[infinispan-dev] Bye bye wrappers, ComparingConcurrentHashMapv8 is here (ISPN-2281)
sanne at infinispan.org
Wed Mar 20 05:49:55 EDT 2013
Great stuff. Another benefit to keep in mind is that this can make it
easy to use equality defined as identity of objects (the "=="):
something I'm looking forward for, as with all use cases I'm familiar
with (Lucene Directory + Hibernate OGM) we should be able to maintain
uniqueness of instances per key.
You could take this a step further (maybe as a second step in future)
and break out of the ConcurrentMap contract.. we don't strictly need
that; as from previous brainstorming we can optimise certain
operations - although I don't remember if any is left as Manik already
applied some great improvements - so just to remind it doesn't have to
strictly guarantee the ConcurrentMap contract.
On 19 March 2013 19:56, Mircea Markus <mmarkus at redhat.com> wrote:
> On 18 Mar 2013, at 12:21, Galder Zamarreño wrote:
>> Hi all,
>> A heads up on what is going on with https://issues.jboss.org/browse/ISPN-2281
>> While discussing this, Tristan and I came to the conclusion that we could avoid the need to create some wrappers required to fulfill requirements in this JIRA, and as a side effect, reduce the memory consumption of Infinispan servers, if we could have internal data containers based on concurrent hash maps that took a custom function for equals/hashCode…etc. By doing that, you could effectively have **byte keys and values for maps**.
>> By doing that, you avoid creating wrappers (yippee!) for keys (bye bye ByteArrayKey), and combined with a better way to pass metadata into Infinispan Caches (i.e. version) that is stored within the internal cache entries, you avoid wrapper values too! (bye bye CacheValue).
> Do you plan to use Versioned*CacheValue for storing values? if so you'd still create a version object to be aggregated in Versioned*CacheValue.
>> Doing the latter was relatively simple (I have this stashed), but having a CHM that could take a byte as key wasn't that easy, since we can't change JDK CHM.
> why's that? copyright?
>> This is why, I've created a new CHM, based on the CHMv8, called ComparingConcurrentHashMapv8 (thx Tristan for the name!). The work for this can be seen in: https://github.com/galderz/infinispan/commit/351e29d327d163ca8e941edf873f6d46b43cfae1
> this relies on sun.misc.Unsafe, so won't work with a non-oracle JVM.
> This class would be aggregated in the DataContainer in the case we use infinispan in server mode, right?
>> I'm sending it here so that I can get feedback early on. I've added some tests as well that verify that ComparingConcurrentHashMapv8, with byte keys and values, works as expected, and checks that the expectations are opposite with JDK CHM. It also tests new function-based methods.
>> To make it easier to track changes as original CHMv8 evolves, I've marked all changes with a marker comment that should make it easy to apply same changes in new CHMv8 versions. Plus, with the tests I've added, it can easily be seen if it works as expected or not.
>> Two important TODOs, which will be most likely separated into separate JIRAs:
>> 1. Note that TreeBin has not been modified to use custom equals/hashCode functions. That is cos I need to implement a way to compare byte arrays, i.e. provide equivalent logic for Comparable.compare().
>> 2. Compare memory consumption of a CHMv8 with wrapper classes for byte arrays versus ComparingConcurrentHashMapv8<byte, byte>. I'll do that once it's closer to CR stages. Right now fulfilling requirements in ISPN-2281 is more priority.
>> Finally, I need to do the same thing with out BoundedConcurrentHashMap, iow, provide a way to do comparison based on custom equals/hashCode. That's gonna be my next task, before I get to transform Infinispan Servers to take a type directly, and avoid relying on ByteArrayKey or CacheValue wrappers.
>> IOW, you'll be able to say: create an Infinsipan Server that has String as key and value of type X, where X is the actual data type, no metadata!! The metadata (version, encoding, whatever is requried to fulfill the compatibility reqs in ISPN-2281) will be passed as part of the put/replace…etc (I will email this around when in place).
>> Galder Zamarreño
>> galder at redhat.com
>> Project Lead, Escalante
>> Engineer, Infinispan
>> infinispan-dev mailing list
>> infinispan-dev at lists.jboss.org
> Mircea Markus
> Infinispan lead (www.infinispan.org)
> infinispan-dev mailing list
> infinispan-dev at lists.jboss.org
More information about the infinispan-dev