A memcached implementation in Java
4 likes2,113 views
This document discusses using memcached for caching data in memory across multiple servers. It introduces ReplCache, which is a Java implementation of memcached that adds redundancy by storing each key on multiple servers. ReplCache improves on memcached by allowing data to be migrated when servers shut down and rebalanced when new servers are added. The document also discusses Infinispan, an open source data grid platform that provides features similar to ReplCache along with additional capabilities such as persistence and querying.
A memcached implementation in Java
- 1. A memcached implementation in Java Bela Ban JBoss 2340
- 2. 2 AGENDA > Introduction > memcached > memcached in Java > Improving memcached > Infinispan > Demo
- 3. 3 Introduction > We want to store all of our data in memory – “Memory is the new disk, disk is the new tape” (Tim Bray) – Reducing DB access speeds up the application A network round trip is at least an order of magnitude faster than accessing the disk The DB becomes the bottleneck > However, we'll likely exceed the memory of a single host > Therefore, we spread our data across multiple hosts > The aggregated memories of all hosts serves as our (large virtual) disk – Given enough data redundancy, we can temporarily afford to lose hosts due to crashes, updates etc
- 4. 4 What is memcached ? > Server (daemon) process written in C, providing an in-memory hashmap > Clients can be written in any language – They talk to memcached servers via the memcached protocol > Simple API: put(key,val), get(key), remove(key) > A key always maps to the same memcached server (consistent hashing) > Clients pick the right memcached server based on consistentHash(K) > Given a good consistent hash function, keys are evenly distributed across memcached servers > Typical use cases – PUT(K, V): write K,V to DB and put K,V into the cache – GET(K): get from cache, return V if found. Else fetch from DB, insert into cache and then return V
- 5. 5 memcached JBoss Servlet memcached A Apache JBoss HTTP Servlet memcached B mod_jk memcached protocol JBoss Servlet memcached C DB
- 6. 6 Issues with memcached > Client have to use the memcached protocol to talk to a server – Protocol is ASCII over HTTP, inefficient, parsing of headers > No L1 cache (some memcached clients cache locally though) > memcached servers don't know about each other > When a memcached server leaves, or a new server is started, some keys may now map to a different server ... which doesn't have them → DB lookup > Keys are not redundantly stored on multiple servers – Server crash (or graceful shutdown) causes keys not to be found → DB lookup – On shutdown of a server, keys are not relocated to a different server
- 7. 7 memcached in Java > L1 caches – Clients run in the same address space as ReplCache – Access to data is fast when in L1 cache, no marshalling overhead > All ReplCaches know about each other (in the same cluster) – Data can be migrated when shutting down – On rehashing, data is rebalanced automatically Avoids a DB access > Fast binary protocol (JGroups) – Can be adapted (compression, encryption, batching) via XML config > Fewer network round trips
- 8. 8 ReplCache: memcached improved > memcached has no redundancy – When a server crashes, the keys on that server need to be re-read from the DB > With ReplCache, every key has a replication count – Defines how many times the key should be stored in the cluster – -1: store key on every server – N: store key on N servers Example: put(key,val, 3) stores {key,val} on 3 servers When one of the 3 servers crashes, {key,val} will be copied to another server > Advantage: – We can define redundancy per data element ! – This allows us to use more of the available aggregated memory
- 9. 9 ReplCache JBoss ReplCache Servlet Apache JBoss ReplCache Cluster HTTP Servlet mod_jk JBoss ReplCache Servlet DB
- 10. 10 ReplCache: memcached improved > Data that can be easily re-read from the DB might use repl-count=1 > Data that isn't available in the DB needs to have a repl-count of -1 > Data where we can tolerate 3 concurrent crashes has repl-count=3 > Think of ReplCache as Dynamic RAID
- 11. 11 Demo
- 12. 12 Infinispan > Open Source Data Grid Platform > Simple JSR-107 interface – Cache extends Map > Content distribution based on Consistent Hashing – Ability to address massive heap – 100 nodes with 2GB each, DIST configured with 1 copy => 100GB addressable space from anywhere! > JTA compliant, support for persistence to cache stores, eviction to prevent OOMs > JMX reporting, (upcoming) mgmt console, migration tools > Extremely high performance core container to support high concurrency – State of the art algorithms with minimal use of mutual exclusion
- 13. 13 Infinispan > memcached server module – Coming soon in Infinispan > Speaks both memcached text protocol + custom binary protocol > Able to use any memcached client – Not just Java! C#, C++, Python, PHP > Custom Java client with ability to load-balance/fail-over – Makes use of binary protocol – We expect more clients like this for other systems (C++, etc) to be contributed > Drop-in replacement for memcached farms – With all the added benefits of Infinispan
- 14. 14 Infinispan > Cool new AsyncAPI – Future<V> putAsync(K key, V value) – Allows you to fire off a number of such operations which operate in parallel Operations can be synchronous – And wait for responses using Future.get() > Other upcoming features include – Querying API – JPA-style interface for “POJO Caching” – Distributed executors Runnables moved to the data, not the other way around Map/reduce on your data!
- 15. 15 Outlook > Implementation of file system (java.io) interfaces to use the grid as a large in-memory file system > JDBC: use the grid as data store rather than the disk ? > Hibernate on a grid
- 16. 16 Links > memcached: http://www.danga.com/memcached > ReplCache: http://www.jgroups.org/memcached.html and http://www.jgroups.org/replcache.html > ReplCache WebStart demo: http://www.jgroups.org/jnlp/replcache.jnlp > ReplCache Flash demo: http://www.jgroups.org/movies/ReplCache.swf > Infinispan: http://www.infinispan.org
- 17. Bela Ban http://www.jboss.com JBoss / Red Hat bela@jboss.com