Hi,
Just to illustrate what Bend's benchmark keys are made of so one can
interpret the results accordingly. The full set of keys consists of
250,000 distinct keys (500*500). With GOOD_HASH multiplier, each of them
has it's own distinct hashCode() and a HashMap filled with all 250k keys
looks like:
Capacity: 524288
Load factor: 0.75
Size: 250000
Bin sizes: 0*274288 1*250000 total=250000
Empty bins: 52.3 %
Unique hash codes per bin: 0*274288 1*250000 total=250000
With BAD_HASH multiplier, there are only 18963 unique hashCodes:
Capacity: 524288
Load factor: 0.75
Size: 250000
Bin sizes: 0*505325 1*74 2*74 3*74 4*74 5*74 6*74 7*74
8*74 9*74 10*74 11*74 12*74 13*8822 14*9253 total=250000
Empty bins: 96.4 %
Unique hash codes per bin: 0*505325 1*18963 total=18963
Although majority of entries constitute the bins of size 13 or 14,
there's only a single hashCode value per bin.
So in this benchmark, treeifying with non-comparable keys is a waste of
effort.
Are there (non-forged) sets of non-comparable keys with hashCodes where
treeifying makes sense?
Regards, Peter
P.S. The below modified benchmark source is update here:
http://cr.openjdk.java.net/~plevart/jdk9-dev/HM.comparableClassFor/HashMapCollision.java
The HashMap simulator that prints stats about bins and hashCodes is here:
http://cr.openjdk.java.net/~plevart/jdk9-dev/HM.comparableClassFor/HashMapSimulator.java
On 01/11/2015 04:09 PM, Peter Levart wrote:
Hi,
I wasn't comfortable with Bernd's HMH benchmark results jitter, so I
changed the mode of operation to be SingleShotTime (since a particular
invocation is from 0.6 to 3sec anyway). GC is triggered before each
invocation (-gc true). I also added -XX:-TieredCompilation VM option
and run 6 forks of 10 iterations of each test. By Doug's suggestion I
also added a variant of unchanged HashMap where TREEIFY_THRESHOLD = 1
<< 20, UNTREEIFY_THRESHOLD = TREEIFY_THRESHOLD - 2,
MIN_TREEIFY_CAPACITY = TREEIFY_THRESHOLD * 4 as a reference to compare
with. Here are the results:
Original JDK9 HashMap:
Benchmark (initialSize) Mode
Samples Score Score error Units
j.t.HashMapCollision.badDistNoComp 16 ss 60
3011.738 78.249 ms
j.t.HashMapCollision.badDistWithComp 16 ss 60
2984.280 48.315 ms
j.t.HashMapCollision.goodDistNoComp 16 ss
60 682.060 52.341 ms
j.t.HashMapCollision.goodDistWithComp 16 ss
60 685.705 55.183 ms
Original JDK9 HashMap with TREEIFY_THRESHOLD = 1 << 20:
Benchmark (initialSize) Mode
Samples Score Score error Units
j.t.HashMapCollision.badDistNoComp 16 ss 60
2780.771 236.647 ms
j.t.HashMapCollision.badDistWithComp 16 ss 60
2541.740 233.429 ms
j.t.HashMapCollision.goodDistNoComp 16 ss
60 757.364 67.869 ms
j.t.HashMapCollision.goodDistWithComp 16 ss
60 671.617 54.943 ms
Caching of comparableClassFor (in ClassRepository - good for
heterogeneous keys too):
Benchmark (initialSize) Mode
Samples Score Score error Units
j.t.HashMapCollision.badDistNoComp 16 ss 60
3014.888 71.778 ms
j.t.HashMapCollision.badDistWithComp 16 ss 60
2279.757 54.159 ms
j.t.HashMapCollision.goodDistNoComp 16 ss
60 760.743 70.674 ms
j.t.HashMapCollision.goodDistWithComp 16 ss
60 725.188 67.853 ms
Caching of comparableClassFor (internally - good for homogeneous keys
only):
Benchmark (initialSize) Mode
Samples Score Score error Units
j.t.HashMapCollision.badDistNoComp 16 ss 60
3026.707 84.571 ms
j.t.HashMapCollision.badDistWithComp 16 ss 60
2137.296 66.140 ms
j.t.HashMapCollision.goodDistNoComp 16 ss
60 635.964 8.213 ms
j.t.HashMapCollision.goodDistWithComp 16 ss
60 685.129 46.783 ms
Regards, Peter
On 01/11/2015 12:55 PM, Peter Levart wrote:
On 01/11/2015 02:27 AM, Martin Buchholz wrote:
Peter,
You are adding the ability to add "app-specific storage" to Class
objects ("Class-local variables"?), which is pretty unusual.
Well, that was my intention, since the logic about what should be
cached is very specific to the usecase and might change in the
future. Anyway, this is only internal API. Users have a public
alternative in ClassValue. That's one reason. The other is space
overhead introduced when caching with ClassValue and inability to
initialize ClassValue so very early in the boot-up sequence.
I was thinking instead of a very dumb 1-element cache, remembering
Class and comparableClassFor, which will work for typical
homogeneous HashMaps.
This seems like a good idea. We would actually need only one field of
type Class<?> and a boolean flag.
Unfortunately, comparableClassFor is a static method used also from
various other contexts that don't have access to HashMap instance,
for example from TreeNode. We would have to extend the internal API
with an additional HashMap argument to pass the HM instance around.
