ofbiz-notifications mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ben Manes (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (OFBIZ-6747) Replace ConcurrentLinkedHashMap by Caffeine
Date Tue, 12 Jun 2018 19:17:00 GMT

    [ https://issues.apache.org/jira/browse/OFBIZ-6747?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16510062#comment-16510062
] 

Ben Manes commented on OFBIZ-6747:
----------------------------------

In regards to fun algorithms, some resources to look into are...
 * For general concurrency, the book [The Art of Multiprocessor Programming|https://www.amazon.com/Art-Multiprocessor-Programming-Revised-Reprint/dp/0123973376]
is excellent. It covers lock-free algorithms and is written by researchers who authored many
well known papers.
 * The concurrency scheme used by my caches is inspired by a database's write-ahead log. The
classic approaches are to either thrash on a global lock to perform a tiny amount of work
(usually pointer flipping) or avoid "movement" by using a CLOCK-based or random sampling policy.
By instead recording into a ring buffer and replaying under a tryLock, the lock contention
is replaced by a cheaper buffer CAS. This lets me uses non-threadsafe algorithms that are
O(1) and more intelligent. The idea is simple and implementation just takes work to squeeze
out performance.
 * For eviction, see the ACM paper on our policy. The pre-published is on the earlier link,
while the peer reviewed version can be freely downloaded from the link on the README. I wrote
an [overview|http://highscalability.com/blog/2016/1/25/design-of-a-modern-cache.html] ([slides|https://docs.google.com/presentation/d/1NlDxyXsUG1qlVHMl4vsUUBQfAJ2c2NsFPNPr2qymIBs/edit#slide=id.p])
for HighScalability. We have a followup paper under review that makes it adaptive. If interested,
I can send that over private email due to submission restrictions (my email is on github).
 * For expiration, see this excellent article by the Kafka folks on [Hierarchical TimerWheel|https://www.confluent.io/blog/apache-kafka-purgatory-hierarchical-timing-wheels].
This is an amortized O(1) priority queue that leverages time to use hashing instead of tree-like
comparisons. Caffeine is the only cache that I'm aware of to use this, but this data structure
is fundamental in kernel subsystems. Its really elegant and fast!

And the rest is all elbow grease! :)

> Replace ConcurrentLinkedHashMap by Caffeine
> -------------------------------------------
>
>                 Key: OFBIZ-6747
>                 URL: https://issues.apache.org/jira/browse/OFBIZ-6747
>             Project: OFBiz
>          Issue Type: Task
>            Reporter: Ben Manes
>            Assignee: Jacques Le Roux
>            Priority: Minor
>
> Similar to OFBIZ-3779, please consider upgrading the library used by [UtilCache|https://github.com/apache/ofbiz/blob/trunk/framework/base/src/org/ofbiz/base/util/cache/UtilCache.java]
(v1.2). The current version is 1.4.2 and is the last major release planned.
> The preferable alternative would be to upgrade to [Caffeine|https://github.com/ben-manes/caffeine].
This is a Java 8 rewrite based on what I've learned since developing CLHM and Guava's cache.
As expected it provides [superior performance|https://github.com/ben-manes/caffeine/wiki/Benchmarks].
It also provides a [near optimal|https://github.com/ben-manes/caffeine/wiki/Efficiency] eviction
policy.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Mime
View raw message