David-
That is very interesting.
Did you also take a look at the one at http://sourceforge.net/
projects/high-scale-lib ? They say its performance only shines for
high thread/cpu counts, but it might be interesting to see where its
numbers lie in the range.
On May 29, 2007, at 11:01 AM, David Ezzio (asmtp) wrote:
Recently, I did some testing of Map implementations under concurrency.
My primary purpose was to verify the reliability of OpenJPA's
ConcurrentHashMap implementation. As I got into it, I saw the
opportunity to get some performance metrics out of the test.
The biggest part of my task was coming up with a reliable and
useful testing framework. I design it with the following two
factors in mind: First, I wanted to test the edge conditions where
an entry had just been added or removed or where a key's value had
just been updated. The idea is that a number of threads add,
remove, and update entries, while other threads check to see if
these recent modifications are visible (or in the case of removals,
not visible). Second, I wanted the testing framework itself to be
free of synchronization. If the testing framework used
synchronization then it would tend to serialize the readers and
writers and thereby mask concurrency issues in the map
implementation under test.
The testing framework uses a non-synchronizing, non-blocking FIFO
queue as the mechanism for the writing threads to communicate their
recent modifications to the reading threads.
To prevent writing threads from overwriting recent modifications
before they could be read and verified, the testing framework walks
the hash map keys in in a linear (or in the case of updates,
circular) order. By using a hash map with a large enough capacity,
readers have the time to verify the recent modifications before the
writer threads come back to modify that part of the key space again.
Using an adapter for the map implementation, the testing framework
starts five writer threads and ten reader threads at the same time.
These threads run wide open for 30 seconds, except that the readers
will give up their time slice if they find nothing on the queue.
The HashMaps were all sized for the needed capacity upon creation,
so no resizing occurred during testing.
I got some interesting results.
Four implementations were tested, Java's unsynchronized HashMap
implementation, Java's synchronized HashMap implementation, Java's
ConcurrentHashMap implementation, and OpenJPA's ConcurrentHashMap
implementation.
Only Java's unsynchronized HashMap failed, as expected, under test.
Under test, this implementation demonstrates its inability to
handle concurrency. The other three implementations worked
flawlessly under test.
The java.util.concurrent.ConcurrentHashMap implementation
(available with Java 5 and 6) was the fastest implementation tested.
Java's synchronized wrapper for the HashMap implementation is one
to two orders of magnitude slower than Java's ConcurrentHashMap
implementation.
OpenJPA's ConcurrentHashMap compares equally with Java's
ConcurrentHashMap in find operations and is 2-4 times slower in
mutating operations.
Implementation Add Remove Update Find-a Find-r Find-u
---------------+------+-------+--------+-------+-------+------
synchronized 103 35 50 40 37 54
concurrent 13.2 6.4 6.1 0.6 0.3 1.1
OpenJPA 29.8 26.6 27.9 0.6 0.6 0.6
Legend:
synchronized:
java.util.Collections.synchronizedMap(new java.util.HashMap())
concurrent: java.util.concurrent.ConcurrentHashMap
OpenJPA: org.apache.openjpa.lib.util.concurrent.ConcurrentHashMap
Add: time for average add operation
Remove: time for average remove operation
Update: time for average update of new value for existing key
Find-a: time to find a recent addition
Find-r: time to NOT find a recent removal
Find-u: time to find a recent update
These times (in microseconds) are representative, but are not the
average of several runs. The tests were run on a Dell Dual Core
laptop under Windows. The performance meter was pegged during the
tests.
David Ezzio