Not to mention that this would be tricky because retaining the last
used comparable Class object in the HM instance could prevent GC from
releasing a ClassLoader in an app server environment for example. A
WeakReference<Class<?>> would have to be used and new WeakReference
object created each time cached value changes. Unless we cache only
the 1st comparableClassFor result and never change it, which has the
same cache-hit ratio for homogeneous keys.
Right, here's what this looks like:
http://cr.openjdk.java.net/~plevart/jdk9-dev/HM.comparableClassFor/HomogeneousKeysCache/webrev.01/
I modified Bernd's JMH benchmark a little to use ThreadLocalRandom
insted of Random, so results express more what is going on with
HashMap and less with Random synchronization:
http://cr.openjdk.java.net/~plevart/jdk9-dev/HM.comparableClassFor/HashMapCollision.java
Results:
Original JDK9 HashMap:
Benchmark (initialSize) Mode
Samples Score Score error Units
j.t.HashMapCollision.badDistNoComp 16 avgt
6 3101.247 435.866 ms/op
j.t.HashMapCollision.badDistWithComp 16 avgt
6 2410.202 478.247 ms/op
j.t.HashMapCollision.goodDistNoComp 16 avgt
6 615.100 7.063 ms/op
j.t.HashMapCollision.goodDistWithComp 16 avgt
6 614.229 159.558 ms/op
Caching of comparableClassFor (in ClassRepository - good for
heterogeneous keys too):
Benchmark (initialSize) Mode
Samples Score Score error Units
j.t.HashMapCollision.badDistNoComp 16 avgt
6 3305.967 652.791 ms/op
j.t.HashMapCollision.badDistWithComp 16 avgt
6 2030.965 241.910 ms/op
j.t.HashMapCollision.goodDistNoComp 16 avgt
6 611.202 6.440 ms/op
j.t.HashMapCollision.goodDistWithComp 16 avgt
6 582.890 4.896 ms/op
Caching of comparableClassFor (internally - good for homogeneous keys
only):
Benchmark (initialSize) Mode
Samples Score Score error Units
j.t.HashMapCollision.badDistNoComp 16 avgt
6 3265.673 660.030 ms/op
j.t.HashMapCollision.badDistWithComp 16 avgt
6 1875.204 224.682 ms/op
j.t.HashMapCollision.goodDistNoComp 16 avgt
6 598.949 25.484 ms/op
j.t.HashMapCollision.goodDistWithComp 16 avgt
6 585.278 8.103 ms/op
Regards, Peter
On Sat, Jan 10, 2015 at 5:01 AM, Peter Levart
<peter.lev...@gmail.com <mailto:peter.lev...@gmail.com>> wrote:
On 01/10/2015 01:20 AM, Doug Lea wrote:
On 01/09/2015 06:29 PM, Martin Buchholz wrote:
Given the prevalence of sub-optimal hashcodes, my own
intuition is also that
raising the treeification threshold from 8 will be a win.
That's what I thought at first. But 8 is a better choice for
String
and other Comparable keys, which account for the majority of
HashMaps
out there. (For non-comparables, infinity is the best threshold.)
How much slower should we make the most common cases to make
the others
faster? The only way to decide empirically is to take a large
corpus of programs and vary thresholds. Short of that, speeding up
comparableClassFor is still the best bet for reducing impact on
non-comparables.
Hi Doug,
comparableClassFor() for non-comparables that don't implement
Comparable is already as fast as it can be (the 1st check is
instanceof Comparable). For other comparables (and
non-comparables) that implement Comparable (except for String
which is special-cased), we could improve the situation by
caching the result.
Here's another attempt at that. This time it uses plain old JDK1
stuff, so it actually works even in HashMap (using
IdentityHashMap so no danger of circular usage if it is to be
applied to CHM also):
http://cr.openjdk.java.net/~plevart/jdk9-dev/Class.getGenericDerivative/webrev.01/
<http://cr.openjdk.java.net/%7Eplevart/jdk9-dev/Class.getGenericDerivative/webrev.01/>
With this patch, the results of Bernd's JMH benchmark do give
some boost to keys that implement Comparable (badDistWithComp case).
These are the results with original JDK9 HashMap:
Benchmark (initialSize) Mode Samples Score Score
error Units
j.t.HashMapCollision.badDistNoComp 16 avgt 6
3104.047 278.057 ms/op
j.t.HashMapCollision.badDistWithComp 16 avgt 6
2754.499 243.780 ms/op
j.t.HashMapCollision.goodDistNoComp 16 avgt 6
1031.992 26.422 ms/op
j.t.HashMapCollision.goodDistWithComp 16 avgt 6
1082.347 30.981 ms/op
And this is with patch applied:
Benchmark (initialSize) Mode Samples Score Score
error Units
j.t.HashMapCollision.badDistNoComp 16 avgt 6
3081.419 386.125 ms/op
j.t.HashMapCollision.badDistWithComp 16 avgt 6
2116.030 281.160 ms/op
j.t.HashMapCollision.goodDistNoComp 16 avgt 6
1015.224 81.843 ms/op
j.t.HashMapCollision.goodDistWithComp 16 avgt 6
1078.719 38.351 ms/op
Caching is performed as part of Class generic types information
caching (ClassRepository), so there's no overhead for those that
don't need generic types information. All logic is kept inside
(C)HM.
Regards, Peter
-Doug
