugin: pragmas stand for in PDictionary's #at:ifAbsent: and #at:put:. Those methods are implemented as primitives in the Pinocchio VM that we are developing. In primitive form most of the methods run ~ twice as fast. This is however not compatible with the way the stack is handled in Squeak/Pharo so can't really be ported. The current HashedCollections don't shrink. Remove is a rarely used operation. Ok, makes sense. One could implement the current dictionaries without associations, that's just less object-oriented. That would generate even less "garbage". True enough. 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. So you change the hash implementation to ensure that they take up the whole array of the Dictionary rather than just the first 4096 + collisions positions. Makes perfect sense for the squeak/pharo implementation but rather strange to enforce this on the whole system imho.
Well, it's only that way in Pharo. In Squeak we did it differently: #identityHash returns the value of the primitive (0..4095) and #scaledIdentityHash does the shifting (0..SmallInteger maxVal). So "old" code that uses #identityHash doesn't break and doesn't have performance issues.
Ok, I hope this clears up some things.
Yes, thanks. Levente
cheers, Toon
_______________________________________________ Pharo-project mailing list Pharo-project@lists.gforge.inria.fr http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
tx levente Stef
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
_______________________________________________ Pharo-project mailing list Pharo-project@lists.gforge.inria.fr http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
do you know if they compare with squeak because levent did a lot of work in that area. Stef On Sep 13, 2010, at 2:55 PM, 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.
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?
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.
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).
_______________________________________________ Pharo-project mailing list Pharo-project@lists.gforge.inria.fr http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
No, they just compared to Pharo 1.1. But the benchmarks are available in a separate package. Adrian On Sep 13, 2010, at 15:52 , Stéphane Ducasse wrote:
do you know if they compare with squeak because levent did a lot of work in that area.
Stef
On Sep 13, 2010, at 2:55 PM, 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.
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?
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.
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).
_______________________________________________ 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
participants (4)
-
Adrian Lienhard -
Levente Uzonyi -
Stéphane Ducasse -
Toon Verwaest